A comprehensive survey on diagnosis and assessment of Parkinson’s disease via plantar pressure analysis Authors: Xiaotian Wang, Zhifu Zhao, Lingquan Lin, Fu Li, Fei Qi, Xinbo Wang, Le Han Published: 2026 ABSTRACT Parkinson’s disease (PD) is a progressive neurodegenerative disorder. Its motor symptoms include bradykinesia, rigidity, and postural instability, while gait abnormalities serve as critical early biomarkers. Traditional diagnostic methods normally rely on subjective clinical assessments, highlighting the need for objective, quantifiable tools. Plantar pressure measurement systems provide high-resolution spatiotemporal and dynamic load data, directly reflecting foot-ground biomechanical interactions and offering insights into balance control and disease progression. This review comprehensively surveys plantar pressure analysis for PD diagnosis and assessment, encompassing data acquisition, existing datasets, preprocessing methodologies, feature engineering, and neural-network-based diagnostic models. We highlight approaches employing artificial intelligence (AI), which have been extensively adopted with promising outcomes for gait-based PD diagnosis and monitoring. Finally, challenges including heterogeneous comorbidities and the interpretability and ethical issues of AI are discussed, alongside future directions in telemonitoring and edge computing. By bridging biomechanics and AI, this survey underscores the transformative potential of plantar pressure analysis for early, objective, and personalized PD diagnostics. Future research should integrate edge intelligence with wearable sensing to develop wearable, continuous monitoring and intervention systems. Simultaneously, multi-center collaborative studies should validate the universal applicability of these systems in real-world clinical settings. Introduction Parkinson’s disease (PD), the second most prevalent neurodegenerative disorder worldwide, poses persistent challenges for achieving accurate early diagnosis and longitudinal monitoring. This critically limits progress in clinical practice and biomedical engineering research1–3. According to the Global Burden of Disease Study, the worldwide prevalence of PD reached ~11.9 million in 2021 and is projected to rise dramatically to 25.2 million by 2050, representing a substantial 112% increase4. PD symptoms are classified into two main categories: motor and non-motor manifestations5. Among these, motor symptoms, including bradykinesia, resting tremor, rigidity, and gait abnormalities like Freezing of Gait (FOG)2,3,6–8, are not only core features but also crucial for clinical diagnosis and evaluation. Currently, the major PD diagnosis approaches are still dependent on the patient’s clinical evaluation9. The gold standard for the evaluation and monitoring of PD is the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS)10,11, followed by the well-established Hoehn and Yahr (H&Y) Scale12,13, both of which incorporate assessment of gait abnormalities as a critical component. Nevertheless, conventional gait assessment predominantly depends on clinician-based observations and patient-reported outcomes, which are inherently subjective, lack sensitivity to subtle motor fluctuations in the early stages of PD, and fail to provide continuous, high-resolution quantitative metrics necessary for reliable monitoring of disease progression. This diagnostic gap has gained increasing interest in leveraging quantifiable biomechanical signals and objective diagnostic approaches for PD assessment14,15. Gait disturbances, as complex biomechanical manifestations, offer a critical window into the pathophysiology of PD progression16–18. As illustrated in Fig. 1, research employing gait analysis for PD has grown rapidly in recent years. This trend has been largely fueled by the advancement of sensor-based gait analysis, which utilizes diverse technologies such as inertial measurement units (IMUs)19, optical motion capture20, surface electromyography (sEMG)21, and force platforms22 to capture high-resolution spatiotemporal, kinetic, and kinematic gait parameters23–25. Among these, the force distribution of plantar pressure allows researchers to understand the human mechanical behavior, providing detailed information for gait analysis. Therefore, the vertical ground reaction force (VGRF)—a noninvasive, high-resolution biomarker capturing gait dynamics17,26 has emerged as a pivotal tool for motor function evaluation and remained a cornerstone for assessing dynamic balance27 and lower-limb coordination28, particularly in neurodegenerative disorders such as PD29,30. VGRF signals, acquired via force plates31 or wearable inertial sensors32,33, enable the quantification of spatiotemporal gait parameters (e.g., stride length, cadence)15, dynamic features (e.g., pressure distribution, force-time integrals), and nonlinear characteristics (e.g., entropy, fractal dimensions)30,34.Fig. 1Annual trends in PD research using gait analysis.This line chart illustrates the annual number of peer-reviewed articles focused on PD research utilizing gait analysis, spanning from 2014 to 2025. The x axis denotes the publication year, and the y axis represents the number of articles. The data show a consistent, year-over-year increase in publication output over the 12-year period. Annual trends in PD research using gait analysis. This line chart illustrates the annual number of peer-reviewed articles focused on PD research utilizing gait analysis, spanning from 2014 to 2025. The x axis denotes the publication year, and the y axis represents the number of articles. The data show a consistent, year-over-year increase in publication output over the 12-year period. With the rapid development of hardware technologies, advances in microelectromechanical systems (MEMS) and wireless communication have enabled portable, low-cost VGRF acquisition through instrumented insoles or smart footwear35–37, facilitating gait monitoring beyond laboratory settings. Advancements in signal processing, especially in Machine Learning and Deep Learning, have propelled VGRF-based PD detection into a promising frontier. Figure 2 presents the comprehensive technology roadmap for PD diagnosis and staging, derived from VGRF data. The workflow initiates with data acquisition, proceeds to signal quality enhancement via preprocessing, and subsequently extracts multidimensional spatiotemporal and dynamic features. Finally, classification models are leveraged to achieve dual objectives: disease identification and severity stratification. The present review is guided by the aforementioned roadmap, which delineates the various steps involved.Fig. 2Overall framework for PD diagnosis and staging based on VGRF analysis.The workflow starts with gait dataset acquisition, where VGRF signals are captured during walking. Raw data are then standardized in the preprocessing stage. Key gait parameters are extracted, and discriminative features are selected in the feature engineering step. Machine learning, deep learning, or statistical models are trained and validated using metrics including accuracy, sensitivity, specificity, and F1-score. The pipeline yields classification and severity assessment via H&Y or UPDRS, stratifying patients into early, moderate, severe, and advanced stages based on clinical scores. Overall framework for PD diagnosis and staging based on VGRF analysis. The workflow starts with gait dataset acquisition, where VGRF signals are captured during walking. Raw data are then standardized in the preprocessing stage. Key gait parameters are extracted, and discriminative features are selected in the feature engineering step. Machine learning, deep learning, or statistical models are trained and validated using metrics including accuracy, sensitivity, specificity, and F1-score. The pipeline yields classification and severity assessment via H&Y or UPDRS, stratifying patients into early, moderate, severe, and advanced stages based on clinical scores. In recent years, artificial intelligence (AI) has demonstrated significant advancements in the application of plantar pressure data to advance research on PD. In accordance with the PRISMA guidelines, we conducted a literature search for studies published between 2008 and October 2025 on the diagnosis of PD using plantar pressure data. The details of the survey methodology are provided in the supplementary information. The overall framework of the article is shown in Fig. 3. Following this overview, Section Introduction presents the introduction to the topic. Section Results presents the results of the literature review. Section Discussion provides a detailed discussion of the current research landscape, challenges, and future directions. Section Methods elaborates on the methodology adopted in this work.Fig. 3Overall framework of the paper.This figure outlines the paper’s complete workflow for PD diagnosis and severity assessment based on VGRF gait data. The framework is structured into four key sections: Introduction, Results, Discussion, and Methods. Overall framework of the paper. This figure outlines the paper’s complete workflow for PD diagnosis and severity assessment based on VGRF gait data. The framework is structured into four key sections: Introduction, Results, Discussion, and Methods. Results Gait and plantar pressure characteristics Motor dysfunction in PD patients manifests as systemic abnormalities in both gait and plantar pressure. In terms of gait, core characteristics include reduced walking speed, shortened stride length, shuffling steps, prolonged double support time, and the hallmark symptom of “FOG”. Regarding plantar pressure, the data reveal a typical abnormal pattern: reduced peak pressure during heel strike, reflecting delayed initiation, a subsequent significant increase in load in the midfoot region, suggesting a collapsed arch or “flat-foot landing”, and finally, weakened propulsive force during the forefoot push-off phase. Furthermore, the COP trajectory also exhibits a shortened progression path, increased medial-lateral sway, and heightened variability, collectively indicating multi-dimensional impairments in patients’ postural regulation and motor execution. Data acquisition devices An analysis of data acquisition technologies shows that in-shoe pressure sensors are the absolute mainstream in this field of research. In studies involving direct data collection, the usage rate of in-shoe pressure sensors is ~77.3%, far exceeding that of force plate systems (~22.7%), with the Tekscan F-Scan series being the most frequently adopted commercial system. This preference is even more pronounced in studies using public datasets: of 67 such studies, a staggering 94.0% (63 studies) utilized data originating from in-shoe sensors, while reliance on force-sensitive switches (4.5%) or force plates (1.5%) was minimal. This highlights the overwhelming advantage of in-shoe technology due to its portability and applicability. Dataset Based on the 102 studies reviewed, the objectives are mainly divided into three categories: diagnosis only (45.1%), a combination of both (32.4%), and severity assessment only (22.5%). Regarding data sources, researchers show a clear preference for public datasets, with an overall usage rate of 81.4% (83 studies). In particular, a core dataset from the PhysioNet platform, containing ~93 PD patients and 73 healthy controls, was used in 82 of the 102 total studies, establishing it as a de facto standard in the field. In contrast, private datasets typically feature smaller sample sizes (e.g., 2-15 PD patients in diagnosis studies) and are more often used for targeted exploratory research or studies focused on specific symptoms like FOG. Signal processing Median filter (n = 13) is a widely adopted signal processing technique, highly robust to outliers. Butterworth filtering (n = 5) is well-suited for frequency-domain noise reduction, while Savitzky-Golay filtering (n = 2) effectively preserves the local characteristics of gait signals. Z-score normalization (n = 11) ensures feature scale consistency, and gait cycle-based data segmentation enables precise modeling of distinct motion phases. Notably, given the limited data volume typical of VGRF time-series data, data augmentation techniques (n = 10), such as window slicing and permutation, are crucial for improving model generalization capabilities. VGRF-based gait feature The application of various gait features in PD diagnosis exhibits a distinct hierarchical structure. Spatiotemporal features (n = 65) are widely used in gait variability analysis and FOG assessment. In addition, frequency-domain analysis (n = 24) is broadly adopted to capture periodic abnormalities, such as tremors, through spectral analysis. Meanwhile, force-domain analysis (n = 17) is primarily utilized to quantify changes in VGRF data, while time-domain analysis (n = 16) offers distinct advantages in reflecting temporal characteristics. While less prevalent, wavelet domain (n = 7) and time-frequency domain analysis (n = 67) are suitable for detecting non-stationary transient changes in VGRF signals. In contrast, kinematic features typically rely on data such as acceleration and angular velocity from devices like IMUs. Anthropometric features (n = 7), serving as demographic variables that reflect individual physiological differences, are primarily employed to normalize or correct for body type variations. Feature selection Filter methods are frequently employed in the analysis of high-dimensional sensor data due to their computational efficiency (n = 11). Among them, MRMR (n = 4) is the most commonly used, as it reduces feature redundancy while preserving relevance to disease labels. This approach is often applied in multi-sensor VGRF datasets for feature selection. Relief-F (n = 3) selects locally discriminative features based on nearest-neighbor distance, facilitating the identification of features that distinguish between individuals with PD and HC. Mutual information (n = 2) and ANOVA (n = 2) provide complementary perspectives from nonlinear and linear statistical associations, respectively. Wrapper methods (e.g., RFECV) incorporate classifier-specific feedback during iterative selection, but their computational complexity poses challenges for application to large-scale datasets. PD diagnosis and assessment methods Research on plantar pressure-based diagnosis and severity assessment of PD has shown a significant upward trend. Among them, 102 studies are primarily dedicated to the two-class diagnosis of PD and severity assessment based on the staging model. In the field of machine learning applications for plantar pressure-based PD diagnosis and severity assessment, SVM demonstrates the highest prevalence at 36.4%, followed closely by RF at 20.2%. Subsequently, DT and KNN account for 17.6% and 13.5%, respectively. Within the deep learning paradigm, CNN establishes absolute dominance at 43.9%. LSTM comprises 19.5% of applications. Notably, GNN and Transformer each represent 6.1% of implementations. A detailed performance analysis of each model can be found in Section Model performance and limitations. Validation and evaluation In research on foot pressure-based diagnostic models for PD, K-Fold cross-validation is the most widely used, accounting for 54%. Hold-out and LOOCV comprise 15% and 11% of the cases, respectively, whereas LODO accounts for 6%. Additionally, some studies have employed supplementary evaluation approaches such as cross-dataset validation, Wilcoxon rank-sum test, and two-tailed t test. In terms of model evaluation metrics, accuracy is the most frequently used indicator, representing 22.9% of usage, sensitivity ranks second with a proportion of 20.2%, and specificity follows at 13.8%, placing third. The F1-score and precision are also commonly adopted, accounting for 12.7% and 11.4%, respectively. The AUC-ROC curve is also employed in some studies, accounting for 5% of the total. Beyond these, metrics such as MCC, CC, RMSE, and MAE are also applied, though their adoption remains relatively limited. Discussion Plantar pressure signal analysis holds significant potential for discriminating PD patients from healthy controls, with the rapid advancements in deep learning techniques and wearable insole-based plantar pressure devices. For clinical applications, the analysis of plantar pressure signals demonstrates promise as a non-invasive tool to: (i) differentiate PD patients from healthy controls, (ii) quantify disease severity through objective metrics such as UPDRS score prediction, and (iii) detect specific motor impairments, including FOG. Plantar pressure monitoring demonstrates potential for early detection of PD through quantification of paraclinical gait/postural perturbations, preceding clinically observable motor symptom onset. Along with the rapid advancements in signal processing and deep learning, current methodologies offer objective, continuous, and quantifiable gait/postural metrics, which are distinct from subjective clinical scales such as the UPDRS-III. Sophisticated algorithms can precisely capture distinct patterns associated with specific motor symptoms, particularly gait asymmetry, stride length variability, and postural instability, thereby providing a technical foundation for quantitative and precision medicine. Furthermore, wearable in-shoe sensors allow chronic, real-world quantification of disease dynamics and intervention outcomes, creating pathways for ecologically valid assessment beyond clinical settings. Continued innovation in pressure sensor systems is anticipated to drive significant cost reduction in production, enhancing cost-effectiveness for population-level deployment in PD management. Consequently, plantar pressure-based gait analysis is positioned to emerge as a clinically significant modality in PD research and management, offering complementary value to established assessment paradigms. Publicly accessible, high-quality datasets with detailed clinical annotations remain scarce, constraining algorithm generalization and technique development. High data acquisition and annotation costs, combined with health data privacy constraints, hinder the construction of large-scale datasets. To solve these problems, on the one hand, standardized protocols for foot pressure and gait analysis should be developed to achieve high-quality, multicenter, homogeneous datasets. On the other hand, advanced computational approaches, including transfer learning, domain adaptation, semi-/self-supervised learning, and data generation techniques, could be studied to alleviate data limitations while enhancing data diversity. Notably, the dataset constructed in the present study mainly consists of clinically diagnosed PD patients and healthy elderly subjects and does not yet include individuals with other neurological disorders presenting similar motor symptoms, such as Huntington’s disease and dementia with Lewy bodies. In future work, multi-disease clinical datasets could be further expanded and combined with advanced machine learning strategies such as contrastive learning, which may further improve the generalization ability and diagnostic robustness of the models. The absence of unified standards for indicators in plantar pressure data acquisition compromises comparability across studies. To promote the establishment of standardized and homogeneous datasets, we analyzed relevant studies38,39 and proposed several methodological recommendations. First, data acquisition protocols should be tailored according to research objectives. For example, natural single-task walking is recommended for fall risk assessment, whereas dual-task or fast-walking paradigms may be more suitable for disease severity grading, incorporating cognitive tasks such as serial subtraction or backward counting. Second, the recommended test distance should not exceed 10 meters, with each trial lasting at least 1 minute, and appropriate rest intervals should be provided to minimize fatigue effects and maintain gait stability. Following data collection, rigorous preprocessing is essential, particularly accurate gait cycle segmentation. Common approaches include defining gait cycles based on the interval between ipsilateral heel strikes and adaptive sliding window segmentation methods, which can be selected according to specific research scenarios and analytical requirements. Current algorithms suffer from limited generalization capability and sub-optimal robustness on external validation cohorts, revealing prevalent over-fitting to training data distributions. This is mainly caused by: inter-subject variability (age, body mass, comorbidities, gait preferences), environmental perturbations (ground surfaces, footwear), and heterogeneous sensor platforms, a combination that significantly compromises real-world deployment reliability. To address these limitations, first, more robust models with enhanced generalization capability should be explored, such as capturing spatio-temporal dependencies through graph neural networks or incorporating foundation models to boost diagnostic performance. Second, data collection should enable longitudinal tracking of individual patients. Through continuous monitoring of changes in each patient’s condition, the model can better capture personalized disease characteristics and progression patterns, thereby facilitating precise patient-specific monitoring and early detection of abnormalities. Third, ensemble learning can be adopted to fuse the prediction results of multiple models, so as to reduce the variance of individual models and improve the overall stability and generalization ability. Currently, most PD diagnostic methods based on VGRF rely on single-modal data, which tends to result in incomplete characterization of movement features and further limits the generalization ability and diagnostic stability of models. The first fusion strategy is the integration of kinematic and kinetic information. IMU and VGRF reflect kinematic and kinetic information, respectively, and they are highly complementary in describing movement patterns, thus providing an ideal solution for more comprehensive PD assessment. The second approach enables the fusion of VGRF data, biomarkers, and neuroimaging data. Biomarkers, including α-synuclein and dopa decarboxylase, have been applied to PD detection and severity assessment40,41. Dopaminergic molecular imaging techniques, such as dopamine transporter PET imaging, provide support for identifying pathological changes42,43. However, these indicators still exhibit limitations in the continuous progression of the disease. As a high-precision and dynamic motion biomarker, VGRF can sensitively reflect changes in motor function in real time. Its combination with biological and imaging data is expected to further improve the overall accuracy of PD diagnosis. Traditional ML-based methods rely on labor-intensive handcrafted feature extraction, risking information loss and sub-optimal biomarker discovery. Deep learning models with particularly complex architectures suffer from epistemological opacity, hindering clinical adoption due to unverifiable decision logic. Critical questions remain unresolved: Which features drive diagnostic decisions? How do they map to pathophysiological mechanisms? To address the challenges of feature engineering and interpretability, practical and actionable solutions can be developed from three perspectives: automated feature extraction, enhanced model interpretability, and integration of pathophysiological mechanisms—aiming to break through current technical bottlenecks. On the one hand, an adaptive multi-scale feature extraction framework (incorporating attention mechanisms and GNNs) can enable automated feature mining from plantar pressure signals. On the other hand, a cross-modal knowledge fusion model constructed by integrating multi-source clinical data, introducing knowledge graphs, and embedding pathophysiology-consistent regularization terms enables biologically credible feature representations rather than relying purely on data-driven outputs. These measures collectively resolve the inefficiency of traditional feature engineering and the lack of mechanism-based mapping. For clinical applications, systems should be designed to (1) automatically mine latent biomarkers from plantar pressure data; (2) intuitively present diagnostic decision-making bases (e.g., annotating abnormal plantar pressure regions); and (3) predict disease risks using extracted plantar pressure features. Such designs provide reliable support for clinical diagnosis and facilitate the translation of plantar pressure signal analysis into practical clinical scenarios. First, current research on foot pressure data acquisition and analysis mainly remains in laboratory validation stages, lacking large-scale, multi-center, and daily data collection coupled with clinical validation studies. Acquiring foot pressure data in unsupervised home environments introduces higher noise levels, which leads to greater challenges for data cleaning and management. Secondly, the correlation between existing performance metrics (such as Accuracy and AUC) and actual clinical needs (such as sensitivity and specificity for early diagnosis, and predictive value) requires further in-depth investigation and standardization. At last, a key challenge lies in achieving effective integration of the acquisition systems and algorithmic outputs into existing clinical workflows. To address these issues, first, context-specific indications such as screening in high-risk cohorts, diagnostic support in community clinics, and progression tracking need to be established. Second, rigorously designed clinical trials should be implemented to validate the algorithm’s diagnostic performance and clinical utility in real-world settings. At last, applications of wearable plantar pressure insoles and biomechanical data in home-based health monitoring, telemedicine, digital biomarker development, and differential diagnosis of neurodegenerative disorders should be investigated. This review methodically appraises the research advancements in the domain of PD diagnosis and severity assessment employing plantar pressure signals. The review emphasizes the pivotal technical intricacies, encompassing data acquisition, preprocessing, feature extraction, model construction, and algorithm evaluation. The study demonstrates that plantar pressure signals, as a high-resolution, dynamic biomechanical indicators reflecting gait and posture stability, can provide an important basis for objective diagnosis and disease grading of PD. The incorporation of advanced machine learning and deep learning algorithms into this domain has been shown to enhance diagnostic accuracy and severity assessment precision, underscoring its potential for broad applications in clinical-assisted diagnosis and remote monitoring. Despite the advancements demonstrated in the present study, the clinical implementation of PD detection based on plantar pressure analysis remains encumbered by numerous challenges. These challenges encompass the heterogeneity of patient co-morbidities, the limited interpretability of deep models, and the absence of standardized data resources. Future research should focus on enhancing the generalizability of algorithms across various clinical settings and improving the transparency and clinical acceptability of the models' decision-making process. Furthermore, with the rapid development of telemedicine and edge computing technologies, the construction of efficient and low-power smart wearable systems will provide new ideas for continuous monitoring and home intervention in PD. Integration of biomechanical and AI methodologies holds promise for enhancing the efficacy of plantar pressure analysis in facilitating early identification, precise assessment, and personalized diagnosis and treatment of PD. Methods Data collection and datasets for PD diagnosis Section Gait and plantar pressure characteristics in PD introduces the characteristic gait abnormalities and plantar pressure alterations associated with PD, laying the physiological foundation for pressure-based assessment methods. Section Plantar pressure data acquisition presents an overview of the devices used for plantar pressure data acquisition, including both laboratory-grade force plates and wearable insole systems, highlighting their sensing technologies and application contexts. Section Plantar pressure datasets describes several publicly available and self-constructed plantar pressure datasets, outlining their sensor configurations, participant demographics, and relevance to PD diagnosis and monitoring research. Gait and plantar pressure characteristics in PD PD significantly impacts patients’ walking patterns, resulting in distinctive gait abnormalities. Typical symptoms of PD primarily encompass resting tremor, rigidity, bradykinesia, postural instability and FOG16,44–46. The observed kinematic deficits are associated with measurable alterations in ground reaction forces and plantar pressure distribution. Individuals with PD often exhibit a reduced peak pressure or impact force, sometimes presenting as more hesitant ground contact compared to healthy controls. Such a characteristic can be clearly illustrated by VGRF signals as shown in Fig. 4. The kinetic aspect of this hesitant contact is evident in panel (A), which shows that the normalized peak heel force (%BW) is significantly lower in the PD group. Furthermore, this cautious gait strategy alters the overall temporal structure of the gait cycle. As panel (B) demonstrates, the double support time is significantly increased to maximize stability. Concurrently, PD patients often show a shift of plantar load toward the medial aspect or an increase under the midfoot region, possibly as a compensatory mechanism to maintain balance during unstable gait, a phenomenon also visible in Fig. 5. Furthermore, the propulsive phase of gait is typically affected, with a reduction in the peak VGRF generated during push-off (primarily from the forefoot), indicating decreased capacity for forward propulsion. Additionally, the trajectory of the Center-of-Pressure (COP) beneath the foot often displays characteristic abnormalities, including increased variability, reduced path length, or greater medial-lateral excursion, as depicted in Fig. 5. These COP alterations reflect the impaired postural control and increased gait instability often associated with PD gait. Therefore, plantar pressure can provide objective and quantifiable insights into quantify the biomechanical consequences of pathological gait patterns in PD.Fig. 4Comparison between PD and HC in gait-related plantar pressure parameters.a shows the standardized peak heel forces (as a percentage of body weight) for the left and right feet; b illustrates double support time percentages across grouped recordings. The data are sourced from the Gait in PD dataset available on PhysioNet50.Fig. 5Comparison of plantar pressure distribution and COP trajectories between PD and HC.Each sub-image shows the pressure heatmap and COP trajectory for one subject (PD and HC). The data are sourced from the Ga group in the Gait in PD dataset available on PhysioNet50. Comparison between PD and HC in gait-related plantar pressure parameters. a shows the standardized peak heel forces (as a percentage of body weight) for the left and right feet; b illustrates double support time percentages across grouped recordings. The data are sourced from the Gait in PD dataset available on PhysioNet50. Comparison of plantar pressure distribution and COP trajectories between PD and HC. Each sub-image shows the pressure heatmap and COP trajectory for one subject (PD and HC). The data are sourced from the Ga group in the Gait in PD dataset available on PhysioNet50. Plantar pressure data acquisition Different tools are available for analyzing how we walk and how pressure is distributed under our feet. Force plates are considered the gold standard in biomechanics for accurately measuring VGRF and the COP47. These rigid platforms, equipped with underlying sensors (commonly strain gauge or piezoelectric), quantify 3D forces and moments. In PD research, force plates are widely utilized in laboratory settings to evaluate postural stability, dynamic gait features, and gait initiation difficulties, such as AMTI and Kistler. They also serve as a reference standard for validating wearable technologies. Complementing these systems for laboratory or clinical assessment, pressure-sensitive walkways (or mats), like GAITRite or Tekscan HR Mat, feature an extended sensorized surface. This design captures plantar pressure distribution and comprehensive spatiotemporal gait parameters from multiple consecutive footfalls, proving valuable for analyzing continuous walking patterns in individuals with PD. However, both force plates and walkways are generally expensive and primarily confined to laboratory or clinical environments, making them unsuitable for continuous gait monitoring in daily life48. For daily-life and continuous monitoring, insole-based plantar pressure sensors represent a common and convenient alternative for collecting gait data in PD patients49. Typically embedded within thin, flexible insoles designed for comfortable use in everyday footwear, these systems facilitate ambulatory gait monitoring. For instance, the widely utilized PhysioNet PD gait dataset50 was collected using the Ultraflex Computer Dyno Graphy system (Infotronic Inc., Netherlands), a capacitive system measuring VGRF via 8 sensors per foot sampling at 100 Hz. Common sensor technologies in such insoles include piezoresistive (Force Sensitive Resistors, FSR) and capacitive types. Piezoresistive sensors, like the F-Scan™ system (Tekscan, South Boston, MA, USA) or ParoTech system (Paromed, Neubauer, Germany), are generally lower-cost but may offer comparatively lower accuracy and reliability. In contrast, capacitive sensors, such as those employed by the Pedar® system (Novel GmbH, Germany), tend to provide greater accuracy and stability, making them preferred for precise pressure distribution and COP measurements in demanding research applications like PD studies51. Notably, a high-performance plantar pressure data acquisition device is realized by the latest mantis leg-inspired smart insole52 with an innovative closed-loop power supply. It exhibits an ultra-low detection limit of 0.10 Pa, a broad working range up to 1.4 MPa, and exceptional stability over 12,000 loading cycles. Capable of capturing high-precision 16-channel spatiotemporal plantar pressure data—including real-time pressure values, dynamic pressure distribution, COP trajectory, and time-series gait signals—it provides solid and reliable data support for foot arch abnormality detection and pathological gait classification, demonstrating promising application prospects in intelligent gait analysis and clinical diagnosis. In addition, integrated systems that combine plantar pressure sensing with IMUs have become widely adopted, with representative products from manufacturers such as FeetMe® insoles (Paris, France) and Moticon OpenGo (Munich, Germany). The inclusion of IMUs allows for the capture of kinematic information (e.g., segment orientation, acceleration), which complements the kinetic data from pressure sensors to enhance the accuracy of gait event detection and provide a more comprehensive gait analysis49,53. Data derived from these advanced insoles enable effective analysis of spatiotemporal parameters, dynamic pressure patterns, and key PD-related features like FOG54. With the rapid development of computer vision technology, estimating VGRF from videos has become a feasible approach to acquiring VGRF data. This method features low cost and simple equipment requirements, while enabling the synchronous assessment of other motor symptoms from video streams, such as reduced arm swing and abnormal posture47,55–57. However, compared with methods for direct VGRF signal measurement, video-based VGRF estimation still exhibits certain limitations. First, several key abnormal gait characteristics in PD patients are often reflected in fine-grained changes in plantar pressure, which are difficult to fully capture relying solely on video information. In addition, video-based analysis methods generally cannot directly obtain the spatial distribution information of plantar pressure. In the future, it is necessary to further improve the estimation accuracy of vision-based biomechanical parameters, so as to provide more reliable technical support for the early screening and severity evaluation of PD. Plantar pressure dataset High-quality plantar pressure datasets are the foundation for the development and validation of auxiliary diagnostic algorithms for PD. These datasets not only provide the essential raw input for models but, more importantly, capture the fine-grained pathological differences in gait dynamics between PD patients and healthy controls. An ideal dataset should include clearly defined participant information (such as age, gender, UPDRS score, and disease stage), standardized data collection protocols (e.g., specified walking speeds and gait tasks), and a sufficiently large sample size to ensure the effectiveness, stability, and generalizability of the trained models. This section will, therefore, review and evaluate the key plantar pressure datasets commonly used in this field, analyzing their characteristics and limitations to provide a reference for future researchers in the selection, construction, or application of datasets. A detailed summary of the datasets reviewed in this section is presented in Table 1.Table 1Public VGRF datasets for PDDatasetParticipantsCollection frequencySignal typeLinkPhysioNet5093 PD100 HzVGRFhttps://physionet.org/content/gaitpdb/1.0.0/73 HCPDPDS5832 PD100 HzVGRFhttps://doi.org/10.6084/m9.figshare.13530587PDG5926 PD300 HzVGRFhttps://doi.org/10.6084/m9.figshare.14896881.v4Kinematic FeaturesSmart-Insole378 PD100 HzVGRFhttps://bmi.hmu.gr/the-smart-insole-dataset/13 HCKinematic Features9 ElderlyWearGait-PD6061 PD100 HzVGRFhttps://www.synapse.org/Synapse:syn52540892/65 HCKinematic DataThis table summarizes five publicly available datasets containing VGRF and related gait data for PD research. Each entry details the dataset name, participant composition, collection frequency, signal type, and access link. Table 1Public VGRF datasets for PDDatasetParticipantsCollection frequencySignal typeLinkPhysioNet5093 PD100 HzVGRFhttps://physionet.org/content/gaitpdb/1.0.0/73 HCPDPDS5832 PD100 HzVGRFhttps://doi.org/10.6084/m9.figshare.13530587PDG5926 PD300 HzVGRFhttps://doi.org/10.6084/m9.figshare.14896881.v4Kinematic FeaturesSmart-Insole378 PD100 HzVGRFhttps://bmi.hmu.gr/the-smart-insole-dataset/13 HCKinematic Features9 ElderlyWearGait-PD6061 PD100 HzVGRFhttps://www.synapse.org/Synapse:syn52540892/65 HCKinematic DataThis table summarizes five publicly available datasets containing VGRF and related gait data for PD research. Each entry details the dataset name, participant composition, collection frequency, signal type, and access link. Public VGRF datasets for PD This table summarizes five publicly available datasets containing VGRF and related gait data for PD research. Each entry details the dataset name, participant composition, collection frequency, signal type, and access link. The following paragraphs present a comparative overview of several publicly available and self-collected plantar pressure datasets, focusing on their sensor configurations, subject characteristics, and relevance to PD diagnosis and monitoring. The Gait in PD Database50 is a publicly available dataset that provides VGRF recordings from 93 PD patients and 73 healthy controls. Data was collected using 16 in-shoe pressure sensors (8 per foot) at a sampling rate of 100 Hz while participants walked at their usual, self-selected pace for about two minutes on level ground, representing a free walking paradigm under laboratory conditions. In addition, a subset of participants performed a dual-task paradigm (walking while completing serial-7 subtractions), enabling analysis of cognitive-motor interactions. The dataset also includes demographic and medical history information, as well as clinical assessments based on the H&Y staging system and the UPDRS. From these recordings, gait measures such as stride time, swing time, and variability indices can be derived. It has been widely adopted as a benchmark resource for investigating gait variability and disease-related motor impairments in PD. The PDPDS dataset58 provides valuable resources for research on postural control in individuals with PD. It includes static balance data from 32 PD patients assessed in both ON and OFF medication states. The core data consist of GRF collected at 100 Hz using a force platform under four conditions (rigid/unstable surface, eyes open/closed). Processed data includes GRF components, moments, and COP trajectories. Crucially, the dataset is complemented by comprehensive clinical assessments, including UPDRS-II, UPDRS-III, H&Y, New Freezing of Gait Questionnaire (NFoG-Q), Montreal Cognitive Assessment (MoCA), Mini-BESTest, and FES-I scales, along with detailed demographic and medical history information. Being publicly available, this dataset facilitates quantitative analysis of balance impairments in PD, investigation of medication effects on postural stability, and supports the development of improved diagnostic and therapeutic strategies. The PDG dataset59 contains full-body gait data from 26 individuals with idiopathic PD, 13 of whom experience FOG. It provides three-dimensional kinematic data sampled at 150 Hz and kinetic data from ground reaction forces sampled at 300 Hz. The collection paradigm involved participants walking overground at a self-selected comfortable speed under two randomized conditions: an ON state (1 hour after taking dopaminergic medication) and an OFF state (after at least 12 hours without medication). In addition to biomechanical data, the dataset includes comprehensive metadata for each participant, featuring demographic, anthropometric, and clinical assessments using scales such as UPDRS, H&Y, the NFoG-Q, and the MoCA. This dataset is distinguished as the first publicly available resource to offer a three-dimensional, full-body analysis of both kinematics and kinetics in individuals with PD, specifically comparing gait between ON and OFF medication states. While the authors acknowledge the small and heterogeneous sample as a limitation, this resource provides invaluable data for understanding the biomechanical effects of medication on gait. It serves as a critical foundation for enhancing diagnostic methods, monitoring symptoms, and developing more effective rehabilitation and treatment strategies. The “Smart-Insole Dataset”37, is a public resource for gait analysis, containing data from 29 participants across three groups: healthy adults, the elderly, and PD. Data was acquired using Moticon SCIENCE wearable insoles, each equipped with 16 pressure sensors and a 6-axis IMU, sampling at a rate of 100 Hz. The measurement protocol required participants to perform two main tests: the “Walk Straight and Turn” (WST) test at slow, normal, and fast speeds, and a modified “Timed Up and Go” (TUG) test. The dataset provides raw sensor values and features a two-level annotation scheme, labeling both Activities of Daily Living (e.g., walking, turning) and specific gait cycle events (e.g., heel strike, Toe Off). A key feature is the inclusion of clinical evaluations, where a neurologist specialized in movement disorders rated participants on four items of the MDS-Unified PD Rating Scale (MDS-UPDRS) based on video recordings. According to the authors, this is the first public gait dataset to be evaluated by a neurologist, offering a valuable tool for developing computational methods for gait analysis in real-life settings. The WearGait-PD dataset60, is an extensive, open-access resource designed for gait and balance analysis in PD and age-matched controls. It includes multi-modal data from 126 participants (61 PD patients and 65 controls) performing various gait and balance tasks, such as self-paced walking, hurried walking, tandem gait, Timed Up and Go (TUG), and free walking. Data streams, synchronized at 100 Hz, encompass measurements from full-body IMUs, pressure-sensitive insoles, and a pressure-sensing walkway, supplemented by synchronized video for detailed annotations. In addition to raw sensor data, the dataset features extensive clinical information, including MDS-UPDRS scores, and incorporates frame-by-frame annotations of activities and clinical events (like FOG) reviewed by neurologists, ensuring high clinical relevance and reliability. WearGait-PD addresses critical gaps by providing full-body sensor coverage, ground-truth reference data, rich clinical context, and detailed annotations often missing in previous datasets. Consequently, it offers valuable resources for developing and validating computational methods for assessing gait and balance impairments in PD and older adults. While publicly available datasets provide a crucial foundation for research, numerous studies also develop and utilize self-built datasets. However, it is crucial to pay attention to the quality of these public datasets. Currently, there is a lack of standardization among them in terms of collection protocols, annotation methods, and data formats. Future work needs to focus on standardizing these aspects to enhance the comparability and usability of the data. Indeed, several studies utilize such self-constructed datasets for various investigations. For instance, Ji et al.61 designed an integrated device combining pressure sensors and IMUs to collect kinematic data from 25 PD patients and 25 healthy controls; their dataset included spatiotemporal and kinematic gait features derived from this custom setup. In Wahid et al.’s study25, a dataset of self-selected gait data was constructed from 23 PD patients and 26 age-matched healthy controls. Xu et al.62 developed a dataset comprising gait data from 13 patients diagnosed with FOG and 13 patients without FOG, specifically for the purpose of quantifying FOG severity. Marcante et al.63 created a dataset by employing a novel system combining pressure insoles with a 3D accelerometer for FOG detection in 20 PD participants; the data was collected to facilitate FOG detection in both clinical and ecological settings. Despite the specific utility of these self-built datasets, their primary issue is the lack of public availability, which limits broader validation and reproducibility of the research findings. VGRF data preprocessing As a dynamic time-series signal with distinct periodicity, VGRF requires the preservation of complete gait cycles for meaningful analysis. However, its utility is often constrained by insufficient sample size and vulnerability to noise during data acquisition. Therefore, applying appropriate preprocessing methods for specific tasks to enhance data quality is essential for ensuring improved accuracy and reliability in subsequent analysis and model training. This section summarizes several preprocessing techniques adopted in recent studies based on the general framework shown in Fig. 6.Fig. 6Preprocessing methodology for VGRF data.This figure outlines the core stages and corresponding techniques in VGRF data preprocessing. It is structured into five key modules. Preprocessing methodology for VGRF data. This figure outlines the core stages and corresponding techniques in VGRF data preprocessing. It is structured into five key modules. Noise Suppression and Normalization are two routine steps to ensure data quality. The former mitigates discrete and continuous noise to enhance signal stability, while the latter normalizes data scales to eliminate the influence of the non-pathological factors (e.g., body weight and height), thereby improving cross-subject comparability. As these techniques are relatively mature and have been widely adopted, this review will not elaborate on them in detail but instead focus on how data augmentation can address the small-sample limitations inherent in VGRF data. Due to the collection methods and paradigms of VGRF data being constrained by signal specificity, there are inherent limitations that directly result in VGRF data exhibiting significant small-sample characteristics. This limitation leads to two critical issues: overfitting of the model and weakening of its generalization capability. Signal Segmentation is an important step in VGRF analysis. By dividing long time series into shorter segments via time-windowing or gait-cycle-based partitioning, multiple sub-samples can be generated, thereby alleviating the problem of limited sample size to some extent. However, segmentation alone is insufficient to overcome the inherent “small-sample” nature of VGRF. Researchers have introduced more diverse data augmentation techniques, which can be broadly classified into three groups: (i) generative approaches. Eguchi et al.35 proposed a probabilistic statistics-based approach that treats single-step measurements as mean vectors while extracting gait variability from other users’ data to form a covariance matrix. Multiple virtual gait samples are generated at each time point using a multivariate normal distribution. Li et al.64 adopted an augmentation strategy based on Conditional Generative Adversarial Networks (cGAN) to generate synthetic samples consistent with the original data distribution, thereby balancing the class sizes. (ii) reconstructive approaches. Window Slicing augments local features and time-scale variations by extracting fixed-length subsequences65–67. Permutation exchanges sequence segments by randomly selecting a pivot point, thereby exploiting the periodicity of the gait cycle to simulate anomalous gait patterns. (iii) methods for class imbalance. Some studies employed the Synthetic Minority Oversampling Technique (SMOTE) to expand healthy samples and balance the binary classes to address the specific problem of class imbalance in original datasets (e.g., more patients than healthy controls and uneven distribution across disease stages)68–70. Feature extraction and selection In PD diagnosis, deep learning approaches excel at automatically extracting key features and achieving high accuracy. However, these approaches often fail to establish explicit connections between gait features and pathological mechanisms, thereby limiting the ability to provide interpretable and traceable diagnostic evidence. To address this issue, the following content analyzes gait signals from a feature engineering perspective, focusing on the extraction of physiologically defined indicators to enhance clinical interpretability. We summarize the types of VGRF features, extraction methods, and selection strategies (see Table 2). The basic framework of feature engineering for PD diagnosis is shown in Fig. 7.Fig. 7Feature engineering framework designed for PD diagnosis.This figure presents a hierarchical feature engineering framework tailored for PD diagnosis and assessment, encompassing three core components: gait features, feature extraction, and feature selection. Gait features are categorized into four types, including anthropometric, spatiotemporal, kinetic, and combined features. Feature extraction adopts two strategies: hand-crafted extraction and automatic extraction. Feature selection employs three categories of filter, wrapper, and embedded methods to identify the most discriminative attributes.Table 2Feature extraction and selection methods for PD diagnosisFeature classYear/refFeaturesFeature Ext/select methodSpatio-temporal201996Mean in duration(s) of the stride time, mean of the long swing time, mean of the gait asymmetry, coefficients of variation (CV) in percentage(%) of swing time of the left foot, etc.Filter202192Time domain and frequency domain joint features of left and right gait.CNN Bi-LSTM202284Multi-level force points transfer information about gait.Optical Flow Method PCA2025125Stride time, stance time, swing time, double support time, SI(Symmetry index)stance time, SIswing time, CoVstride time, etc.Statistical computation formulaKinetic2012103Statistical features: mean, median of coefficients, energy, standard deviation(SD), skewness, kurtosis, etc.WT201383The mean frequency moments, mean frequency, and variance of frequency from each time series force differences.STFT2019137Entropy, energy, variance, SD and waveform length.WPT(Wavelet Packet Transform)202134Statistical features: mean, standard deviation(SD), energy, skewness, kurtosis, log entropy, etc.SWLNGP202194Force domain: average value, SD, mean, etc. Peak domain: the count and interval time, SD of peak data. Abnormality domain: ITabnormal, PVabnormal.ANOVA202366Frequency domain representations.FFT RFdGAD2025100Statistical domain: mean, median, std, variance, skewness, kurtosis. Frequency domain: mean, median, bandpower. Entropy features: wavelet energy, Shannon wavelet entropy, Rényi wavelet entropy, Tsallis wavelet entropy.FFT DWT BGWO BWOA BDASpatio-temporal kinetic202082Time domain: number, duration, length of COP reversals, CV of COP, number of weight shifts, etc. FFT features: total power in the FFT signal, dominant frequency, max, min, and mean. DWT features: variances of coefficients, max, min, mean energy, etc.FFT DWT MRMR202154Time domain: number, duration, length of COP reversals, CV of COP position, velocity, acceleration, number of weight shifts. FFT features: total power in FFT signal, dominant frequency, freeze index, power in locomotion, freeze bands, etc. DWT features: variances of coefficients, max, min, and mean energy.FFT DWT MRMR Relief-F202177Time-frequency spectrogram.CWT STFT CNN WSST PCA202295Frequency domain: Freezing Index(FI) energy. Time domain: mean, absolute mean, zero-crossing rate, SD, range, RMS. Statistical domain: sum power, maximum, minimum, principal direction eigenvalue, entropy.ANOVA MI (Mutual Information)202568Time domain: stance-duration, stride-domain, swing-domain, step-time. Spatial domain: stride-length, step-length. Temporal: swing-to-stance ratio, std stance ratio, standardized swing ratio, heel-strike force, toe-of-force. Frequency domain: cadence.RFEKinematic kinetic202063Fraction of the total force, COP trajectory, dominant frequency of COP curve, and range of vertical axis acceleration.FFTSpatio-temporal kinetic kinematic202279Dominant FFT frequency of COP velocity in ML/AP directions, dominant FFT frequency of thigh accelerometers in AP direction, and mean energy of WT approximation coefficient of COP position in AP direction.DWT FFT Relief-F202461The mean frequency moments, mean frequency, and variance of frequency from each time series force differences.Preset algorithmThis table summarizes feature extraction and selection methods reported in gait-based PD diagnosis and assessment studies, grouped by feature class: Spatiotemporal, Kinetic, Spatiotemporal Kinetic, Kinematic Kinetic, and Spatiotemporal Kinetic Kinematic. (Ref Reference, Feature Ext/Select Method Feature Extraction/Selection Method). Feature engineering framework designed for PD diagnosis. This figure presents a hierarchical feature engineering framework tailored for PD diagnosis and assessment, encompassing three core components: gait features, feature extraction, and feature selection. Gait features are categorized into four types, including anthropometric, spatiotemporal, kinetic, and combined features. Feature extraction adopts two strategies: hand-crafted extraction and automatic extraction. Feature selection employs three categories of filter, wrapper, and embedded methods to identify the most discriminative attributes. Table 2Feature extraction and selection methods for PD diagnosisFeature classYear/refFeaturesFeature Ext/select methodSpatio-temporal201996Mean in duration(s) of the stride time, mean of the long swing time, mean of the gait asymmetry, coefficients of variation (CV) in percentage(%) of swing time of the left foot, etc.Filter202192Time domain and frequency domain joint features of left and right gait.CNN Bi-LSTM202284Multi-level force points transfer information about gait.Optical Flow Method PCA2025125Stride time, stance time, swing time, double support time, SI(Symmetry index)stance time, SIswing time, CoVstride time, etc.Statistical computation formulaKinetic2012103Statistical features: mean, median of coefficients, energy, standard deviation(SD), skewness, kurtosis, etc.WT201383The mean frequency moments, mean frequency, and variance of frequency from each time series force differences.STFT2019137Entropy, energy, variance, SD and waveform length.WPT(Wavelet Packet Transform)202134Statistical features: mean, standard deviation(SD), energy, skewness, kurtosis, log entropy, etc.SWLNGP202194Force domain: average value, SD, mean, etc. Peak domain: the count and interval time, SD of peak data. Abnormality domain: ITabnormal, PVabnormal.ANOVA202366Frequency domain representations.FFT RFdGAD2025100Statistical domain: mean, median, std, variance, skewness, kurtosis. Frequency domain: mean, median, bandpower. Entropy features: wavelet energy, Shannon wavelet entropy, Rényi wavelet entropy, Tsallis wavelet entropy.FFT DWT BGWO BWOA BDASpatio-temporal kinetic202082Time domain: number, duration, length of COP reversals, CV of COP, number of weight shifts, etc. FFT features: total power in the FFT signal, dominant frequency, max, min, and mean. DWT features: variances of coefficients, max, min, mean energy, etc.FFT DWT MRMR202154Time domain: number, duration, length of COP reversals, CV of COP position, velocity, acceleration, number of weight shifts. FFT features: total power in FFT signal, dominant frequency, freeze index, power in locomotion, freeze bands, etc. DWT features: variances of coefficients, max, min, and mean energy.FFT DWT MRMR Relief-F202177Time-frequency spectrogram.CWT STFT CNN WSST PCA202295Frequency domain: Freezing Index(FI) energy. Time domain: mean, absolute mean, zero-crossing rate, SD, range, RMS. Statistical domain: sum power, maximum, minimum, principal direction eigenvalue, entropy.ANOVA MI (Mutual Information)202568Time domain: stance-duration, stride-domain, swing-domain, step-time. Spatial domain: stride-length, step-length. Temporal: swing-to-stance ratio, std stance ratio, standardized swing ratio, heel-strike force, toe-of-force. Frequency domain: cadence.RFEKinematic kinetic202063Fraction of the total force, COP trajectory, dominant frequency of COP curve, and range of vertical axis acceleration.FFTSpatio-temporal kinetic kinematic202279Dominant FFT frequency of COP velocity in ML/AP directions, dominant FFT frequency of thigh accelerometers in AP direction, and mean energy of WT approximation coefficient of COP position in AP direction.DWT FFT Relief-F202461The mean frequency moments, mean frequency, and variance of frequency from each time series force differences.Preset algorithmThis table summarizes feature extraction and selection methods reported in gait-based PD diagnosis and assessment studies, grouped by feature class: Spatiotemporal, Kinetic, Spatiotemporal Kinetic, Kinematic Kinetic, and Spatiotemporal Kinetic Kinematic. (Ref Reference, Feature Ext/Select Method Feature Extraction/Selection Method). Feature extraction and selection methods for PD diagnosis This table summarizes feature extraction and selection methods reported in gait-based PD diagnosis and assessment studies, grouped by feature class: Spatiotemporal, Kinetic, Spatiotemporal Kinetic, Kinematic Kinetic, and Spatiotemporal Kinetic Kinematic. (Ref Reference, Feature Ext/Select Method Feature Extraction/Selection Method). VGRF-based gait features Gait serves as a critical biomechanical indicator of human coordination and balance, offering valuable insights into an individual’s health status. In PD studies using VGRF data, gait features are commonly classified into four categories: anthropometric, spatiotemporal, kinetic, and combined features. The subsequent content briefly outlines these categories and their application in PD diagnosis. Anthropometric features (e.g., age, sex) reflect core physiological characteristics of individuals. Many studies have employed normalization to reduce the influence of these variables on gait analysis71. Zhao et al.24,72,73 standardized VGRF data by body weight (e.g., VGRF/BW) to minimize weight-related effects. Other studies normalized gait features by height or constructed multivariate regression models incorporating speed, age, and sex to decouple physiological factors from gait measurements25,74–76. To account for both height and weight, Farashi et al.26 normalized VGRF data using the body mass index (BMI). Such strategies effectively reduce individual physiological variability, enhancing the generalizability and comparability of gait features. Spatiotemporal features describe gait parameters related to temporal and spatial variations, and have been shown to exhibit strong correlations with disease progression. Temporal features focus on the duration of each phase within the gait cycle, capturing the timing pattern of walking. Spatial features describe step-related parameters such as stride length, foot placement, and gait symmetry. Various statistical methods have been used across studies to compute these features. Common spatiotemporal features are summarized in Table 3.Table 3Basic description of seven spatiotemporal-related featuresFeature nameCalculation formulaDescriptionCadence/The walking rate (normal: 90–110 steps/min) is determined by counting the number of heel strike peaks within a 60-second period.Stride timeICright/left(i + 1) − ICright/left(i)The time interval between two consecutive heel strikes of the same foot, IC (Initial Contact), refers to the moment of initial contact in the gait cycle.Step timeICright(i) − ICleft(i)The time interval between consecutive heel strikes of opposite feet.Stride lengthWalking Speed ⋅ Stride TimeThe distance covered from one heel strike to the next heel strike of the same foot.Double-leg stance timeTOleft/right(i) − ICright/left(i)The duration of the phase during which both feet are simultaneously in contact with the ground within a single gait cycle. TO (Toe Off) denotes the moment when the foot leaves the ground.Stride regularity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1-\frac{| {\rm{Stride}}\,{{\rm{Time}}}_{{\rm{right}}}(i)-{{\rm{Stride}}{\rm{Time}}}_{{\rm{left}}}(i)| }{\max ({\rm{Stride}}\,{{\rm{Time}}}_{{\rm{right}}}(i),{\rm{Stride}}\,{{\rm{Time}}}_{{\rm{left}}}(i))}$$\end{document}1- ∣StrideTimeright(i)-StrideTimeleft(i)∣max(StrideTimeright(i),StrideTimeleft(i))The similarity between consecutive strides.Step regularity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1-\frac{| \,{\rm{Step}}\,{{\rm{Time}}}_{{\rm{right}}}(i)-{\rm{Step}}\,{{\rm{Time}}}_{{\rm{left}}}(i)\,| }{\max ({\rm{Step}}\,{{\rm{Time}}}_{{\rm{right}}}(i),{\rm{Step}}\,{{\rm{Time}}}_{{\rm{left}}}(i))}$$\end{document}1-∣StepTimeright(i)-StepTimeleft(i)∣max(StepTimeright(i),StepTimeleft(i))The similarity between consecutive steps.Symmetry\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{{\rm{Average}}}_{{\rm{left}}}}{{{\rm{Average}}}_{{\rm{right}}}}$$\end{document}AverageleftAveragerightThe comparison of gait parameters between left and right feet for assessing imbalance and symmetry determination.This table defines eight key spatiotemporal gait features, presenting their calculation formulas and functional descriptions. The features include cadence, stride time, step time, stride length, double-leg stance time, stride regularity, step regularity, and symmetry. Table 3Basic description of seven spatiotemporal-related featuresFeature nameCalculation formulaDescriptionCadence/The walking rate (normal: 90–110 steps/min) is determined by counting the number of heel strike peaks within a 60-second period.Stride timeICright/left(i + 1) − ICright/left(i)The time interval between two consecutive heel strikes of the same foot, IC (Initial Contact), refers to the moment of initial contact in the gait cycle.Step timeICright(i) − ICleft(i)The time interval between consecutive heel strikes of opposite feet.Stride lengthWalking Speed ⋅ Stride TimeThe distance covered from one heel strike to the next heel strike of the same foot.Double-leg stance timeTOleft/right(i) − ICright/left(i)The duration of the phase during which both feet are simultaneously in contact with the ground within a single gait cycle. TO (Toe Off) denotes the moment when the foot leaves the ground.Stride regularity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1-\frac{| {\rm{Stride}}\,{{\rm{Time}}}_{{\rm{right}}}(i)-{{\rm{Stride}}{\rm{Time}}}_{{\rm{left}}}(i)| }{\max ({\rm{Stride}}\,{{\rm{Time}}}_{{\rm{right}}}(i),{\rm{Stride}}\,{{\rm{Time}}}_{{\rm{left}}}(i))}$$\end{document}1- ∣StrideTimeright(i)-StrideTimeleft(i)∣max(StrideTimeright(i),StrideTimeleft(i))The similarity between consecutive strides.Step regularity\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1-\frac{| \,{\rm{Step}}\,{{\rm{Time}}}_{{\rm{right}}}(i)-{\rm{Step}}\,{{\rm{Time}}}_{{\rm{left}}}(i)\,| }{\max ({\rm{Step}}\,{{\rm{Time}}}_{{\rm{right}}}(i),{\rm{Step}}\,{{\rm{Time}}}_{{\rm{left}}}(i))}$$\end{document}1-∣StepTimeright(i)-StepTimeleft(i)∣max(StepTimeright(i),StepTimeleft(i))The similarity between consecutive steps.Symmetry\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{{\rm{Average}}}_{{\rm{left}}}}{{{\rm{Average}}}_{{\rm{right}}}}$$\end{document}AverageleftAveragerightThe comparison of gait parameters between left and right feet for assessing imbalance and symmetry determination.This table defines eight key spatiotemporal gait features, presenting their calculation formulas and functional descriptions. The features include cadence, stride time, step time, stride length, double-leg stance time, stride regularity, step regularity, and symmetry. Basic description of seven spatiotemporal-related features This table defines eight key spatiotemporal gait features, presenting their calculation formulas and functional descriptions. The features include cadence, stride time, step time, stride length, double-leg stance time, stride regularity, step regularity, and symmetry. Force-domain features are pivotal in gait analysis, capturing GRF dynamic variations throughout the gait cycle. VGRF signals are key for quantifying plantar load distribution and evaluating postural stability. In addition, frequency-domain features are also critical in gait analysis. Through spectral analysis techniques, such as Fast Fourier Transform (FFT) or Wavelet Transform (WT) to VGRF signals, key features including power spectral density, mean and median frequency, and energy distribution can be extracted to reveal the rhythmicity and steadiness of gait. Moreover, time-frequency features, integrating time and frequency analysis, comprehensively characterize gait dynamics. Previous research77 shows these features differ between healthy and pathological gaits, making them potential clinical biomarkers for PD diagnosis. Beyond single-domain analysis, researchers have increasingly turned to adopt multi-feature fusion strategies to better capture gait abnormalities and improve model performance63,78,79. Combining features across domains helps address the limitations of unimodal features. Vimalajeewa et al.78 fused time-domain features with frequency-domain features, achieving better classification with fewer features than single-domain approaches. Integrating features from different data sources has shown practical value. Marcante et al.63 combined plantar pressure, COP trajectories, 3D acceleration, and frequency features to achieve stable FOG detection in both real-world and clinical tasks. Cross-modal feature fusion has become an increasingly studied strategy. Some studies jointly encode gait pressure and visual signals8, employing multimodal graph networks to improve FOG detection. Feature extraction methods Feature extraction is the process of identifying and extracting key patterns or features from preprocessed data that are crucial for the target task, enabling subsequent classification models to effectively utilize these features for better predictive accuracy. Based on different technical backgrounds, researchers typically employ the following approaches to extract relevant gait features. Feature transformation methods Frequency-domain transformation helps identify abnormal gait rhythmicity and periodicity in PD patients. The basic Discrete Fourier Transform (DFT) decomposes time-domain signals into combinations of sine waves. Due to its high computational cost, studies often employ the more efficient FFT73,80–82 to map VGRF signals into the frequency domain to analyze the main frequency components. FFT can effectively capture the frequency features of gait, but it fails to capture dynamic changes in frequency over time, limiting its applicability to non-stationary signals. To address this, the Short-Time Fourier Transform (STFT) applies Fourier analysis within a sliding window, producing a time-frequency representation suitable for detecting local rhythmic changes in VGRF signals76,83. To enhance time-frequency resolution, Continuous Wavelet Transform (CWT) enables detection of frequency variation patterns in VGRF signals at multiple temporal scales. Setiawan et al.81 showed that CWT spectrograms effectively distinguish gait differences between PD patients at different stages and healthy controls, with marked irregular fluctuations in moderate-to-advanced cases. Discrete Wavelet Transform (DWT) efficiently extracts structured VGRF features (e.g., energy, variance, and extrema of wavelet coefficients). Pardoel et al.80 verified that DWT-derived features substantially improve FOG detection accuracy. To address the limitations of traditional wavelet methods in energy concentration, the Wavelet Synchrosqueezed Transform (WSST) enhances the resolution of time-frequency representations. The analysis in Setiawan et al.77 revealed that PD patients exhibit distinct abnormalities in low-force amplitude patterns within the 2.5–5 Hz frequency range, and that WSST can reduce frequency blurring by using phase information, making key frequency components more prominent. Nevertheless, its classification performance degrades with long time windows. Specialized spatial feature extraction methods The Directional Histogram (DH) encodes the displacement of force points between frames, which are represented as optical flow vectors, into an eight-dimensional feature for each point. The approach can yield a 16 × 8-dimensional vector, which captures gait motion patterns84. The Local Binary Pattern (LBP) encodes textural information by comparing the intensity of a center pixel with its neighbors. However, its sensitivity to random noise and fixed-weight encoding limits its ability to capture subtle gradient variations in PD gait data. Priya et al.34 proposed the Symmetric Weighted Local Neighborhood Gradient Pattern (SWLNGP), which integrated symmetric weighting and neighborhood gradient strategies to extract robust local gradient features from gait signals. The Phase Space Reconstruction (PSR)85–89 effectively preserves the nonlinear dynamics of gait signals by extracting features such as Euclidean Distance (ED), Shannon Energy Envelope (SEE), and Coefficient of Variation (CV) to characterize abnormal gait patterns of PD patients. It captures complex nonlinear features that traditional methods often miss and shows strong robustness to individual variability and noise. However, the extracted features lack intuitive interpretability, making it difficult for clinicians to grasp their physiological relevance. Linear projection methods Principal component analysis (PCA)77,81,90 can capture the intrinsic structure of the dataset and extract the most representative features without requiring labeled data. Because it does not consider class information, PCA introduces components irrelevant to classification tasks, limiting its effectiveness. Dong et al.84 reported that applying PCA led to a 1.7% improvement in PD classification accuracy, suggesting that the original spatiotemporal and optical flow features contained redundant information, whose negative effects outweighed the loss of information due to dimensionality reduction. Deep learning methods Convolutional neural networks (CNNs) extract local spatial variation patterns through convolution and pooling operations, enabling the learning of high-level features from plantar pressure maps in PD diagnosis and FOG prediction. The hierarchical structure of CNNs enables them to capture abnormal pressure distribution patterns, with translation invariance allowing for strong generalization across subjects. Long short-term memory (LSTM) networks are well-suited for modeling VGRF time series by capturing temporal information and long-term dependencies through a gated mechanism. The structure of forget, input, and output gates allows LSTM to selectively retain key features within gait cycles, effectively addressing the vanishing gradient problem in traditional RNNs. Temporal convolutional networks (TCNs) extract both short- and long-term temporal patterns through temporal and dilated convolutions, enabling multi-scale feature learning from gait data and significantly improving FOG detection accuracy. Compared to LSTMs, TCNs exhibit advantages in parallel computation and receptive field control, allowing efficient modeling of temporal dependencies across scales. Transformer leverage self-attention mechanisms to model long-range dependencies in VGRF time series, using multi-head attention to capture temporal feature relationships across different time steps in VGRF signals. The architecture offers strong representational power for capturing temporal pattern variations in gait cycles, especially in FOG prediction. Unlike RNN-based models, Transformer’s parallel processing overcomes sequential computation bottlenecks, preserving the temporal order of gait data using its positional encoding mechanism. By constructing a topological graph of plantar sensor distribution, graph neural networks (GNNs)8,65,67,91 learn PD-related features via node and edge representations for accurate classification. Nodes correspond to sensor positions, and edges represent spatial correlations and pressure transfer. Graph convolutional networks (GCNs)66 and graph attention networks (GATs) effectively capture spatial pressure dynamics, well-suited for analyzing COP shifts and abnormal loading patterns in PD. These deep learning architectures offer complementary analytical perspectives, extracting different features of VGRF data. Recent studies have also explored hybrid models, such as CNN-LSTM92, spatiotemporal graph convolution networks66, and CNN-LSTM-DNN hybrid models93, further enhancing the performance of diagnostic systems. Figure 8 depicts the distribution of various feature extraction methods in PD diagnosis. Automated extraction methods account for the largest proportion (65%), with CNNs only reaching 25%, presenting their advantages in processing gait images and time-frequency data. Feature transformation (27%) methods are also important for their efficiency in capturing frequency domain features. Additional specialized methods like PSR (5%) also have their own strengths and can be useful in specific situations.Fig. 8Proportional distribution of feature extraction methods.This pie chart illustrates the relative prevalence of different feature extraction methods in VGRF-based PD gait analysis, which are grouped into three broad categories: automated extraction, feature transformation, and specialized methods. Proportional distribution of feature extraction methods. This pie chart illustrates the relative prevalence of different feature extraction methods in VGRF-based PD gait analysis, which are grouped into three broad categories: automated extraction, feature transformation, and specialized methods. Feature selection methods As data dimensionality increases, processing, storage, and computation become increasingly complex, potentially leading to the so-called “curse of dimensionality.” Feature selection alleviates the phenomenon by reducing redundant dimensions and extracting the most informative features, thereby enhancing the efficiency of models. The following methods are used to select and evaluate extracted features, removing those with low relevance. Filter methods Filter methods54,76,94,95 evaluate individual features based on relevance or divergence with respect to the target variable, deciding whether to filter out variables according to their own characteristics or relationships with other features, without relying on other machine learning models. ANOVA94,95 evaluates feature discriminability via F-value and P-value based on variance analysis. The MRMR filter method54 selects informative features by balancing feature-target correlation and reducing feature redundancy using mutual information. Wrapper methods Wrapper Methods typically use the objective function as a black box to assist feature selection, often involving multiple rounds of training on feature subsets. Important features are iteratively pruned through recursive procedures, which results in high computational cost96–98,99. Heuristic search is a stepwise feature selection approach that adds or removes features gradually to identify the optimal subset. Sequential Forward Selection (SFS) [99]starts with an empty feature set and iteratively removes and adds features based on model classification performance to determine the optimal subset. Recursive feature elimination (RFE)68,97 starts with the full set of candidate features and iteratively eliminates the least important ones based on feature importance scores, until a predefined number of features is reached or other stopping criterion is met. For PD severity classification, Agrawal et al.100 proposed an ensemble feature selection framework integrating three meta-heuristic algorithms, which reduces feature dimensionality and boosts subsequent classifier performance. Furthermore, Li et al.64 employed a deep learning architecture for end-to-end feature selection, automatically identifying key features related to PD pathological mechanisms to provide high-quality inputs for H&Y stage prediction. Embedded methods Unlike the aforementioned methods, embedded methods integrate feature selection into model training by optimizing the objective function. Lasso101 employs L1 regularization to drive regression coefficients of irrelevant features to zero, automatically identifying critical features. Although such methods balance accuracy and feature relevance, they are model-dependent and thus limited in generalizability. We conducted a data analysis of the collected literature on VGRF-assisted diagnosis. The results show that filter-based methods are the most commonly used due to their simplicity and efficiency, with the Relief-F method being the most frequently applied for evaluating feature importance. In contrast, wrapper-based and embedded methods are used less frequently but exhibit significant advantages in specific tasks. Advanced methods for PD diagnosis and severity assessment via VGRF In recent years, the use of plantar pressure measurement systems for the diagnosis and severity assessment of PD has garnered considerable attention. Figure 9 illustrates the temporal publication trend of studies on PD research utilizing gait data, as published in high-impact international journals or conferences between 2014 and October 2025. This rapid growth can be largely attributed to the emergence and advancement of deep learning technologies, which have demonstrated exceptional capacity in feature extraction and representation learning, thereby enabling novel approaches for analyzing complex patterns within plantar pressure signals. We focused on analyzing a total of 102 relevant studies using VGRF signals for PD diagnosis and severity assessment.Fig. 9Annual trends in PD research using VGRF.This line chart illustrates the annual number of peer-reviewed articles focused on PD research VGRF data, spanning from 2014 to 2025. The x axis denotes the publication year, and the y axis represents the count of published articles. Annual trends in PD research using VGRF. This line chart illustrates the annual number of peer-reviewed articles focused on PD research VGRF data, spanning from 2014 to 2025. The x axis denotes the publication year, and the y axis represents the count of published articles. Research in the field of PD primarily focuses on two objectives: diagnosis and severity assessment. Studies addressing these objectives typically employ three technological paths: statistical analysis23,102–105, machine learning25,88,89,96,106 and deep learning14,24,65,66,81. Accordingly, the content that follows is organized into three parts. Section PD diagnosis and Section PD severity assessment systematically review the technical development progress from early statistical analysis to the mainstream models, in terms of PD diagnosis and severity assessment. The section Algorithm performance evaluation focuses on summarizing and analyzing the performance of various commonly used algorithms, while discussing their limitations to provide references for subsequent research. The general framework of the following content is illustrated in Fig. 10 to provide readers with a clear structural overview.Fig. 10Framework of section methods.This schematic diagram outlines the methodological framework for Section Methods, which focuses on VGRF-based PD research. The work is divided into two core tasks: PD diagnosis and PD severity assessment, both built on three analytical categories: statistical analysis, machine learning, and deep learning. Framework of section methods. This schematic diagram outlines the methodological framework for Section Methods, which focuses on VGRF-based PD research. The work is divided into two core tasks: PD diagnosis and PD severity assessment, both built on three analytical categories: statistical analysis, machine learning, and deep learning. Meanwhile, Table 4 primarily summarizes studies that employed the PhysioNet dataset and groups the references according to their validation methods. Similarly, Table 5 presents studies based on other datasets, also organized by the same validation criteria for systematic comparison. These tables encapsulate essential details, including publication year, research objectives, datasets used, core algorithmic approaches, validation methodologies, and performance metrics. This structured presentation facilitates a holistic understanding of the current research landscape.Table 4Summary of studies based on the PhysioNet PD gait dataset50RefYrObjBackbonePerf.DSHold-out732018✓SVMAcc = 0.948, Sen = 0.970, Spe = 0.8701372019✓SVMAcc = 0.903782023✓SVMAcc = 0.974, Sen = 0.990, Spe = 0.950862023✓DTAcc = 0.880, Sen = 0.945, Spe = 0.805852021✓Ensemble DTAcc = 0.986, Sen = 0.996, Spe = 0.9951382025✓RFAcc = 0.960, Sen = 0.9601132022✓KNNAcc = 0.76722022✓Ensemble KNNAcc = 0.9501002025✓✓Weighted VotingAcc = 0.985, Sen = 0.9861392024✓LRAcc = 0.817, Sen = 0.897, Spe = 0.6301282020✓✓ANNAcc = 0.972, Sen = 0.9701272020✓✓ANN(D) Acc = 0.974, Sen = 0.977, Spe = 0.970 (S) Acc = 0.865292018✓✓CNN(D) Acc = 0.986182019✓✓CNNAcc = 0.887922021✓CNNAcc = 0.992, Sen = 1.000, Spe = 0.9801302022✓✓CNNAcc = 0.983, Sen = 0.982, Spe = 0.976942023✓CNNAcc = 0.9841322024✓✓CNNAcc = 0.985, Sen = 0.964, Spe = 0.9881402024✓CNN-XGBoostAcc = 0.9770 (Ga), 0.981 (Ju), 0.953 (Si)1412025✓CNNAcc = 0.9341422025✓CNNAcc = 0.990, Sen = 1.000, Spe = 0.9901432025✓CNNAcc = 0.991, Sen = 1.000, Spe = 1.0001092020✓LSTMAcc = 0.920, Sen = 0.9671182021✓✓LSTM(D) Acc = 0.986 (S) Acc = 0.966, Sen = 0.962, Spe = 0.9811442025✓LSTMAcc = 0.977, Sen = 0.990, Spe = 0.9601332023✓TransformerAcc = 0.988, Sen = 0.988912025✓GNNAcc = 0.995, Sen = 0.9971452025✓GRUAcc = 0.875, Sen = 0.875, Spe = 0.875LOOCV1102017✓SVMAcc = 1.000, Sen = 1.000, Spe = 1.000962019✓SVMAcc = 0.903, Sen = 0.902742022✓✓DT(D) Acc = 0.9931062023✓KNNAcc = 0.876, Sen = 0.851, Spe = 0.8891462016✓ANNAcc = 0.933, Sen = 0.924, Spe = 0.945892019✓ANNAcc = 0.988, Sen = 0.989, Spe = 0.9861272020✓ANN(D) Acc = 0.949, (S) Acc = 0.8765-fold1032012✓Statistical AnalysisAcc = 0.7731022016✓Statistical AnalysisAcc = 0.869, Sen = 0.720, Spe = 0.810842023✓✓DT(D) Acc = 0.933, Sen = 0.946, Spe = 0.917 (S) Acc = 0.914, Sen = 0.855, Spe = 0.9541462016✓ANNAcc = 0.963, Sen = 0.967, Spe = 0.9581272020✓✓ANN(D) Acc = 0.909, (S) Acc = 0.7301012021✓✓CNN+ SVM/KNN(D, KNN) Acc = 0.990, Sen = 0.990, Spe = 0.980 (S, SVM) Acc = 0.980, Sen = 0.940, Spe = 0.980872023✓✓CNN(D) Acc = 0.990, Sen = 0.995, Spe = 0.987 (S) Acc = 0.980, Sen = 0.987, Spe = 0.992RefYrObjBackbonePerf.DSFivefold1472025✓✓LSTM(D) Acc = 0.994, Sen = 0.994 (S) Acc = 0.982, Sen = 0.982662022✓GNNAcc = 0.938652023✓GNNAcc = 0.9591212025✓GNNAcc = 0.9991202019✓✓CNN+LSTM(D) Acc = 0.990, Sen = 0.991, Spe = 0.990 (S) Acc = 0.9801472025✓✓CNN+LSTM(D,LSTM) Acc = 0.994, Sen = 0.994 (S,CNN) Acc = 0.966, Sen = 0.96610-fold1482021✓✓SVM752021✓✓SVMAcc = 0.987, Sen = 0.976, Spe = 0.991762021✓✓SVMAcc = 0.984, Sen = 0.966, Spe = 0.996982022✓✓SVM(D) Acc = 0.985, (S) Acc = 0.977882024✓✓SVM(D) Acc = 0.988, Sen = 0.989, Spe = 0.986 (S) Acc = 0.966, Sen = 0.933, Spe = 0.977262020✓✓DTAcc = 1.000, Sen = 1.000, Spe = 1.0001242020✓✓DTAcc = 0.994, Sen = 0.986, Spe = 0.978712024✓DTAcc = 0.941, Sen = 0.942, Spe = 0.9831112017✓RFAcc = 0.980, Sen = 0.991, Spe = 0.957362018✓✓RF(D) Acc = 0.990, Sen = 0.978, Spe = 0.9951262023✓RFAcc = 0.990, Sen = 0.970, Spe = 0.9901492021✓RSRNAAcc = 0.810742022✓✓DT+ER(D) Acc = 0.999682025✓✓RFT+ER(D) Acc = 0.975, Sen = 0.970, Spe = 0.950 (S) Acc = 0.964892019✓ANNAcc = 0.969, Sen = 0.967, Spe = 0.9721142020✓ANNAcc = 0.983, Sen = 0.986, Spe = 0.9751272020✓✓ANN(D) Acc = 0.879, (S) Acc = 0.760342021✓ANNAcc = 0.963, Sen = 0.966, Spe = 0.959142020✓✓CNN(D) Acc = 0.987, Sen = 0.981, Spe = 1.000 (S) Acc = 0.853152020✓CNNAcc = 0.955812021✓✓CNN(D) Acc = 0.966, Sen = 0.944, Spe = 0.976 (S) Acc = 0.945, Sen = 0.920, Spe = 0.9561012021✓✓CNN(D) Acc = 0.990, Sen = 0.990, Spe = 0.980 (S) Acc = 0.980, Sen = 0.940, Spe = 0.9801152023✓✓CNN(D) Acc = 0.967, Sen = 0.981, Spe = 0.933 (S) Acc = 0.923, Sen = 0.9221502025✓✓CNN(D) Acc = 0.986, Sen = 0.992, Spe = 0.9851192024✓✓LSTM(D) Acc = 0.992, Sen = 0.991, Spe = 0.997 (S) Acc = 0.923, Sen = 0.924, Spe = 0.9771342024✓✓Transformer(D) Acc = 0.973, Sen = 0.960, Spe = 1.0001162024✓TransformerAcc = 0.878, Sen = 0.8801512025✓TransformerAcc = 0.923, Sen = 0.990, Spe = 0.8871522025✓TransformerAcc = 0.908, Sen = 0.910662022✓GNNAcc = 0.959652023✓GNNAcc = 0.9601212025✓GNNAcc = 0.9991532016✓DNN(Ga) Acc = 0.914, Sen = 0.770, Spe = 0.550 (Si) Acc = 0.921, Sen = 0.940, Spe = 0.890 (Ju) Acc = 0.909, Sen = 0.740, Spe = 0.8301292020✓DNN(D) Acc = 0.994(S) Acc = 0.991RefYrObjBackbonePerf.DS292018✓✓CNN+LSTM(D) Acc = 0.986932024✓CNN+LSTM +DNN(UPDRS) Acc = 0.993, Sen = 0.993 (H&Y) Acc = 0.993, Sen = 0.9931542025✓✓CNN+Transformer +GRU(D) Acc = 0.957, Sen = 0.990, Spe = 0.878 (S) Acc = 0.870, Sen = 0.870702025✓CNN+TransformerAcc = 0.966, Sen = 0.941692025✓GRUAcc = 0.987Other methodsRepeated random832013✓SVMAcc = 0.912, Sen = 0.917, Spe = 0.8992-fold1462016✓ANNAcc = 0.915, Sen = 0.914, Spe = 0.916892019✓ANNAcc = 0.914, Sen = 0.913, Spe = 0.9161272020✓✓ANN(D) Acc = 0.84015-fold1272020✓ANNAcc = 0.7736-fold1452025✓GRUAcc = 0.860, Sen = 0.8905 Independent Repetitions642025✓ANNAcc = 0.92810 Independent Repetitions1352022✓SVMAcc = 0.859, Sen = 0.940, Spe = 0.650LODO662022✓GNN(Ga_Ju) Acc = 0.750, (Ga_Si) Acc = 0.791 (Si_Ju) Acc = 0.787652023✓GNN(Ga_Ju) Acc = 0.778, (Ga_Si) Acc = 0.824 (Si_Ju) Acc = 0.7901212025✓GNN(Ga_Ju) Acc = 0.796, (Ga_Si) Acc = 0.865 (Si_Ju) Acc = 0.803672025✓GNN(Ga_Ju) Acc = 0.812, (Ga_Si) Acc = 0.872 (Si_Ju) Acc = 0.814Statistical Analysis1072009✓t-TEST172016✓Wilcoxon1042019✓Pearson correlation analysis1482021✓✓WilcoxonThis table summarizes PD gait analysis studies based on the PhysioNet PD Gait Dataset, organized by validation method: Hold-out, LOOCV, 5-fold, 10-fold and Other Methods. Each entry lists the reference (Ref), publication year (Yr), research objective (Obj: D: Diagnosis/S: Severity Assessment), model backbone (Backbone), and performance metrics (Perf.).Table 5Summary of studies from other datasetsRefYrObjDBBackbonePerf.DSHold-out612024✓✓25PD, 25HCSVMAcc = 0.954, Sen = 0.9541362022✓5PD, 5HCCNNAcc = 0.950, Sen = 0.900, Spe = 0.960492024✓14PDCNNAcc = 0.999, Sen = 0.880, Spe = 0.9901552024✓Synthetic DataCNN82021✓20PDGNNAcc = 0.821, Sen = 0.818, Spe = 0.88272023✓21PDCNN+GRUAcc = 0.757, Sen = 0.834, Spe = 0.7294-fold242011✓12PD, 20HCANNAcc = 0.9565-fold1122022✓✓19PD, 25HCRF+AdaBoost(D, RF) Acc = 0.880 (S, AdaBoost) Acc = 0.810222022✓42PD, 43HCKNN1562020✓5PDCNNSen = 0.943, Spe = 0.9511572025✓26PDGNN+ TransformerAcc = 0.978, Sen = 0.93310-fold1232021✓29PD, 18HCSVMAcc = 1.000, Sen = 1.000, Spe = 1.000252015✓23PD, 26HCRFAcc = 0.926, Sen = 0.960, Spe = 0.9001582018✓15PD, 15HCANNAcc = 0.933, Sen = 0.933, Spe = 0.9331172025✓✓111PD, 105HCCNN(D) Acc = 0.839, Sen = 0.839 (S) Acc = 0.70015-fold902008✓21PD, 17HCSVMAcc = 0.917, Sen = 0.752, Spe = 0.999Statistical analysis312008✓6PD, 14HCContribution rate1082022✓2PD, 6HCClustering232022✓108PD, 88HCt-TEST1222023✓30PD, 30HCPearson correlation analysisAcc = 0.883, Sen = 0.80, Spe = 0.967This table summarizes PD gait analysis studies based on other datasets, organized by validation method: hold-out, 4-fold, 5-fold, 10-fold, 15-fold, and Statistical Analysis. Each entry lists the reference (Ref), publication year (Yr), research objective (Obj: D diagnosis/S Severity assessment), model backbone (Backbone), and performance metrics (Perf.). Table 4Summary of studies based on the PhysioNet PD gait dataset50RefYrObjBackbonePerf.DSHold-out732018✓SVMAcc = 0.948, Sen = 0.970, Spe = 0.8701372019✓SVMAcc = 0.903782023✓SVMAcc = 0.974, Sen = 0.990, Spe = 0.950862023✓DTAcc = 0.880, Sen = 0.945, Spe = 0.805852021✓Ensemble DTAcc = 0.986, Sen = 0.996, Spe = 0.9951382025✓RFAcc = 0.960, Sen = 0.9601132022✓KNNAcc = 0.76722022✓Ensemble KNNAcc = 0.9501002025✓✓Weighted VotingAcc = 0.985, Sen = 0.9861392024✓LRAcc = 0.817, Sen = 0.897, Spe = 0.6301282020✓✓ANNAcc = 0.972, Sen = 0.9701272020✓✓ANN(D) Acc = 0.974, Sen = 0.977, Spe = 0.970 (S) Acc = 0.865292018✓✓CNN(D) Acc = 0.986182019✓✓CNNAcc = 0.887922021✓CNNAcc = 0.992, Sen = 1.000, Spe = 0.9801302022✓✓CNNAcc = 0.983, Sen = 0.982, Spe = 0.976942023✓CNNAcc = 0.9841322024✓✓CNNAcc = 0.985, Sen = 0.964, Spe = 0.9881402024✓CNN-XGBoostAcc = 0.9770 (Ga), 0.981 (Ju), 0.953 (Si)1412025✓CNNAcc = 0.9341422025✓CNNAcc = 0.990, Sen = 1.000, Spe = 0.9901432025✓CNNAcc = 0.991, Sen = 1.000, Spe = 1.0001092020✓LSTMAcc = 0.920, Sen = 0.9671182021✓✓LSTM(D) Acc = 0.986 (S) Acc = 0.966, Sen = 0.962, Spe = 0.9811442025✓LSTMAcc = 0.977, Sen = 0.990, Spe = 0.9601332023✓TransformerAcc = 0.988, Sen = 0.988912025✓GNNAcc = 0.995, Sen = 0.9971452025✓GRUAcc = 0.875, Sen = 0.875, Spe = 0.875LOOCV1102017✓SVMAcc = 1.000, Sen = 1.000, Spe = 1.000962019✓SVMAcc = 0.903, Sen = 0.902742022✓✓DT(D) Acc = 0.9931062023✓KNNAcc = 0.876, Sen = 0.851, Spe = 0.8891462016✓ANNAcc = 0.933, Sen = 0.924, Spe = 0.945892019✓ANNAcc = 0.988, Sen = 0.989, Spe = 0.9861272020✓ANN(D) Acc = 0.949, (S) Acc = 0.8765-fold1032012✓Statistical AnalysisAcc = 0.7731022016✓Statistical AnalysisAcc = 0.869, Sen = 0.720, Spe = 0.810842023✓✓DT(D) Acc = 0.933, Sen = 0.946, Spe = 0.917 (S) Acc = 0.914, Sen = 0.855, Spe = 0.9541462016✓ANNAcc = 0.963, Sen = 0.967, Spe = 0.9581272020✓✓ANN(D) Acc = 0.909, (S) Acc = 0.7301012021✓✓CNN+ SVM/KNN(D, KNN) Acc = 0.990, Sen = 0.990, Spe = 0.980 (S, SVM) Acc = 0.980, Sen = 0.940, Spe = 0.980872023✓✓CNN(D) Acc = 0.990, Sen = 0.995, Spe = 0.987 (S) Acc = 0.980, Sen = 0.987, Spe = 0.992RefYrObjBackbonePerf.DSFivefold1472025✓✓LSTM(D) Acc = 0.994, Sen = 0.994 (S) Acc = 0.982, Sen = 0.982662022✓GNNAcc = 0.938652023✓GNNAcc = 0.9591212025✓GNNAcc = 0.9991202019✓✓CNN+LSTM(D) Acc = 0.990, Sen = 0.991, Spe = 0.990 (S) Acc = 0.9801472025✓✓CNN+LSTM(D,LSTM) Acc = 0.994, Sen = 0.994 (S,CNN) Acc = 0.966, Sen = 0.96610-fold1482021✓✓SVM752021✓✓SVMAcc = 0.987, Sen = 0.976, Spe = 0.991762021✓✓SVMAcc = 0.984, Sen = 0.966, Spe = 0.996982022✓✓SVM(D) Acc = 0.985, (S) Acc = 0.977882024✓✓SVM(D) Acc = 0.988, Sen = 0.989, Spe = 0.986 (S) Acc = 0.966, Sen = 0.933, Spe = 0.977262020✓✓DTAcc = 1.000, Sen = 1.000, Spe = 1.0001242020✓✓DTAcc = 0.994, Sen = 0.986, Spe = 0.978712024✓DTAcc = 0.941, Sen = 0.942, Spe = 0.9831112017✓RFAcc = 0.980, Sen = 0.991, Spe = 0.957362018✓✓RF(D) Acc = 0.990, Sen = 0.978, Spe = 0.9951262023✓RFAcc = 0.990, Sen = 0.970, Spe = 0.9901492021✓RSRNAAcc = 0.810742022✓✓DT+ER(D) Acc = 0.999682025✓✓RFT+ER(D) Acc = 0.975, Sen = 0.970, Spe = 0.950 (S) Acc = 0.964892019✓ANNAcc = 0.969, Sen = 0.967, Spe = 0.9721142020✓ANNAcc = 0.983, Sen = 0.986, Spe = 0.9751272020✓✓ANN(D) Acc = 0.879, (S) Acc = 0.760342021✓ANNAcc = 0.963, Sen = 0.966, Spe = 0.959142020✓✓CNN(D) Acc = 0.987, Sen = 0.981, Spe = 1.000 (S) Acc = 0.853152020✓CNNAcc = 0.955812021✓✓CNN(D) Acc = 0.966, Sen = 0.944, Spe = 0.976 (S) Acc = 0.945, Sen = 0.920, Spe = 0.9561012021✓✓CNN(D) Acc = 0.990, Sen = 0.990, Spe = 0.980 (S) Acc = 0.980, Sen = 0.940, Spe = 0.9801152023✓✓CNN(D) Acc = 0.967, Sen = 0.981, Spe = 0.933 (S) Acc = 0.923, Sen = 0.9221502025✓✓CNN(D) Acc = 0.986, Sen = 0.992, Spe = 0.9851192024✓✓LSTM(D) Acc = 0.992, Sen = 0.991, Spe = 0.997 (S) Acc = 0.923, Sen = 0.924, Spe = 0.9771342024✓✓Transformer(D) Acc = 0.973, Sen = 0.960, Spe = 1.0001162024✓TransformerAcc = 0.878, Sen = 0.8801512025✓TransformerAcc = 0.923, Sen = 0.990, Spe = 0.8871522025✓TransformerAcc = 0.908, Sen = 0.910662022✓GNNAcc = 0.959652023✓GNNAcc = 0.9601212025✓GNNAcc = 0.9991532016✓DNN(Ga) Acc = 0.914, Sen = 0.770, Spe = 0.550 (Si) Acc = 0.921, Sen = 0.940, Spe = 0.890 (Ju) Acc = 0.909, Sen = 0.740, Spe = 0.8301292020✓DNN(D) Acc = 0.994(S) Acc = 0.991RefYrObjBackbonePerf.DS292018✓✓CNN+LSTM(D) Acc = 0.986932024✓CNN+LSTM +DNN(UPDRS) Acc = 0.993, Sen = 0.993 (H&Y) Acc = 0.993, Sen = 0.9931542025✓✓CNN+Transformer +GRU(D) Acc = 0.957, Sen = 0.990, Spe = 0.878 (S) Acc = 0.870, Sen = 0.870702025✓CNN+TransformerAcc = 0.966, Sen = 0.941692025✓GRUAcc = 0.987Other methodsRepeated random832013✓SVMAcc = 0.912, Sen = 0.917, Spe = 0.8992-fold1462016✓ANNAcc = 0.915, Sen = 0.914, Spe = 0.916892019✓ANNAcc = 0.914, Sen = 0.913, Spe = 0.9161272020✓✓ANN(D) Acc = 0.84015-fold1272020✓ANNAcc = 0.7736-fold1452025✓GRUAcc = 0.860, Sen = 0.8905 Independent Repetitions642025✓ANNAcc = 0.92810 Independent Repetitions1352022✓SVMAcc = 0.859, Sen = 0.940, Spe = 0.650LODO662022✓GNN(Ga_Ju) Acc = 0.750, (Ga_Si) Acc = 0.791 (Si_Ju) Acc = 0.787652023✓GNN(Ga_Ju) Acc = 0.778, (Ga_Si) Acc = 0.824 (Si_Ju) Acc = 0.7901212025✓GNN(Ga_Ju) Acc = 0.796, (Ga_Si) Acc = 0.865 (Si_Ju) Acc = 0.803672025✓GNN(Ga_Ju) Acc = 0.812, (Ga_Si) Acc = 0.872 (Si_Ju) Acc = 0.814Statistical Analysis1072009✓t-TEST172016✓Wilcoxon1042019✓Pearson correlation analysis1482021✓✓WilcoxonThis table summarizes PD gait analysis studies based on the PhysioNet PD Gait Dataset, organized by validation method: Hold-out, LOOCV, 5-fold, 10-fold and Other Methods. Each entry lists the reference (Ref), publication year (Yr), research objective (Obj: D: Diagnosis/S: Severity Assessment), model backbone (Backbone), and performance metrics (Perf.). Summary of studies based on the PhysioNet PD gait dataset50 This table summarizes PD gait analysis studies based on the PhysioNet PD Gait Dataset, organized by validation method: Hold-out, LOOCV, 5-fold, 10-fold and Other Methods. Each entry lists the reference (Ref), publication year (Yr), research objective (Obj: D: Diagnosis/S: Severity Assessment), model backbone (Backbone), and performance metrics (Perf.). Table 5Summary of studies from other datasetsRefYrObjDBBackbonePerf.DSHold-out612024✓✓25PD, 25HCSVMAcc = 0.954, Sen = 0.9541362022✓5PD, 5HCCNNAcc = 0.950, Sen = 0.900, Spe = 0.960492024✓14PDCNNAcc = 0.999, Sen = 0.880, Spe = 0.9901552024✓Synthetic DataCNN82021✓20PDGNNAcc = 0.821, Sen = 0.818, Spe = 0.88272023✓21PDCNN+GRUAcc = 0.757, Sen = 0.834, Spe = 0.7294-fold242011✓12PD, 20HCANNAcc = 0.9565-fold1122022✓✓19PD, 25HCRF+AdaBoost(D, RF) Acc = 0.880 (S, AdaBoost) Acc = 0.810222022✓42PD, 43HCKNN1562020✓5PDCNNSen = 0.943, Spe = 0.9511572025✓26PDGNN+ TransformerAcc = 0.978, Sen = 0.93310-fold1232021✓29PD, 18HCSVMAcc = 1.000, Sen = 1.000, Spe = 1.000252015✓23PD, 26HCRFAcc = 0.926, Sen = 0.960, Spe = 0.9001582018✓15PD, 15HCANNAcc = 0.933, Sen = 0.933, Spe = 0.9331172025✓✓111PD, 105HCCNN(D) Acc = 0.839, Sen = 0.839 (S) Acc = 0.70015-fold902008✓21PD, 17HCSVMAcc = 0.917, Sen = 0.752, Spe = 0.999Statistical analysis312008✓6PD, 14HCContribution rate1082022✓2PD, 6HCClustering232022✓108PD, 88HCt-TEST1222023✓30PD, 30HCPearson correlation analysisAcc = 0.883, Sen = 0.80, Spe = 0.967This table summarizes PD gait analysis studies based on other datasets, organized by validation method: hold-out, 4-fold, 5-fold, 10-fold, 15-fold, and Statistical Analysis. Each entry lists the reference (Ref), publication year (Yr), research objective (Obj: D diagnosis/S Severity assessment), model backbone (Backbone), and performance metrics (Perf.). Summary of studies from other datasets This table summarizes PD gait analysis studies based on other datasets, organized by validation method: hold-out, 4-fold, 5-fold, 10-fold, 15-fold, and Statistical Analysis. Each entry lists the reference (Ref), publication year (Yr), research objective (Obj: D diagnosis/S Severity assessment), model backbone (Backbone), and performance metrics (Perf.). PD diagnosis Plantar pressure has emerged as a highly promising biomechanical marker for early screening and auxiliary diagnosis of PD. Over the past decade, methods spanning traditional statistical analysis, machine learning, and deep learning have been developed to identify clinically relevant features from plantar pressure signals for PD diagnosis. The following methods systematically review advancements in PD diagnosis research from the three aforementioned aspects. Statistical analysis methods In studies utilizing statistical analysis for the diagnosis of PD based on plantar pressure data, existing research can be broadly categorized into three methodological classes. Within feature difference analysis, multiple studies have compared gait parameters between PD patients and healthy controls using one-way ANOVA, t-tests, and ROC analysis. These studies consistently demonstrated that PD patients commonly exhibit changes in spatio-temporal features, including shortened stride length, reduced swing phase duration, and prolonged stance phase. Moreover, kinematic features have been reported to provide greater diagnostic value than kinetic features23,102,103. In the domain of signal complexity quantification, autoregressive models have been applied to analyze the frequency characteristics of plantar pressure signals. When the associated discrete index is used as a diagnostic indicator, PD patients demonstrated a more compact and similar power spectral distribution within specific frequency bands. Moreover, multiscale approximate entropy has been employed to assess the complexity of gait signals, revealing that the MAE of PD patients exhibits distinctive patterns of change with increasing scale, which differentiates them from patients with amyotrophic lateral sclerosis and Huntington’s disease104,107. Within the realm of unsupervised learning and data fusion, the integration of plantar pressure signals with hierarchical clustering algorithms, as well as the application of PCA to fuse sensor data from multiple sources, has achieved promising classification performance108,109. Machine learning methods Traditional machine learning approaches play an important role in the diagnosis of PD. The following content provides a systematic overview of the applications and performance outcomes of various ML-based models in the context of PD diagnosis using plantar pressure data. Support Vector Machine: As a classical machine learning algorithm recognized for its efficacy in managing small-sample, high-dimensional, and nonlinear data, the SVM has been widely adopted in PD diagnostic studies utilizing plantar pressure data. Capitalizing on its capacity to delineate complex decision boundaries through kernel function mapping, SVM consistently demonstrated superior performance over other methods such as KNN, DT, RF, and NB in terms of both classification accuracy and robustness88,96,99. Furthermore, the integration of SVM with tensor decomposition techniques or the design of application-specific kernel functions has shown effectiveness in processing multi-dimensional temporal signals and frequency domain features, leading to further improvements in diagnostic capability83,110. In addition, when combined with WT or frequency domain feature extraction strategies, SVM is better equipped to capture long-range dependencies and nonlinear characteristics within gait signals, which contributes to the early detection and classification of PD73,78. Decision Tree: DT models have been widely adopted as a classical machine learning approach in plantar pressure analysis owing to their interpretable structure, computational transparency, and robustness to feature redundancy and outliers. In multiple studies, DT and its ensemble variants have demonstrated strong diagnostic performance and practical application value26,85,86. Among these, the DT classifier constructed using multi-domain features achieved 100% classification accuracy in a binary classification task26. Additionally, by combining weight-normalized VGRF signals with nonlinear features such as Euclidean distance and incorporating the Gradient Boosting Decision Trees (GBDT) model, the diagnostic accuracy was further improved to 99.52%, demonstrating the integrated decision tree structure’s excellent adaptability to individual differences85. Random Forest: In research on the diagnosis of PD, RF models and their derivative ensemble tree architectures have been extensively adopted due to their strong nonlinear modeling capabilities, high tolerance for feature redundancy and minimal requirement for complex normalization procedures. Multiple studies have shown that feeding regression-normalized, multi-dimensional gait features into RF classifiers can achieve high accuracy and robustness across various neurodegenerative disease classification tasks25,89,111. Moreover, in binary classification tasks for identifying PD patients, RF has outperformed both AdaBoost and Extremely Randomized Trees, highlighting its efficacy in PD detection112. K-Nearest Neighbors: various studies have focused on optimizing feature design and distance metric strategies to improve the diagnostic performance of the KNN method86,106,113. To enhance the effectiveness of KNN in capturing temporal and nonlinear characteristics, time-frequency domain features were extracted from plantar pressure signals, and the performance of Euclidean distance and Dynamic Time Warping was compared within the KNN framework. A classification accuracy of 85% was attained at K=5113. Furthermore, from the perspective of distance metrics, KNN was applied for the selection of fuzzy rule centers. A local clustering structure was constructed in the feature space using the minimum mean square distance, demonstrating the ability of KNN to support the organization of high-dimensional feature spaces106. Deep learning methods In contrast to conventional machine learning models that rely on handcrafted feature engineering, deep learning methods can automatically learn discriminative features directly from raw plantar pressure sequences. This end-to-end learning paradigm enhances their capability to model complex nonlinear relationships and capture dynamic temporal patterns inherent in gait signals. Artificial Neural Network: In research on the diagnosis of PD using plantar pressure, ANN has been widely employed due to its nonlinear modeling capabilities. Extensive research has focused on improving ANN architectures to enhance classification performance34,114. Specifically, an NR-LBP feature extraction framework combined with a single-hidden-layer ANN achieved an accuracy of 98.3% by leveraging variance features based on max-min representation, emphasizing the importance of feature optimization even within simple ANN architectures114. In another study, researchers integrated the SWLNGP local pattern transformation with an ANN, and applied 48-dimensional features and the application of the trainscg training function, attaining an accuracy of 96.28%, demonstrating that dimensionality reduction and algorithmic optimization can significantly enhance the efficiency of ANN models34. Convolutional Neural Network: Research on PD diagnosis using CNNs applied to plantar pressure data has developed into a diverse technical framework. The first approach involves using 1D-CNNs to directly process raw VGRF time-series signals for unimodal signal modeling14,81. For instance, a system employing 18 parallel 1D-ConvNet branches to fuse multi-sensor features achieved a detection accuracy of 98.7%14. In another study, VGRF signals were transformed into time-frequency spectrograms and classified using ResNet-50, attaining an accuracy of 96.52%81. The second direction focuses on multimodal spatiotemporal modeling87,115. A 3D-CNN was used to integrate spatial and temporal information, and a dual-stream 3D-CNN incorporating squeeze-and-excitation modules and symmetry analysis achieved an accuracy of 99.07%87. Additionally, static-dynamic network using a hybrid 1D/2D-CNN architecture was constructed to capture sensor time-series patterns and COP transitions, achieving 96.7% accuracy115. The third approach involves hybrid architecture optimization by combining CNNs with other modules92,116. For example, spatial features extracted by CNNs were fed into Bi-LSTM networks for sequential modeling92. A CNN-Transformer hybrid model was also proposed to capture both local and global discriminative features, achieving a classification accuracy exceeding 99%116. Furthermore, a Federated Transferring Multi-Channel CNN was developed, which employs the self-collected XJTU dataset as the target domain and leverages a layer-adaptive transfer loss for fine-tuning the pre-trained model, while simultaneously mitigating spurious performance gains arising from data leakage, thereby improving its clinical applicability117. Long Short-Term Memory: Among deep learning algorithms for plantar pressure based on PD diagnosis, LSTM networks have attracted widespread attention due to their strong capability in modeling temporal dependencies in sequential biomechanical signals. Several studies have adopted raw VGRF time-series as inputs, employing stacked LSTM and dense layers to effectively capture intricate dependencies within gait cycles, thus validating the effectiveness of LSTM in end-to-end classification tasks109,118. Meanwhile, several improved LSTM variants have been proposed, such as lightweight LSTM models and dual-modal attention-enhanced LSTM networks119,120. The lightweight LSTM architecture achieved 99.19% classification accuracy in PD detection using only four key features derived from single-dimensional gait signals, with hyperparameters optimized via Bayesian optimization. The AE-LSTM model combines CNN-based spatial feature extraction with attention mechanisms that dynamically weight critical time steps, while also incorporating asymmetry analysis between the left and right foot. This approach achieved a maximum accuracy of 99.31% in binary classification tasks. Graph Neural Network: In recent years, adaptive graph structure modeling has emerged as a key approach to enhance the representational capacity of GNNs. Frequency domain graph adaptive networks (RFdGAD) and adaptive spatiotemporal directed graph networks (AST-DGNN) have both attempted to dynamically construct graph topologies based on intrinsic relationships in plantar pressure data, effectively addressing the limitations of static, predefined adjacency matrices65,66. Specifically, RFdGAD constructed dynamic graph structures from frequency domain features and employed a spatial dependency learning module to adaptively model inter-sensor relationships on the foot using a probabilistic graph framework. In contrast, AST-DGNN introduced directed edge mechanisms to encode pressure transmission directions within the foot into the graph topology, and utilizes a trainable global adaptive matrix to model the overall spatial connectivity. When integrated with a TCN, this architecture achieved an average accuracy improvement of 2.13% in cross-dataset validation experiments. Building upon AST-DGNN, Multi-Scale Adaptive Directed Graph Neural Network introduced a multi-scale perspective by modeling VGRF signals as multi-scale directed graph structures for PD detection121. Recently, a Global-Local Dynamic Directed Graph Neural Network has been proposed, which incorporates a Dynamic Graph Learning unit to learn the temporal dynamics of topological structures from gait signals, thereby capturing the evolving nature of plantar pressure propagation67. PD severity assessment With the rapid advancement of deep learning and machine learning technologies, an increasing number of studies have focused on applying these techniques to the objective assessment of PD severity. In clinical practice, PD severity is typically assessed using the H&Y staging system and MDS-UPDRS, where higher scores typically indicate more advanced disease progression. Among the various motor symptoms, FOG is closely linked to disease progression, with its frequency and severity showing a positive correlation with clinical scores. VGRF provides rich quantitative data for assessing PD severity in conjunction with FOG characteristics, making modeling based on deep learning and machine learning an emerging research focus. The following content reviews recent advances and representative work on PD severity assessment using plantar pressure signals and quantitative FOG characteristics via diverse computational methods. Statistical analysis methods The quantitative detection of FOG is considered a critical aspect of assessing PD severity and identifying the progression of motor impairments. To this end, recent research has extensively investigated techniques such as wearable sensor-based data acquisition, heterogeneous data fusion, and nonlinear signal modeling to improve the sensitivity and clinical relevance of FOG detection. Although the specific methodologies vary, these studies demonstrate strong consistency in their core objectives to identify subtle motor abnormalities, and capture individual-specific patterns of variability. Firstly, traditional linear statistical features are widely considered insufficient for the identification of FOG in its early stages. To address this limitation, various nonlinear dynamic analysis and symbolic modeling methods have been developed to better capture the complexity of motor disorders in PD104,105. For example, multiscale approximate entropy has been applied to quantify the complexity of VGRF signals, revealing disordered motor control patterns of neurodegenerative diseases from a global dynamic perspective104. In addition, genetic programming has been utilized to model plantar pressure signals nonlinearly, demonstrating superior fitting performance and generalization capability compared to conventional linear methods105. Secondly, multimodal data fusion and cross-channel dynamic co-modeling are regarded as key strategies to improve the robustness of FOG detection17,23,122. For example, gait influence diagrams have been used to capture the cooperative relationships between different plantar regions, uncovering anterior-posterior and medial-lateral gait asymmetries in patients with mild PD. This approach provided novel perspectives for modeling early-stage symptoms17. Furthermore, wearable systems integrating multiple sensor types, such as force-sensitive, piezoelectric, and IMUs have been developed to extract heterogeneous features including pressure, dynamic deformation, and postural angles. Experimental results demonstrated that such fusion strategies can significantly improve model performance122. Additionally, studies utilizing force-sensitive insoles to collect kinematic and dynamic data have revealed that gait rhythm dispersion exhibits high discriminative power in patients with advanced PD. Specifically, compared to traditional mean-based features, certain rhythm-based indicators achieve an AUC greater than 0.8 in ROC analysis. This finding suggests that gait rhythm instability may constitute a significant biomarker of PD progression23. Machine learning methods Traditional machine learning approaches remain fundamental to the task of assessing PD severity. Typically, these studies extract quantitative features from plantar pressure signals and use them to construct multi-class classification models for PD severity staging and scoring. The following content reviews various ML-based methodologies employed for PD severity assessment. Support Vector Machine: Studies utilizing SVM for assessing PD severity have primarily followed three methodological directions: kernel function optimization, feature engineering, and multi-class classification strategies. In terms of kernel optimization, particle swarm optimization has been employed to optimize the parameters of linear, polynomial, and radial basis function (RBF) kernels. Experimental results indicated that the RBF kernel outperforms others in PD severity classification, achieving an accuracy of 95.35%61. Another study demonstrated that the quadratic kernel SVM attained the highest accuracy of 98.56% in H&Y staging tasks75. For feature engineering, one approach involves converting gait signals into grayscale texture images, followed by deep feature extraction and Lasso-based feature selection, which yielded 100% accuracy in experiments stratified by gender101. Alternatively, the application of persistent entropy and topological feature extraction achieved a classification accuracy of 98.08% under small-sample conditions123. Regarding multi-class classification, the One-vs-One strategy has been commonly adopted to address the multi-class nature of PD severity staging61,75,123. Decision Tree: In plantar pressure-based analysis using DT, numerous studies have validated the high efficiency and accuracy of DT models in assessing PD severity. These studies typically begin with feature extraction, employing correlation analysis to identify key spatio-temporal gait biomarkers or constructing multidimensional feature sets from the time, frequency, and time-frequency domains. Following feature selection and integration, the processed data are fed into DT classifiers to effectively differentiate between PD patients and HC26,124. Building on this foundation, some studies have introduced structural enhancements to standard DT models. For instance, the Information Set Decision Tree framework incorporates uncertainty modeling of features along with the MMOTE oversampling technique to address class imbalance, achieving an overall accuracy of 94.14%71. Additionally, a two-stage hybrid strategy has been proposed, wherein a DT classifier is first used to distinguish PD from HC, followed by an ensemble regressor to quantify PD severity. This method attained an accuracy as high as 99.9% under 10-fold cross-validation74. The model employed Hannan-Amirban entropy as the splitting criterion for constructing the decision tree, which demonstrated significantly superior performance in disease classification and severity prediction compared with conventional decision trees based on the Gini index or information gain125. For the detection of FOG, a DT model was built using ReliefF feature selection and random undersampling to enhance decision tree construction. It was found that bilateral plantar pressure data enabled the prediction of 94.2% of FOG episodes approximately 0.8 seconds in advance80. Random Forest: RF, a robust and interpretable ensemble learning algorithm, has been extensively utilized36,84,126. A method combining RF with a Mahalanobis distance-based oversampling strategy was proposed, where more representative new samples were generated from the PhysioNet dataset, resulting in an outstanding accuracy of 99% and a recall of 97%126. Furthermore, by integrating optical flow and spatio-temporal features, a multi-class model based on Classification and Regression Trees was constructed, achieving an overall accuracy of 91.4% and outperforming conventional RF models84. K-Nearest Neighbors: KNN algorithm has been widely applied due to its simplicity, efficiency, and sensitivity to nonlinear data distributions. Notably, the application of KNN in detecting motor symptoms of PD, such as postural instability and bradykinesia, has received considerable attention22,98. Experimental results revealed that KNN performed effectively in assessing both postural stability and bradykinesia, particularly after features related to the spatial range of COP and base of support were extracted. These findings suggested that KNN could sensitively detect subtle and nonlinear postural variations. Deep learning methods Deep learning methods have become a prominent focus in PD severity assessment due to their strong capabilities in temporal sequence modeling and complex feature representation. The content that follows provides a systematic overview of various DL architectures applied in PD severity evaluation and offers a comparative analysis of their performance. Artificial Neural Network and Deep Neural Network: A two-stage model based on VGRF signals and ANN was constructed to perform binary classification for PD diagnosis, as well as multi-class classification for severity assessment. The ANN was trained using Resilient Backpropagation and the Levenberg-Marquardt algorithm for optimization, while the SMOTE was applied to mitigate class imbalance. An accuracy of 87.1% was achieved for severity level classification127. By incorporating frequency-domain features, statistical measures were extracted from 16-channel VGRF signals via FFT, and a deep multi-hidden-layer ANN was employed for PD stage classification, resulting in an overall classification accuracy of 97%128. In contrast, studies involving DNN have focused on leveraging deep architectures to extract spatiotemporal features directly from raw VGRF signals, thus eliminating the need for manual feature engineering. An end-to-end CLDNN model integrating CNN, LSTM, and DNN was developed to achieve hierarchical feature fusion across spatial, temporal, and fully connected layers, yielding high-performance results93. Additionally, a 6–9 layer DNN architecture was designed, incorporating Bayesian hyperparameter optimization and feature enhancement techniques, which led to a multi-class classification accuracy of 99.1%129. Long Short-Term Memory: LSTM networks have been widely adopted due to their powerful capability to model temporal dynamics in gait sequences. Hybrid CNN-LSTM architectures are frequently utilized29,130. IMFs were extracted from gait signals via empirical mode decomposition, and dominant IMFs were identified through power spectral analysis to train the CNN-LSTM model, achieving a multi-class classification accuracy of 98.32%130. In one dual-channel model, LSTM was responsible for capturing the temporal dependencies of gait data, while CNN extracted spatial features using 5 × 5 convolutional kernels and max pooling. The fusion of these features led to a maximum PD detection accuracy of 98.88%29. Moreover, in models for detecting and predicting FOG in PD patients based on VGRF, a two-layer LSTM network demonstrated superior performance. For FOG prediction, a three-layer LSTM model achieved an average specificity of 81.2% in binary classification under leave-one-out cross-validation131. Convolutional Neural Network: CNN has been widely applied and demonstrated remarkable effectiveness in the analysis of high-dimensional and multimodal gait data due to its superior spatial feature extraction capabilities. In practical implementations, researchers have explored various CNN variants, such as the design and comparative evaluation of multiple Deep Convolutional Neural Network (DCNN) architectures to identify the optimal configuration, the proposal of a framework integrating Variational Mode Decomposition and Recurrence Plots with DCNN, and the development of TCNs. Through architectural optimization, expansion of network depth, and tailored design of convolutional kernel sizes, these approaches have significantly enhanced the performance of CNN-based models in gait signal representation15,49,132. Moreover, CNNs are often integrated with other models. For instance, the CLDNN model combines CNN with LSTM, while parallel DCNN structures incorporate multiple feature extraction pathways. The CNN component consistently extracts highly discriminative features from gait signals that are strongly correlated with PD severity, demonstrating robust encoding capabilities across hybrid architectures15,93. Transformer: In the field of PD severity assessment, Transformer-based architectures have been increasingly adopted for gait signal analysis due to their advantages in modeling long-range dependencies and parallel sequence processing. A key emphasis in current research lies in the Transformer’s ability to model long-term temporal characteristics inherent in gait data. By incorporating mechanisms such as positional encoding and a [CLS] token, the model is capable of effectively modeling the global dynamic patterns within complex time-series data like VGRF, thereby enhancing its discriminative ability across different stages of PD116,133. Moreover, Transformer-based models demonstrate strong flexibility and performance in multimodal data fusion. By integrating both raw VGRF signals and handcrafted spatio-temporal and frequency domain features into the Transformer framework, and using genetic algorithms for hyperparameter optimization, high precision regression prediction of PD severity levels has been achieved134. Additionally, to further enhance representational capacity and discriminative performance, several studies have explored hybrid architectures that combine Transformers with local feature extractors. For instance, a 1D-ConvNet module was employed to capture local patterns, followed by temporal and spatial Transformer encoders to model long-range dependencies in gait signals and inter-sensor spatial correlations. This “local-global” fusion strategy demonstrated high classification accuracy and F1-scores in PD staging tasks116. Model performance and limitations We have reviewed the applications of various algorithms in PD diagnosis and severity assessment from the perspectives of statistical analysis, machine learning, and deep learning. This section summarizes and compares the performance and limitations of these models. Accuracy, sensitivity and specificity are the most widely used performance metrics for machine learning and deep learning models in PD diagnosis, and are also key indicators of clinical concern. The detailed definitions and clinical implications of these metrics are elaborated in Section Performance metrics. Given the research status of PD, PD diagnosis studies are abundant, leading to sufficient reports of accuracy, sensitivity, and specificity. In contrast, PD severity assessment is still in rapid development, with few studies reporting Sensitivity and Specificity. As a box plot requires a sufficient sample size to reliably reflect the distribution and dispersion of performance data, limited samples would result in an unrepresentative distribution. Therefore, we only present the three metrics for PD diagnosis, and visualize only Accuracy for PD severity staging as shown in Fig. 11. In addition, AUC-ROC is a valuable comprehensive metric that reflects the overall discriminative ability of the model, especially for imbalanced clinical datasets. It quantifies the probability that the model ranks a random positive sample higher than a random negative sample, providing a robust and unbiased measure for clinical screening tasks. However, it is rarely reported in current studies. Future research should adopt AUC-ROC as a supplementary evaluation metric to enrich model performance verification and improve the comparability of research results.Fig. 11Visualization of accuracy, sensitivity, and specificity for PD diagnosis and severity assessment models.This figure presents four box plots visualizing key performance metrics of various models for PD diagnosis and severity assessment using VGRF data. a Shows accuracy for PD diagnosis. b Displays sensitivity for PD diagnosis. c Presents specificity for PD diagnosis. d Illustrates accuracy for PD severity assessment. Visualization of accuracy, sensitivity, and specificity for PD diagnosis and severity assessment models. This figure presents four box plots visualizing key performance metrics of various models for PD diagnosis and severity assessment using VGRF data. a Shows accuracy for PD diagnosis. b Displays sensitivity for PD diagnosis. c Presents specificity for PD diagnosis. d Illustrates accuracy for PD severity assessment. As shown in Fig. 11, for the PD diagnosis task, SVM, RF, CNN, and LSTM achieve excellent and stable performance across Accuracy, Sensitivity, and Specificity. Among them, CNN delivers the best overall performance with a narrow box range and concentrated data distribution, demonstrating strong classification capability and clinical applicability. In contrast, KNN and GNN exhibit obvious performance fluctuations and insufficient robustness. PD severity assessment is a more challenging task, and the overall accuracy of all algorithms is lower than that in the diagnosis task. Among them, SVM, CNN, and LSTM perform prominently, while RF and ANN show large performance variations. As a popular method in current research, GNN still requires further improvement in performance. Overall, deep learning models generally outperform traditional machine learning models in stability. However, as black-box models, deep learning approaches have inherent limitations in interpretability, which can hardly meet the transparency requirements of clinical diagnosis and treatment. In summary, existing models still face several critical challenges. Table 6 summarizes the common limitations identified in the literature review, aiming to provide methodological guidance and directions for improvement in future research.Table 6Limitations of commonly used models for PD diagnosis and PD severity assessmentCategoryModelIdentified limitationsStatisticalTraditional statisticalLinear assumptions fail to capture nonlinear and dynamic relations. Frequency domain features have limited biological interpretability. PCA-based fusion may discard locally discriminative information.MLSVMKernel and parameter tuning rely on manual settings, lacking adaptability to variations. Feature selection is manually driven, not integrated with SVM’s margin-based optimization. One-vs-one strategy leads to classifier explosion, reducing real-time applicability.DTGreedy splitting is sensitive to noise, causing false feature dependencies. Overfitting risk due to overly deep trees and small leaf sample sizes.RFStatic hyperparameters limit minority class recognition in imbalanced datasets. Poor at modeling high-order nonlinear temporal patterns. High-dimensional features demand numerous trees, increasing model complexity.KNNUses static Euclidean distance, insensitive to nonlinear, dynamic gait changes. Highly affected by unbalanced datasets, with reduced specificity. Suffers from “curse of dimensionality” and poor scalability to large datasets.DLANNShallow architectures limit nonlinear gait representation. Network design is manual, lacking systematic optimization. Feature selection is decoupled from classification, leading to possible information loss. Model performance is unstable due to sensitivity to initialization and learning rate.CNNCaptures local but not global temporal dependencies. Fixed kernel size limits multi-scale spatiotemporal feature extraction. Large parameter count hinders deployment on edge devices.LSTMGradient vanishing reduces the ability to model long-term degradation trends. Network configurations are fixed and lack quantitative exploration across disease stages.TransformerRequires large-scale, balanced datasets; current PD gait datasets are limited. Quadratic time complexity of self-attention impairs long sequence modeling. Weak in aligning multi-cycle, multi-modal temporal patterns across individuals/devices.GNNGraph design relies on prior knowledge and lacks adaptability. High computational cost due to dynamic graph generation and message passing.This table systematically categorizes and summarizes the key limitations of widely adopted models for PD diagnosis and severity assessment, grouping them into three methodological categories: statistical, machine learning, and deep learning. Table 6Limitations of commonly used models for PD diagnosis and PD severity assessmentCategoryModelIdentified limitationsStatisticalTraditional statisticalLinear assumptions fail to capture nonlinear and dynamic relations. Frequency domain features have limited biological interpretability. PCA-based fusion may discard locally discriminative information.MLSVMKernel and parameter tuning rely on manual settings, lacking adaptability to variations. Feature selection is manually driven, not integrated with SVM’s margin-based optimization. One-vs-one strategy leads to classifier explosion, reducing real-time applicability.DTGreedy splitting is sensitive to noise, causing false feature dependencies. Overfitting risk due to overly deep trees and small leaf sample sizes.RFStatic hyperparameters limit minority class recognition in imbalanced datasets. Poor at modeling high-order nonlinear temporal patterns. High-dimensional features demand numerous trees, increasing model complexity.KNNUses static Euclidean distance, insensitive to nonlinear, dynamic gait changes. Highly affected by unbalanced datasets, with reduced specificity. Suffers from “curse of dimensionality” and poor scalability to large datasets.DLANNShallow architectures limit nonlinear gait representation. Network design is manual, lacking systematic optimization. Feature selection is decoupled from classification, leading to possible information loss. Model performance is unstable due to sensitivity to initialization and learning rate.CNNCaptures local but not global temporal dependencies. Fixed kernel size limits multi-scale spatiotemporal feature extraction. Large parameter count hinders deployment on edge devices.LSTMGradient vanishing reduces the ability to model long-term degradation trends. Network configurations are fixed and lack quantitative exploration across disease stages.TransformerRequires large-scale, balanced datasets; current PD gait datasets are limited. Quadratic time complexity of self-attention impairs long sequence modeling. Weak in aligning multi-cycle, multi-modal temporal patterns across individuals/devices.GNNGraph design relies on prior knowledge and lacks adaptability. High computational cost due to dynamic graph generation and message passing.This table systematically categorizes and summarizes the key limitations of widely adopted models for PD diagnosis and severity assessment, grouping them into three methodological categories: statistical, machine learning, and deep learning. Limitations of commonly used models for PD diagnosis and PD severity assessment This table systematically categorizes and summarizes the key limitations of widely adopted models for PD diagnosis and severity assessment, grouping them into three methodological categories: statistical, machine learning, and deep learning. Performance metrics During model development, validation, and evaluation are essential steps to ensure the model’s reliability and generalization ability in practical applications. The following content will provide an in-depth discussion of the main methods and significant roles of model validation and evaluation. Validation This survey examines commonly used validation methods, including K-Fold cross-validation, Hold-out validation, Leave-one-out cross-validation (LOOCV), and leave-one-dataset-out cross-validation (LODO). Additionally, other approaches such as cross-dataset validation, Wilcoxon rank-sum test, and the two-tailed t test are also utilized. K-fold cross-validation involves dividing the dataset into K subsets. Each subset is sequentially utilized as the test set, while the remaining (K-1) subsets are used as the training set. This process is repeated until each subset has been used as the test set exactly once. The Hold-out method involves splitting the dataset into training and testing sets according to a predefined ratio, with common splits being 80% and 20% or 70% and 30%. LOOCV selects one subject’s data at a time as the test set, with the remaining data from all other subjects used for training. This process is repeated N times, where N represents the total number of subjects. LODO is similar to LOOCV but differs in that an entire dataset is used as the test set, while all other datasets are used for training. In PD diagnosis models, different validation methods are applied based on the characteristics of the data and the specific requirements of the model. K-fold cross-validation is suitable for datasets of moderate size and relatively uniform distribution, typically containing hundreds of patients. The Hold-out method is appropriate for large datasets, preliminary validation, and rapid model screening. LOOCV is particularly effective for extremely limited datasets, such as those with only a few dozen patients, as it maximizes data utilization, enhancing the model’s learning capability. LODO is most appropriate for datasets with significant source variations, such as patient data collected from different hospitals, geographic regions, or devices. While methods like LODO begin to address variations between data sources, a more rigorous evaluation involves testing cross-dataset performance using entirely independent datasets. This validation is critical for assessing true generalizability and ensuring models are robust enough for diverse clinical environments, adding significant value beyond within-dataset testing methods. Evaluation In the task of PD diagnosis, selecting appropriate performance evaluation metrics is of critical importance for assessing the effectiveness and clinical usability of models. Given that PD diagnosis often involves binary classification tasks with class imbalance, this section categorizes evaluation metrics into four types based on clinical requirements and model application objectives: missed diagnosis control metrics, misdiagnosis control metrics, overall performance assessment metrics, and regression error control metrics. The distribution of metric usage across these categories is illustrated in Fig. 12, while their medical implications are detailed in Table 7.Fig. 12Usage charts for four types of metrics.This bar chart illustrates the usage frequency of performance metrics across four categories in PD research using VGRF data. The metrics are grouped into four distinct categories, each represented by colored bars. Missed diagnosis control (pink bars), misdiagnosis control (cyan bars), overall performance (orange bars), and regression error (blue bars).Table 7Description of the medical implications of each indicatorTypeMetricDescription of the medicalMissed diagnosis control metricsSensitivityAbility to detect patients.NPVThe proportion of truly healthy individuals within the group that tested negative.FNRProbability of being sick but incorrectly classified as disease-free (greater clinical risk).Misdiagnosis control metricsSpecificityAbility to exclude healthy people.PrecisionThe proportion of actually sick individuals within the group that tested positive.FPRProbability of being detected as sick among healthy individuals (may trigger unnecessary intervention).Overall performance evaluation metricsAccuracyThe proportion of all forecasts that are correctly predicted.F1-scoreIt is often used to balance recall and precision.AUC-ROCA composite measure of the model’s ability to classify positive and negative.MCC—G-means—Kappa—Regression error metricsCCIt is often used to analyze the degree of linear correlation between two continuous variables.RMSEError comparisons are often used in medical research.MAEThe units are the same as the actual measurements and are clinically interpretable.SDUsed to represent the variability of measurements in different subjects.This table categorizes and defines the clinical meanings of key performance metrics used to evaluate diagnostic models. Metrics are grouped into four categories: missed diagnosis control, misdiagnosis control, overall performance evaluation, and regression error metrics. Usage charts for four types of metrics. This bar chart illustrates the usage frequency of performance metrics across four categories in PD research using VGRF data. The metrics are grouped into four distinct categories, each represented by colored bars. Missed diagnosis control (pink bars), misdiagnosis control (cyan bars), overall performance (orange bars), and regression error (blue bars). Table 7Description of the medical implications of each indicatorTypeMetricDescription of the medicalMissed diagnosis control metricsSensitivityAbility to detect patients.NPVThe proportion of truly healthy individuals within the group that tested negative.FNRProbability of being sick but incorrectly classified as disease-free (greater clinical risk).Misdiagnosis control metricsSpecificityAbility to exclude healthy people.PrecisionThe proportion of actually sick individuals within the group that tested positive.FPRProbability of being detected as sick among healthy individuals (may trigger unnecessary intervention).Overall performance evaluation metricsAccuracyThe proportion of all forecasts that are correctly predicted.F1-scoreIt is often used to balance recall and precision.AUC-ROCA composite measure of the model’s ability to classify positive and negative.MCC—G-means—Kappa—Regression error metricsCCIt is often used to analyze the degree of linear correlation between two continuous variables.RMSEError comparisons are often used in medical research.MAEThe units are the same as the actual measurements and are clinically interpretable.SDUsed to represent the variability of measurements in different subjects.This table categorizes and defines the clinical meanings of key performance metrics used to evaluate diagnostic models. Metrics are grouped into four categories: missed diagnosis control, misdiagnosis control, overall performance evaluation, and regression error metrics. Description of the medical implications of each indicator This table categorizes and defines the clinical meanings of key performance metrics used to evaluate diagnostic models. Metrics are grouped into four categories: missed diagnosis control, misdiagnosis control, overall performance evaluation, and regression error metrics. Missed diagnosis control metrics Missed diagnosis refers to the failure to detect true PD patients, which may cause patients to miss early intervention opportunities and represents a high-risk scenario in PD diagnosis. This is particularly critical in early screening contexts. To control missed diagnosis risks, commonly used evaluation metrics include: Sensitivity (True Positive Rate, TPR): This metric denotes the proportion of positive samples correctly predicted by the classification model among all positive samples. It reflects the model’s ability to identify true PD patients and serves as a key indicator for measuring missed diagnosis rates34. Negative Predictive Value (NPV): This is the proportion of true negative cases among all samples predicted as negative by the model, representing the probability that an individual predicted as non-PD is actually healthy. It is particularly useful for follow-up confirmation in screening34. False Negative Rate (FNR): This represents the proportion of positive samples incorrectly predicted as negative by the model, the proportion of actual PD patients misclassified as healthy. A lower FNR indicates a smaller missed diagnosis risk, making it an important indicator for evaluating the safety of early screening systems. Misdiagnosis control metrics Misdiagnosis refers to the misclassification of healthy individuals as PD patients, which may lead to non-patients being misled into unnecessary examinations or drug interventions. Key evaluation metrics for misdiagnosis include: Specificity (True Negative Rate, TNR): This metric represents the proportion of negative samples correctly predicted by the classification model among all negative samples. It reflects the model’s ability to identify healthy individuals and is suitable for scenarios where minimizing additional burdens from misdiagnosis is critical135. Precision: This denotes the proportion of true positive samples among those predicted as positive by the model, which reflects the probability that an individual predicted to have PD actually has the disease. It is applicable for controlling misdiagnosis risks during further diagnostic procedures. A trade-off exists between precision and sensitivity; if both are critical, the F1-score can be used for evaluation34. False positive rate (FPR): This indicates the proportion of negative samples incorrectly predicted as positive by the model, which means that it reflects the proportion of healthy individuals misclassified as PD patients. A lower FPR signifies a smaller misdiagnosis risk, making it particularly relevant for assessing the acceptability of preclinical screening135. Overall performance assessment metrics These metrics are used to measure a model’s predictive capability across overall samples, typically without bias toward any single class, making them suitable for comparative evaluation among multiple models—particularly in scenarios with class imbalance. Key indicators include: Accuracy: The proportion of correctly predicted samples among all instances136. Error rate (ERR): The proportion of incorrectly predicted samples among all instances. These two metrics are suitable for balanced datasets and are complementary to each other, with one typically chosen for reporting. F1-Score: This is the harmonic mean of precision and recall, used to comprehensively evaluate model performance, reflecting both the control of misdiagnoses and the prevention of missed diagnoses.74. Area under the receiver operating characteristic curve (AUC-ROC): Used to globally evaluate a model’s ability to distinguish between PD patients and healthy individuals. AUC-ROC is ideal for comparing the comprehensive discriminative power of different models36. Matthews correlation coefficient (MCC): Accounts for true positives, false positives, true negatives, and false negatives, making it suitable for imbalanced datasets and scenarios requiring holistic model performance analysis34. G-means and Kappa: Also applicable to imbalanced data. G-means focuses on balancing a model’s sensitivity and specificity, while Kappa excludes the influence of accidental agreement. These can be used interchangeably with F1-score, AUC, or MCC to further validate model robustness34. Regression error control metrics For PD studies targeting continuous variables such as risk probability or symptom scores as prediction objectives, regression error metrics should be employed. Correlation Coefficient: A statistical index measuring the strength and direction of the linear relationship between two variables36. The most commonly used is the Pearson correlation coefficient, which ranges between −1 and 1. It is suitable for assessing the correlation between model predictions and actual outcomes, particularly for evaluating the effectiveness of continuous variable prediction. Root mean square error (RMSE): A widely used metric for measuring the discrepancy between model predictions and actual observations, which is more sensitive to larger prediction errors and emphasizes the capability to control outliers74. Mean absolute error (MAE) is another indicator for predicting errors, calculating the average of absolute differences between predicted and actual values. In PD diagnosis, both RMSE and MAE are used to evaluate the magnitude of prediction errors, but RMSE is more sensitive to large errors, while MAE is more intuitive. Therefore, one of them is typically selected for reporting74. Standard Deviation (SD): A statistical measure of the dispersion of numerical values in a dataset. It quantifies the degree of deviation between data points and their mean. It is not directly used to assess model performance but can reflect the dispersion of data distributions, which may be helpful for comparing sample volatility. If only error evaluation is of concern, SD may not be necessary34. It is worth noting that most existing relevant studies take individual gait cycle samples as the analytical unit, whereas research and analysis focusing on independent human subjects as a whole remain relatively insufficient. In practical clinical settings, studies conducted at the level of independent subjects possess greater practical diagnostic and therapeutic value as well as clinical guiding significance, as they can more authentically reflect individual gait characteristics and disease status. Therefore, future research may further strengthen the analysis and exploration at the independent subject level, so as to provide more reliable evidence for clinical applications. Supplementary information Supplementary Information Supplementary Information