Electroencephalography‐Based Clustering Reveals Robust Neurophysiological Subtypes in Parkinson's Disease Authors: Daniel Vered, Zoya Katzir, Idan Daniel Grosbard, Avner Thaler, Inbal Maidan Published: 2026 ABSTRACT AbstractBackgroundParkinson's disease (PD) is clinically heterogeneous, with substantial variability in motor and cognitive features. Conventional clinical scales provide limited insight into underlying neural mechanisms and show poor longitudinal stability. Scalp electroencephalography (EEG) offers a direct and scalable measure of brain activity that may support mechanistically informed stratification of PD, yet EEG‐based subtyping remains largely unexplored.ObjectivesTo determine whether EEG‐derived features can identify neurophysiological subtypes of PD and to examine their clinical and genetic correlates.MethodsEEG recordings were obtained from 116 PD patients and 30 healthy controls during resting state and treadmill walking. Spectral power (theta–gamma), aperiodic components, and temporal complexity measures were extracted across cortical regions. Multiple dimensionality reduction and unsupervised clustering approaches were evaluated, with internal validity and stability assessed using leave‐one‐out resampling. Clinical, cognitive, gait, and genetic features were compared across clusters.ResultsA three‐cluster solution demonstrated optimal separation and robustness across models. The clusters exhibited distinct neurophysiological profiles characterized by differences in spectral slowing, signal complexity, and aperiodic activity across both behavioral states. Clusters did not differ in motor severity, gait speed, or disease duration, but showed differences in cognitive performance and genetic composition. One cluster demonstrated preserved cognition and included more LRRK2‐mutation carriers, whereas two cognitively impaired clusters displayed divergent electrophysiological signatures despite similar clinical profiles.ConclusionsEEG‐based clustering reveals robust neurophysiological subtypes of PD that capture heterogeneity not reflected by standard clinical measures. These findings support EEG as a scalable tool for mechanism‐based stratification, with implications for prognosis and precision clinical trials. © 2026 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.Keywords: Parkinson's disease, electroencephalography (EEG), neurophysiological subtypes, clustering analysis, neural heterogeneity The clinical and pathophysiological heterogeneity of Parkinson's disease (PD) is well recognized, with substantial variability in motor and cognitive manifestations across individuals. 1 Efforts to capture this variability have included both a priori subtype classifications based on features such as age at onset or motor phenotype (TD, tremor dominance or PIGD, postural instability gait difficulty), and data‐driven clustering approaches integrating broader clinical, motor, and non‐motor characteristics. 1 , 2 However, most of these frameworks rely on clinical rating scales, including the Movement Disorder Society‐Unified Parkinson's Disease Rating Scale (MDS‐UPDRS), 3 which reflect complex behavioral outcomes rather than the neural processes underlying symptom expression. 4 As a result, clinically defined groupings often show limited longitudinal stability and provide only indirect insight into disease mechanisms. 5 , 6 , 7 , 8 Neuroimaging studies have attempted to link clinical heterogeneity to underlying brain organization. Structural magnetic resonance imaging (MRI) studies have associated variability in PD with differences in regional atrophy and cognitive profiles, 9 , 10 , 11 while functional MRI studies have identified alterations in large‐scale networks related to motor and cognitive function. 12 , 13 . However, these studies have often focused on differentiating predefined clinical subtypes rather than characterizing interindividual variability in neural function in a data‐driven manner. In addition, structural imaging captures relatively static disease features, and functional MRI (fMRI) assessments are constrained by scanner environments that limit the study of naturalistic behaviors such as walking, that are central to functional impairment in PD. Electroencephalography (EEG) offers a complementary approach for addressing disease heterogeneity by directly measuring large‐scale neuronal activity with high temporal resolution. 14 Moreover, EEG can be recorded during naturalistic tasks such as walking, providing insight into the neural dynamics underlying both motor and cognitive impairments. Previous studies have identified EEG alterations in PD, including spectral slowing 15 , 16 , 17 , 18 and reduced signal complexity, 17 , 19 which have been associated with disease severity and cognitive impairment. However, the extent to which these electrophysiological features relate to interindividual differences in clinical presentation remains largely unexplored. To our knowledge, only one study utilized EEG‐based functional connectivity measures to data‐driven clustering in PD. 20 However, this work was limited by a small sample size, a single closed‐eyes resting‐state condition, and the absence of systematic validation and stability analyses necessary for assessing reproducibility and generalizability. 20 In this study, we systematically used EEG‐derived features to explore variability in large‐scale neural dynamics in PD patients. We leveraged a large dataset collected during both resting and walking conditions, extracting well‐established electrophysiological metrics including spectral, aperiodic, and temporal complexity measures. Using a comprehensive machine learning pipeline that evaluates multiple dimensionality reduction and clustering algorithms, we aimed to characterize patterns of electrophysiological variation and relate them to motor, cognitive, and genetic features. Data‐driven analyses were employed as an exploratory approach to assess whether individuals with similar clinical profiles exhibited distinct neural dynamics, thereby providing insight into the neural bases of phenotypic heterogeneity in PD. Methods Participants We enrolled 116 PD patients and 30 healthy controls (HC), recruited from the Laboratory of Early Markers of Neurodegeneration (LEMON) at Tel Aviv Sourasky Medical Center (TASMC). The study protocol was approved by the local ethics committee and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to participation. Patients were diagnosed with PD according to the updated MDS clinical diagnostic criteria. Inclusion criteria for PD patients were: (1) aged between 40 and 90 years; (2) a Hoehn and Yahr stage of 2 or less; 21 and (3) a Montreal Cognitive Assessment (MoCA) score of 21 or higher. 22 Importantly, patients presenting with clinical features suggestive of atypical parkinsonian syndromes were excluded during the diagnostic evaluation. Patients were assessed during their ‘ON’ medication state, approximately 1 hr after their last dose of dopaminergic medication, to minimize variability and ensure comparable conditions during EEG acquisition. Exclusion criteria included prior brain surgery, a history of stroke or significant head injury, recurrent loss of consciousness, other significant central nervous system disorders, or inability to walk independently for at least 5 min. Healthy controls were required to be aged between 40 and 90 years of age, community‐dwelling, and functionally independent. The exclusion criteria included any significant surgical, orthopedic, or psychiatric conditions that could interfere with the study protocol. Data Collection All participants underwent a comprehensive assessment protocol. We collected demographic and clinical measures, including age, sex, disease duration, MDS‐UPDRS, 3 and concurrent medication enabling the calculation of levodopa equivalent daily dose (LEDD). Cognitive performance was assessed using the MoCA as a measure of global cognition. The Color Trails Test (CTT) 23 was used to evaluate executive function, processing speed, visual attention, and set‐shifting. In addition, a dual‐task walking paradigm, in which participants performed serial‐7 subtractions, assessed divided attention, working memory, and executive control under increased attentional load. 24 Motor function was evaluated using two walking over‐ground tasks, 1 min of normal walking and 1 min of walking while performing serial‐7 subtractions (dual‐task), along a 20‐m corridor. In addition, genetic screening identified carriers of mutations in the GBA1 and LRRK2 genes. 25 EEG Acquisition and Preprocessing EEG data were acquired from 64 scalp electrodes positioned according to the international 10–20 system, during two distinct conditions: 4 min of seated resting‐state with eyes open and 4 min of walking on a treadmill. Preprocessing was performed using the standard EEGLAB toolbox in MATLAB. 26 Preprocessing steps included bandpass filtering (0.5–40 Hz), channel rejection, average re‐referencing, independent component analysis (ICA), and artifact correction. All preprocessing code can be found at EEG-Preprocessing-Toolkit. Feature Engineering From the preprocessed EEG, we extracted features across the frequency and time domains. Spectral features included relative power in canonical frequency bands (theta, alpha, beta, gamma), along with parametrizations of the aperiodic component of the power spectrum, estimated using the FOOOF algorithm. 27 In the time domain, signal complexity was assessed using permutation entropy (PermEn) 28 and Lempel–Ziv complexity (LZC). 29 For interpretability and dimensionality reduction, features were aggregated within 13 anatomical regions, defined by combinations of general cortical region (frontal, central, parietal, temporal, occipital) and hemisphere (left, midline, right). A full description of the mapping between electrodes and regions may be found in Table S1. Before clustering, all features were standardized across participants using Z‐scores and missing values were handled using mean imputation. Dimensionality Reduction To address the challenges of high‐dimensional feature sets, we first explored several dimensionality reduction techniques. We compared three dimensionality reduction approaches: clustering directly on the standardized features (vanilla), projecting the data into an orthogonal latent space using principal component analysis (PCA), and nonlinearly embedding the data into a low‐dimensional manifold using Uniform Manifold Approximation and Projection (UMAP). 30 The latent dimensionality of both PCA and UMAP, as well as other model hyperparameters, was tuned based on the data. Clustering Analysis We evaluated three unsupervised clustering approaches with different underlying assumptions. K‐Means was used as a baseline method, partitioning the data into K clusters by minimizing within‐cluster variance. 31 Gaussian mixture model (GMM) provided a probabilistic alternative, allowing clusters with different shapes and covariances. Finally, we tested density‐based spatial clustering of applications with noise (DBSCAN), 32 which identifies clusters of arbitrary shape while labeling low‐density points as noise. Final models were evaluated using several internal validity indices, including the Silhouette score, 33 Davies–Bouldin index (DBI), 34 and Calinski–Harabasz index (CH). 35 We assessed clustering robustness for both within‐model stability and between‐model agreement. Stability was evaluated using a leave‐one‐out resampling scheme with the adjusted Rand index (ARI), 36 while agreement was measured by pairwise ARI calculations on cluster assignments. For detailed descriptions of the training process and evaluation schemes see Supplementary Clustering analysis section and Tables S2–S4. Feature Importance We used a permutation‐based analysis to assess the relative importance of each neurophysiological feature. Importance was quantified by the drop in clusters separation (Δ Silhouette) and stability (ARI) after permuting the values of each feature. This analysis was conducted separately for the sitting and walking conditions, and results were compared across features and cortical regions. Clinical Profile To interpret the clinical meaning of the identified clusters, we compared groups using one‐way analysis of variance (ANOVA) on demographic, clinical, and genetic measures. Measures included age, disease duration, age at onset, motor severity (MDS‐UPDRS‐III), cognitive performance (MoCA, CTT), and LEDD. Furthermore, domain‐specific motor severity scores were computed using the TD/PIGD phenotyping framework. 37 In addition, genetic status (GBA1, LRRK2) was examined to explore potential enrichment of mutation carriers in specific clusters. Participants were genotyped for the G2019S‐LRRK2 mutation and for nine common mutations in the GBA1 gene: N370S and R496H, considered mild mutations; L444P, 84GG, IVS2 + 1G>A, V394L, and 370Rec, considered severe mutations; and E326K and T369M, considered risk variants. 38 These mutations account for over 96% of mutant alleles among Ashkenazi Jews. 39 Only heterozygous carriers were included in this study, while individuals with dual mutations (LRRK2 + GBA) or homozygous carriers were excluded due to their small sample size. Patients with no detectable mutations were classified as idiopathic PD (iPD). For simplicity, results are reported only for the K‐Means clusters. Complete results for the remaining models are detailed in the Supplementary Material. Neurophysiological Profile To profile neurophysiological differences, we compared the clusters within the original feature space. We assessed both global‐level differences using features averaged across all electrodes, and regional‐level differences by comparing the topographical distribution of feature values on the scalp. This approach allowed us to determine if the clusters diverged in overall brain dynamics or in specific, localized patterns. Statistical Testing We assessed group differences using one‐way ANOVA for numerical variables and chi‐squared (χ2) tests for categorical distributions. Following a significant ANOVA result, pair‐wise Tukey post‐hoc analysis was performed. To control for multiple comparisons, all resulting P‐values were corrected using the Benjamini–Hochberg False Discovery Rate (FDR) procedure. This correction was applied separately within each variable category: demographic (age, sex), motor (MDS‐UPDRS‐Total, MDS‐UPDRS‐III, single‐task gait speed, dual‐task gait speed, domain‐specific motor scores), cognitive (MoCA, CTT1, CTT2, number of S7 subtractions during walking), genetic (LRRK2/GBA1 mutation status), and disease characteristics (age at onset, disease duration, LEDD). Correction was applied separately to the resting‐state and active‐walking behavioral conditions. All data analysis and machine learning pipelines were implemented in Python using standard libraries (numpy, pandas, sklearn, matplotlib, seaborn, scipy, statsmodels, and MNE). Permutation entropy was calculated using Antropy, 40 while LZC was calculated using EntropyHub. 41 The only exception was the EEG preprocessing, which was conducted in MATLAB. All code and detailed documentation are available in our public repository (EEG-basedPD-Subtyping). Results Participants' Characteristics We included 116 PD patients and 30 HC who completed motor, cognitive, gait, and EEG assessments. As shown in Table 1, the PD group was older than the HC group, demonstrated lower cognitive performance as measured by the MOCA, CTT1, and CTT2, and had lower gait speed. As expected, they had higher MDS‐UPDRS scores compared with the HC. TABLE 1Patient characteristicsCharacteristicVariableHC (n=30)PD (n=116) P‐value DemographicsAge (years) 60.82±7.47 65.97±9.5 0.006 Sex (M/F) 14/16 72/44 0.126CognitionMoCA 27.44±1.96 25.17±3.12 0.002 CTT1 (s) 45.61±14.2 65.93±36.51 0.006 CTT2 (s) 84.84±29.66 116.14±54.5 0.006 S7 subtraction walking 13.67±5.86 12.61±7.76 0.510GaitSingle‐task walk (m/s) 1.25±0.18 1.05±0.19 <0.001 Dual‐task walk (m/s) 1.15±0.18 0.88±0.22 <0.001 PD characteristicsMDS‐UPDRS Total 3.14±3.33 40.16±20.47 <0.001 MDS‐UPDRS‐Part III 1.33±1.73 24.7±12.44 <0.001 TD score NA 0.41±0.36 NA PIGD score NA 0.5±0.44 NA Disease duration (months) NA 58.34±47.03 NA Age at onset (years) NA 61.11±9.3 NA LEDD (mg) NA 418.6±383.3 NA Genetic mutations (Idiopathic/LRRK2/GBA1) NA 73/25/18 NA Note: Values in bold type indicate statistical significance.Abbreviations: HC, healthy control, PD, Parkinson's disease; M, male; F, female; MoCA, Montreal Cognitive Assessment; CTT, Color Trails Test; S7, Serial 7s subtraction; MDS‐UPDRS, Movement Disorder Society‐Unified Parkinson's Disease Rating Scale; TD, tremor‐dominance; NA, not available; PIGD, postural instability gait difficulty; LEDD, levodopa equivalent daily dose. Patient characteristics Genetic mutations (Idiopathic/LRRK2/GBA1) Note: Values in bold type indicate statistical significance. Abbreviations: HC, healthy control, PD, Parkinson's disease; M, male; F, female; MoCA, Montreal Cognitive Assessment; CTT, Color Trails Test; S7, Serial 7s subtraction; MDS‐UPDRS, Movement Disorder Society‐Unified Parkinson's Disease Rating Scale; TD, tremor‐dominance; NA, not available; PIGD, postural instability gait difficulty; LEDD, levodopa equivalent daily dose. Clustering Performance After hyperparameter optimization within each clustering algorithm, three final models were retained for comparison: K‐Means, GMM, and DBSCAN. In all clustering algorithms, UMAP reduction provided better results compared with PCA or vanilla. For K‐Means and GMM, a three‐cluster solution yielded the best performance (K‐Means: n0=41,n1=40,n2=35; GMM: n0=47,n1=36,n2=33), whereas DBSCAN converged to two clusters (n0=83,n1=26) with seven subjects classified as noise. Internal validity scores are summarized in Table S5. Among the models, K‐Means achieved the strongest overall separation, with superior Silhouette and CH scores and a lower DBI. Clustering Robustness Clustering stability was evaluated by using leave‐one‐out resampling and quantified by the ARI. Within‐model robustness was highest for K‐Means (ARI=0.77±0.15), followed by DBSCAN (ARI=0.72±0.23) and GMM (ARI=0.57±0.17). These findings indicate that the K‐Means solution was the most consistent across repeated subsamples of the cohort. Between‐model agreement was evaluated by pairwise comparisons of clustering solutions. Agreement was highest between K‐Means and GMM (ARI=0.69±0.15), whereas agreement was weaker between GMM and DBSCAN (ARI=0.38±0.12) and between K‐Means and DBSCAN (ARI=0.37±0.1). These results indicate moderate to high agreement between models, showing that although patient assignments are not identical, the resulting groupings are highly similar and reflect a consistent underlying structure in the data (Fig. S1). Clinical Profiles Given the similarity between the models, we present the results based on the K‐Means analysis (Fig. 1 and Table S6). The complete results for the GMM and DBSCAN analysis are provided in Figures S2, S3 and Tables S7 and S8. No significant differences were observed between clusters in age (F=1.34;P=0.532;η2=0.02), sex distribution (χ2=0.1;P=0.952;V<0.1), disease duration (F=1.25;P=0.325;η2=0.02), and age at onset (F=1.14;P=0.325;η2=0.02). No significant differences were found in LEDD (F=1.37;P=0.325;η2=0.02), MDS‐UPDRS‐Total (F=2.15;P=0.364;η2=0.04), MDS‐UPDRS‐III (F=2.32;P=0.364;η2=0.04), and gait speed in both single (F=0.13;P=0.915;η2<0.01) and dual task (F=0.23;P=0.915;η2<0.01) conditions. In addition, both TD (F=0.09;P=0.915;η2<0.01) and PIGD (F=0.26;P=0.915;η2<0.01) were comparable between clusters. Clinical profiles of K‐Means clusters. *Significant difference between PD clusters. MoCA and S7 values were negated so that lower values indicate better performance, to match the directionality of other measures. DD, disease duration; PD, Parkinson's disease; LEDD, levodopa equivalent daily dose; UPDRS‐III, Unified Parkinson's Disease Rating Scale‐Part III; MoCA, Montreal Cognitive Assessment; CTT, Color Trails Test; S7, Serial 7 s subtraction during walking; TD, tremor‐dominance; PIGD, postural instability gait difficulty. [Color figure can be viewed at wileyonlinelibrary.com] In contrast, cognitive measures across different domains differed between the clusters. Differences were observed in MoCA (F=3.62;P=0.04;η2=0.06), CTT Part 1 and 2 (CTT1: F=5.86;P=0.008;η2=0.10, CTT2: F=6.94;P=0.006;η2=0.12), and in the number of serial‐7 subtractions during walking (F=3.32;P=0.04;η2=0.06). Cluster PD‐2 exhibited better cognitive performance across all measures compared with the other clusters. Genetic composition differed significantly across clusters (χ 2 = 9.22, P = 0.055, V = 0.2), with the highest proportion of LRRK2 mutation carriers observed in PD‐2, intermediate representation in PD‐0, and the lowest in PD‐1. The proportion of GBA1 mutation carriers was comparable across clusters. Among the 18 patients carrying GBA1 mutations, 14 were carriers of the N370S variant, one carried the 84GG variant, one the E326K variant, and two carried other mutations. Neurophysiological Profiles Spectral power distributions differed significantly between clusters across all relative canonical frequency bands under both behavioral conditions (P < 0.001). During resting state, in absolute frequency bands (Fig. 2A), Cluster PD‐1 showed pronounced spectral slowing, characterized by elevated theta and reduced alpha and beta power compared with the other clusters and HC (Figs 2A and 3A). Cluster PD‐2 displayed low theta and high alpha power, whereas Cluster PD‐0 demonstrated low theta and alpha but elevated beta and gamma power, compared with the other clusters and HC (Figs 2A and 3A). During walking, differences were observed only in relative power: both PD‐0 and PD‐2 exhibited higher beta power compared with HC, whereas PD‐1 showed lower beta power (Figs 2B and 3B). Topographical inspection indicated that these effects reflected global power shifts rather than spatial reorganization, with broadly similar scalp distributions across clusters (Fig. 3). Neurophysiological profiles of K‐Means clusters during resting state and active walking. (A) Power spectral densities during resting state and (B) during active walking. HC, healthy control; PD, Parkinson's disease. [Color figure can be viewed at wileyonlinelibrary.com] Topological distributions of relative powers in canonical frequency bands across K‐Means clusters (A) during resting‐state and (B) during active walking. PD, Parkinson's disease; HC, healthy control. [Color figure can be viewed at wileyonlinelibrary.com] Analysis of the aperiodic exponent revealed significant differences between the clusters across conditions (P<0.001). In both conditions, Cluster PD‐0 exhibited lower aperiodic exponent values relative to HC, whereas Clusters PD‐1 and PD‐2 showed higher or comparable values (Table S9). Topographical maps indicated that these differences reflected global shifts in the spectral slope rather than distinct spatial distributions (Fig. S7). Measures of temporal complexity further differentiated the clusters (Table S9). Both LZC and PermEn revealed significant group effects across behavioral conditions (P<0.001). Cluster PD‐1 consistently exhibited the lowest complexity values, Cluster PD‐0 the highest, and Cluster PD‐2 intermediate. This relative ordering was observed across conditions, with slightly attenuated effect sizes during walking. Topographical patterns indicated that these effects reflected global variations in signal complexity rather than distinct spatial reorganizations (Fig. S7). The neurophysiological profiles of GMM clusters, including canonical frequency bands, aperiodic, and complexity measures, are presented in Figure S5 and Table S10. The corresponding profiles for DBSCAN clusters are presented in Figure S6 and Table S11. Feature Importance Permutation analysis revealed higher overall feature importance during the sitting condition, reflected by a greater mean separation drop (ΔSilhouette=0.036±0.014) (Fig. 4) but similar stability (ARI=0.86±0.05) (Fig. S8) compared with walking (ΔSilhouette=0.02±0.016;ARI=0.88±0.05). During sitting, the greatest importance was observed for relative band power in the theta, alpha, and beta bands, followed by signal‐complexity measures and gamma power, whereas the aperiodic exponent contributed least to the separation. During walking, overall importance values were reduced and the ordering shifted, with beta band power showing the largest effect, followed by theta and alpha. Across both conditions, feature‐importance values were consistent across cortical regions, showing no systematic regional differences (SDΔSilhouette<0.001). Permutation‐based feature importance for K‐Means clustering across conditions and features. Separation loss quantified by the Silhouette score after permutation (higher values indicate better separation). *P < 0.05, **P < 0.01, ***P < 0.001. LZC, Lempel–Ziv complexity; PermEn, permutation entropy. [Color figure can be viewed at wileyonlinelibrary.com] Discussion In this study, we investigated whether electrophysiological features derived from scalp EEG could reveal distinct neurophysiological groupings of PD patients. Our analyses identified three robust clusters (PD‐0, PD‐1, and PD‐2), each characterized by unique neurophysiological profiles across two behavioral states, resting and walking, including distinct patterns of oscillatory power, aperiodic activity, and signal complexity. Importantly, these EEG‐based clusters were associated with differences in cognitive performance and genetic background. This electrophysiology‐based stratification suggests that EEG can capture disease heterogeneity beyond traditional clinical classification (eg, PIGD, TD), reflecting both cognitive‐related slowing and motor‐related beta‐band alterations. 15 , 16 , 17 , 18 , 19 , 42 , 43 Clinical and Genetic Correlates of Neurophysiological Subtypes The main clinical difference across the identified neurophysiological subtypes was in cognition. Beyond global cognition measured by the MoCA, clusters also differed in executive function, cognitive flexibility, processing speed, and divided attention, as assessed by the CTT and dual‐task walking. Cluster PD‐2 demonstrated relatively preserved performance across all cognitive measures, whereas PD‐0 and PD‐1 showed consistent impairments. The consistency of these findings across multiple cognitive domains strengthens the interpretation that the identified subtypes capture meaningful cognitive heterogeneity in PD. Importantly, despite their similar cognitive profiles, PD‐0 and PD‐1 exhibited distinct electrophysiological patterns, suggesting that the neurophysiological differences between subtypes cannot be explained solely by cognitive status. The cognitive profiles also aligned with the distribution of LRRK2 mutation carriers, who are typically characterized by relatively preserved cognition. 44 Cluster PD‐2 included a higher proportion of LRRK2 carriers, consistent with previous reports and extending them by demonstrating associated electrophysiological features, including reduced EEG slowing and relative shift toward higher alpha compared with theta power. 44 In contrast, the proportion of GBA1 mutation carriers did not differ between clusters, suggesting that the observed cognitive and neurophysiological distinctions are not specifically linked to GBA1 mutations. However, the relatively small number of GBA1 carriers limits firm conclusions, and larger cohorts will be required to clarify potential genotype‐specific neurophysiological patterns. Neuropsychological Characteristics The neurophysiological differences distinguishing the clusters were driven by distinct spectral and signal‐complexity patterns. These alterations were global rather than region‐specific, as reflected by the absence of consistent topographical differences and the lack of regional specificity in feature‐importance analyses. This suggests that large‐scale network activity, rather than focal cortical changes, underlies the observed group differences and indicates that such distinctions may be detectable even with reduced‐channel clinical EEG systems. Cluster PD‐2, which showed preserved cognition, was characterized by less spectral slowing, higher alpha relative to theta power, higher signal complexity, and a less steep aperiodic exponent. These features suggest more preserved cortical dynamics that may support the better cognitive function observed in this cluster within the PD cohort. 45 In contrast, the two clusters with lower cognitive performance showed divergent electrophysiological signatures. Cluster PD‐1 exhibited pronounced spectral slowing, reduced alpha, lower temporal complexity, and a steeper aperiodic exponent, a profile commonly associated with cognitive decline. Beta activity in PD‐1 was relatively low across states, including during walking, suggesting attenuation of the typical motor‐related beta modulation observed in the other PD clusters (PD‐0 and PD‐2). 46 Cluster PD‐0 showed a different spectral profile, with comparatively higher temporal complexity and a flatter power spectral density characterized by lower theta and alpha but higher beta and gamma power. These findings indicate that similar cognitive performance in PD‐0 and PD‐1 may arise from distinct underlying neural dynamics. Future studies should examine the long‐term clinical and electrophysiological trajectories of these clusters to evaluate their stability over time and their potential to inform distinct prognostic and mechanistic profiles. Methodological Considerations and Robustness This study provides a framework for stratifying PD using neurophysiological signatures. Cluster stability was confirmed across subsamples and across different clustering algorithms, supporting the robustness of the identified subgroups. Feature‐importance analyses indicated that spectral power contributed most strongly to cluster differentiation, whereas aperiodic features had a comparatively smaller influence in this dataset. 27 An additional aspect of this study was the inclusion of dynamic physiological data across multiple behavioral states, including actual walking, which helped capture functional heterogeneity not evident in resting‐state measures alone. Implications and Interpretation The identification of three robust EEG‐based clusters, defined by a combination of spectral, complexity, and aperiodic EEG features, highlights the potential value of incorporating neurophysiology into PD characterization beyond traditional motor‐based diagnosis. 37 , 47 Notably, EEG distinguished between the two cognitively impaired clusters, PD‐0 and PD‐1, despite similar clinical scores. This divergence suggests that comparable behavioral outcomes may arise from different neural mechanisms, which are not captured by standard clinical assessments. 48 Importantly, we do not interpret these findings as evidence for novel clinical subtypes of PD. Rather, they represent data‐driven neurophysiological clusters that may provide an additional dimension for patient stratification. Longitudinal and external validation studies will be necessary to determine their stability, clinical relevance, and potential prognostic value. Conclusions In summary, scalp EEG enabled the identification of distinct neurophysiological clusters in PD that were associated with cognitive performance and genetic background. Notably, two clusters (PD‐0 and PD‐1) showed comparable levels of cognitive impairment yet displayed markedly different electrophysiological profiles, indicating that the identified groupings reflect underlying neural differences beyond cognitive severity alone. These findings demonstrate that electrophysiological measures can capture clinically meaningful heterogeneity across resting and walking states. However, these clusters should be considered preliminary and require replication in independent cohorts before they can be interpreted as specific disease subtypes or incorporated into clinical frameworks. Overall, this work supports the integration of EEG‐based measures as complementary tools for characterizing PD heterogeneity and guiding future mechanistic investigations. Longitudinal studies will be essential to determine how these neurophysiological patterns evolve with disease progression. Funding The study was funded by the Israel Science Foundation (ISF) (Grant No. 1157/20). Author Roles (1) Research Project: A. Conception, B. Organization, C. Execution; (2) Statistical Analysis: A. Design, B. Execution, C. Review and Critique; (3) Manuscript Preparation: A. Writing of the First Draft, B. Review and Critique. D.V.: 1B, 1C, 2A, 2B, 3A. Z.K.: 1B, 1C, 3B. I.D.G.: 2C, 3B. A.T.: 1A, 3B. I.M.: 1A, 2B, 2C, 3A, 3B. Financial Disclosures I.M. reports receiving grant support from The Michael J. Fox Foundation (MJFF) that has no conflict with the presented work. All the other authors have no financial disclosures. Conflicts of interest The authors report no conflicts of interest. Supporting information Table S1. Mapping electroencephalography electrodes into cortical brain regions. Table S2. Hyperparameter grid for dimensionality reduction models. Table S3. Hyperparameter grid for clustering models. Table S4. Best model hyperparameters within each clustering algorithm. Table S5. Internal validity indices under leave‐one‐out evaluation (mean ± standard deviation). Table S6. Clinical characteristics (mean ± standard deviation) for K‐Means clusters. Table S7. Clinical characteristics (mean ± standard error) for Gaussian mixture model (GMM) clusters. Table S8. Clinical characteristics (mean ± standard error) for density‐based spatial clustering of applications with noise (DBSCAN) clusters. Table S9. Non‐oscillatory neurophysiological features (mean ± standard deviations) between K‐Means clusters. Table S10. Non‐oscillatory neurophysiological features (mean ± standard deviations) between Gaussian mixture model (GMM) clusters. Table S11. Non‐oscillatory neurophysiological features (mean ± standard deviations) between density‐based spatial clustering of applications with noise (DBSCAN) clusters. Figure S1. Clustering robustness. Patient assignment flow between clusters across different clustering models (A) between K‐Means and density‐based spatial clustering of applications with noise (DBSCAN) models and (B) between Gaussian mixture model (GMM) and DBSCAN models. Link width represents the number of subjects transitioning between corresponding clusters, highlighting agreement, splitting, and noise assignment across models. Figure S2. Clinical profiles of Gaussian mixture model (GMM) clusters. *Significant difference between clusters. MoCA and S7 values were negated so that lower values indicate better performance, to match the directionality of other measures. MoCA, Montreal Cognitive Assessment; S7, Serial 7s subtraction during walking; CTT, Color Trails Test; UPDRS‐III, Unified Parkinson's Disease Rating Scale‐Part III; TD, tremor‐dominance; PIGD, postural instability gait difficulty; LEDD, levodopa equivalent daily dose. Figure S3. Clinical profiles of density‐based spatial clustering of applications with noise (DBSCAN) clusters. *Significant difference between clusters. MoCA and S7 values were negated so that lower values indicate better performance, to match the directionality of other measures. MoCA, Montreal Cognitive Assessment; S7, Serial 7s subtraction during walking; CTT, Color Trails Test; UPDRS‐III, Unified Parkinson's Disease Rating Scale‐Part III; TD, tremor‐dominance; PIGD, postural instability gait difficulty; LEDD, levodopa equivalent daily dose. Figure S4. Neurophysiological profiles of K‐Means clusters during resting state and active walking. (A) Relative power in canonical frequency bands (mean ± standard error) during resting state and (B) during active walking. Figure S5. Neurophysiological profiles of Gaussian mixture model (GMM) clusters during resting state and active walking. (A) Relative power in canonical frequency bands (mean ± standard error) during resting state, (B) during active walking, (C) power spectral densities (PSDs) during resting state, and (D) during active walking. Figure S6. Neurophysiological profiles of density‐based spatial clustering of applications with noise (DBSCAN) clusters during resting state and active walking. (A) Relative power in canonical frequency bands (mean ± standard error) during resting state, (B) during active walking, (C) power spectral densities (PSDs) during resting state, and (D) during active walking. Figure S7. Topological distributions of aperiodic and complexity features across K‐Means clusters during (A) resting state and (B) active walking. LZC, Lempel–Ziv complexity; PermEn, permutation entropy. Figure S8. Permutation‐based feature importance for K‐Means clustering across conditions and features. Clustering stability quantified by the adjusted Rand index (ARI) between raw and permuted data (lower values indicate higher stability). *P < 0.05, **P < 0.01, ***P < 0.001. LZC, Lempel–Ziv complexity; PermEn, permutation entropy. Acknowledgments We thank the research participants and the research assistants for their time and effort. Data Availability Statement The data used in this study are publicly available from OpenNeuro (https://doi.org/10.18112/openneuro.ds007526.v1.0.0). All code and detailed documentation are available in our public repository (EEG-basedPD-Subtyping). References Associated Data Supplementary Materials Table S1. Mapping electroencephalography electrodes into cortical brain regions. Table S2. Hyperparameter grid for dimensionality reduction models. Table S3. Hyperparameter grid for clustering models. Table S4. Best model hyperparameters within each clustering algorithm. Table S5. Internal validity indices under leave‐one‐out evaluation (mean ± standard deviation). Table S6. Clinical characteristics (mean ± standard deviation) for K‐Means clusters. Table S7. Clinical characteristics (mean ± standard error) for Gaussian mixture model (GMM) clusters. Table S8. Clinical characteristics (mean ± standard error) for density‐based spatial clustering of applications with noise (DBSCAN) clusters. Table S9. Non‐oscillatory neurophysiological features (mean ± standard deviations) between K‐Means clusters. Table S10. Non‐oscillatory neurophysiological features (mean ± standard deviations) between Gaussian mixture model (GMM) clusters. Table S11. Non‐oscillatory neurophysiological features (mean ± standard deviations) between density‐based spatial clustering of applications with noise (DBSCAN) clusters. Figure S1. Clustering robustness. Patient assignment flow between clusters across different clustering models (A) between K‐Means and density‐based spatial clustering of applications with noise (DBSCAN) models and (B) between Gaussian mixture model (GMM) and DBSCAN models. Link width represents the number of subjects transitioning between corresponding clusters, highlighting agreement, splitting, and noise assignment across models. Figure S2. Clinical profiles of Gaussian mixture model (GMM) clusters. *Significant difference between clusters. MoCA and S7 values were negated so that lower values indicate better performance, to match the directionality of other measures. MoCA, Montreal Cognitive Assessment; S7, Serial 7s subtraction during walking; CTT, Color Trails Test; UPDRS‐III, Unified Parkinson's Disease Rating Scale‐Part III; TD, tremor‐dominance; PIGD, postural instability gait difficulty; LEDD, levodopa equivalent daily dose. Figure S3. Clinical profiles of density‐based spatial clustering of applications with noise (DBSCAN) clusters. *Significant difference between clusters. MoCA and S7 values were negated so that lower values indicate better performance, to match the directionality of other measures. MoCA, Montreal Cognitive Assessment; S7, Serial 7s subtraction during walking; CTT, Color Trails Test; UPDRS‐III, Unified Parkinson's Disease Rating Scale‐Part III; TD, tremor‐dominance; PIGD, postural instability gait difficulty; LEDD, levodopa equivalent daily dose. Figure S4. Neurophysiological profiles of K‐Means clusters during resting state and active walking. (A) Relative power in canonical frequency bands (mean ± standard error) during resting state and (B) during active walking. Figure S5. Neurophysiological profiles of Gaussian mixture model (GMM) clusters during resting state and active walking. (A) Relative power in canonical frequency bands (mean ± standard error) during resting state, (B) during active walking, (C) power spectral densities (PSDs) during resting state, and (D) during active walking. Figure S6. Neurophysiological profiles of density‐based spatial clustering of applications with noise (DBSCAN) clusters during resting state and active walking. (A) Relative power in canonical frequency bands (mean ± standard error) during resting state, (B) during active walking, (C) power spectral densities (PSDs) during resting state, and (D) during active walking. Figure S7. Topological distributions of aperiodic and complexity features across K‐Means clusters during (A) resting state and (B) active walking. LZC, Lempel–Ziv complexity; PermEn, permutation entropy. Figure S8. Permutation‐based feature importance for K‐Means clustering across conditions and features. Clustering stability quantified by the adjusted Rand index (ARI) between raw and permuted data (lower values indicate higher stability). *P < 0.05, **P < 0.01, ***P < 0.001. LZC, Lempel–Ziv complexity; PermEn, permutation entropy. Data Availability Statement The data used in this study are publicly available from OpenNeuro (https://doi.org/10.18112/openneuro.ds007526.v1.0.0). All code and detailed documentation are available in our public repository (EEG-basedPD-Subtyping).