Predicting kinesiophobia in knee osteoarthritis: a head-to-head comparison between machine learning and traditional regression models Authors: Xintong Li, Xiaoyu Huang, Yabing Han, Qianyun Zhang, Shigang Mao, Taolin Zhang, Yang Yu, Junjie Jiang Published: 2026 ABSTRACT AbstractBackgroundAlthough machine learning (ML) holds promise for improving clinical identification, its application to kinesiophobia in knee osteoarthritis (KOA) is limited. Therefore, this study aimed to develop and compare logistic regression (LR)-based nomograms with ML models to determine the optimal approach for identifying kinesiophobia status and guiding personalized interventions.MethodsA cross-sectional study enrolled 590 KOA patients from a tertiary hospital (June 2024 to July 2025) and randomly divided them into training (n = 413) and test (n = 177) sets in a 7:3 ratio. An interpretable LR-based nomogram was developed for assessing kinesiophobia status. Seven ML models (CatBoost, LightGBM, ML-LR, random forest, support vector machine [SVM], extreme gradient boosting, multilayer perceptron) were constructed to compare predictive performance. Models were evaluated via receiver operating characteristic curves with the area under the curve (AUC), calibration curves, and decision curve analysis; Shapley Additive exPlanations (SHAP) was utilized to interpret the ML models.ResultsNo significant differences in baseline characteristics were found between the two sets. LR identified five key predictors: age, educational level, pain intensity, pain catastrophizing, and activity level. The nomogram demonstrated good discrimination (training AUC = 0.844; test AUC = 0.815). Among ML models, SVM exhibited a marginally higher test-set AUC (0.821). SHAP analysis confirmed pain intensity and pain catastrophizing as the strongest predictors in the SVM model.ConclusionsAmong the evaluated ML models, no single model demonstrated universal superiority for classifying current kinesiophobia status in KOA. These models offer complementary strengths. SVM exhibited marginally higher predictive performance, warranting further exploration and potential integration into automated digital health platforms, whereas the LR-based nomogram performed comparably with favorable interpretability for rapid, transparent bedside assessment in primary care. Model selection should be guided by specific clinical contexts.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12891-026-10068-9.Keywords: Knee osteoarthritis, Kinesiophobia, Machine learning, Logistic models, Nomogram, Prediction model, Digital health Background Knee osteoarthritis (KOA) imposes a substantial burden on global health, severely impairing patients’ daily functioning and consequently generating significant pressure on healthcare systems [1, 2]. Within KOA rehabilitation, numerous challenges exist, among which kinesiophobia holds particular importance. Kinesiophobia, defined as an excessive, fear-driven avoidance of movement triggered by anticipated injury or reinjury, frequently leads to muscle weakness, joint instability, accelerated KOA progression, and suboptimal prognosis [3–5]. Early identification of kinesiophobia by healthcare providers facilitates the development of customized interventions and enables careful assessment of its severity and implications for KOA management. Conversely, delayed detection may result in disabilities that not only diminish patients’ quality of life but also adversely affect overall KOA outcomes [6]. In current clinical practice, the Tampa Scale for Kinesiophobia (TSK) is the predominant tool for kinesiophobia assessment, valued for its well-established reliability and ease of implementation [7, 8]. However, as a unidimensional measure focused on fear, it lacks the capacity to integrate multifactorial clinical data (e.g., pain, function, psychological factors) into a unified risk score to guide point-of-care decisions. Similarly, traditional modeling approaches like logistic regression (LR) rely on linear assumptions that may fail to capture the complex, often nonlinear interactions among these biopsychosocial factors, thereby limiting their predictive accuracy and generalizability across diverse KOA populations [9]. These limitations highlight the urgent need for alternative approaches, particularly computational advances deployable on digital health platforms for population-level risk stratification. These limitations and the urgent need for alternative approaches have spurred research into machine learning (ML), making it a promising solution. ML’s unique capacity to model nonlinear relationships and process high-dimensional data offers transformative potential in osteoarthritis research, encompassing domains such as early diagnosis, progression prediction, and phenotype identification [10–13]. Integrating Shapley Additive exPlanations (SHAP) represents a crucial step toward enhancing model transparency. This approach effectively addresses ML’s well-known “black-box” problem, thereby strengthening the models’ utility in clinical practice [14]. Despite these advances, critical knowledge gaps remain. Previous studies developing ML models for kinesiophobia assessment have provided technical insights, but they primarily focused on post-surgical cohorts rather than patients with KOA [15]. More importantly, no research has systematically compared LR and ML models for classifying current kinesiophobia in KOA patients, particularly concerning their potential for integration into clinical digital platforms and population-level screening initiatives. Therefore, this study aimed to (1) compare the predictive performance of LR-based nomograms and ML models for kinesiophobia in KOA patients, and (2) determine which approach is more clinically applicable for early identification of high-risk individuals. The findings could inform more targeted management strategies, contribute to improving rehabilitation outcomes in KOA, and guide the development of scalable, evidence-based interventions for this prevalent condition. Methods Design This cross-sectional study recruited KOA patients from the outpatient and inpatient departments of a tertiary hospital between June 2024 and July 2025. Participants who met the inclusion criteria were randomly assigned (using a computer-generated sequence, 7:3 ratio) to a training set and a test set. Thus, all predictive terms denote concurrent status, not future events. Participants Inclusion criteria were: (1) age 40–80 years; (2) diagnosis of KOA meeting established clinical criteria [16], defined as recurrent knee pain within the past month plus ≥ 2 of the following: (a) X-ray evidence of joint space narrowing, subchondral bone sclerosis and/or cysts, and marginal osteophytes; (b) age ≥ 50 years; (c) morning stiffness ≤ 30 min; (d) palpable or audible crepitus; (3) voluntary participation with signed informed consent. Exclusion criteria included: (1) knee pain caused by rheumatic or other inflammatory diseases; (2) a history of knee surgery or long-term (≥ 3 months) corticosteroid use; (3) severe comorbidities affecting major organs or life-threatening systemic conditions (e.g., severe heart failure [New York Heart Association class III/IV], end-stage renal disease, decompensated cirrhosis, or active malignancy); (4) mental or cognitive impairments precluding cooperation with study procedures. This study was approved by the Ethics Committee of Qingdao Municipal Hospital (2024-KY-083). All procedures were conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent prior to initiating study procedures. Research tools and data collection Data collection tools included a self-designed general information questionnaire and validated scales assessing kinesiophobia, pain, psychological status, joint function, physical activity, and social support (Supplementary Material 1). The general information questionnaire was developed to collect demographic characteristics and disease-related information. It was constructed based on a review of pertinent literature and expert consultation, and its clarity was refined through pilot testing with the target population. Demographic characteristics included gender, age, body mass index, education (junior high and below, senior high/technical school, junior college and above), and marital status (married, or unmarried [including divorced or widowed]). Disease-related information included medical expense payment methods (categorized as urban employee medical insurance, urban resident medical insurance, or others), average monthly household income with a cutoff of USD 700, affected side (left, right, bilateral), disease duration (from KOA diagnosis to enrollment), and the type and number of comorbid chronic diseases (e.g., hypertension, diabetes, coronary heart disease). Kinesiophobia was measured with the TSK-17 (scores 17–68, > 37 clinically significant) [17]. Pain intensity was assessed via the Numerical Rating Scale (NRS) [18], and catastrophizing with the Pain Catastrophizing Scale (PCS; 0–52, ≥ 30 clinically significant) [19]. The Hospital Anxiety and Depression Scale (subscales 0–21, ≥ 11 indicative of anxiety/depression) screened psychological status [20]. Joint function, physical activity, and social support were evaluated using the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC; 0–96) [21], the University of California, Los Angeles (UCLA) Activity Scale (1–10) [22], and the Multidimensional Scale of Perceived Social Support (12–84) [23], respectively. Higher scores consistently indicated greater severity or levels of the respective traits. To ensure data quality, two researchers independently screened the dataset for outliers, missing values, and logical inconsistencies. Any identified discrepancies were adjudicated through medical record review or telephone verification. Data analysis Analyses were conducted using SPSS Version 27.0 (IBM Corporation, Chicago, United States) and R software Version 4.5.0 (R Foundation for Statistical Computing, Vienna, Austria). Data preprocessing involved median imputation for variables with < 5% missing values, multiple imputation for those with 5%–20% missing values, and exclusion of variables with > 20% missing values or low clinical relevance. Outliers were identified via box plots and the Interquartile Range (IQR) method (Q1-1.5IQR to Q3 + 1.5IQR); extreme values (per clinical judgment) were excluded, and data-entry errors corrected via Winsorization (replaced with 1% and 99% percentiles). Continuous data normality was tested via the Shapiro-Wilk test: normally distributed data are presented as mean ± SD (x̄ ± s), non-normal as median (IQR) [M (Q₁, Q₃)]. Group comparisons used independent t-tests or Mann-Whitney U tests (continuous) and χ² or Fisher’s exact tests (categorical). Predictors were first filtered via univariate analysis (p < 0.10) and included in binary LR; for educational level (categorical), “junior college and above” was the reference. Independent predictors (p < 0.05) were used to construct the nomogram. Model evaluation included traditional statistical and ML models. For traditional models, discrimination was evaluated using receiver operating characteristic (ROC) curves with the area under the curve (AUC) as the metric, and the benchmarks were defined as ≤ 0.5 (random), 0.5–0.7 (low), 0.7–0.9 (high), and > 0.9 (excellent). Calibration was assessed using bootstrap-corrected calibration curves, and clinical utility via decision curve analysis (DCA). For ML models, seven algorithms were used: CatBoost, LightGBM, ML-LR, random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP). Models were tuned using ten-fold cross-validation. Performance was compared using AUC, accuracy, sensitivity, specificity, positive predictive values, negative predictive values and F1 score; calibration curves and DCA further validated utility. Feature importance was visualized via SHAP values. Two-tailed p < 0.05 defined significance. Results Baseline characteristics of participants Among 648 patients assessed for eligibility, 22 were excluded for clinical reasons and 36 for extensive missing data (> 20%), yielding 590 patients in the final analysis (Fig. S1, Supplementary Material 2). A subsequent clinical review of potential outliers led to no further exclusions but identified two data-entry errors, and these were corrected. Patients were randomly allocated to training (n = 413) and testing (n = 177) sets in a 7:3 ratio. No significant differences were observed in baseline characteristics between the sets (all p > 0.05; Table S1, Supplementary Material 2). Development and validation of a nomogram with the LR model Screening of predictors Univariate analysis of the training set (n = 413) revealed six variables significantly associated with kinesiophobia (Table 1). A subsequent binary LR model, incorporating variables significant at p < 0.10, confirmed five independent predictors (all p < 0.05; Table 2). Increasing age (OR 1.099, 95% CI 1.065–1.133), lower educational attainment (OR 2.206 for junior high and below vs. reference), and higher scores on the NRS (OR 1.399, 95% CI 1.179–1.660) and PCS (OR 1.088, 95% CI 1.049–1.128) all increased the odds of kinesiophobia. Conversely, a higher UCLA Activity Score was associated with a protective effect (OR 0.690, 95% CI 0.584–0.816). The model demonstrated good fit (Hosmer-Lemeshow χ² = 6.492, p = 0.592) with no evidence of multicollinearity (all variance inflation factors < 2.0). Table 1Comparison of characteristics between KOA patients with and without kinesiophobia in the training set (n = 413)VariablesKinesiophobia(n = 181)Non-kinesiophobia(n = 232)χ2/t/z p Gender, n(%)0.1250.723 Male52 (28.73)63 (27.16) Female129 (71.27)169 (72.84)Age (years), median(IQR)69.00 (63.00, 73.00)65.50 (56.00, 71.00)-4.513< 0.001*BMI (kg/m²), median(IQR)22.50 (20.36, 25.37)22.81 (20.75, 25.93)-0.8720.383Education, n(%)11.1500.004* Junior high and below105 (58.01)97 (41.81) Senior high/technical school51 (28.18)84 (36.21) Junior college and above25 (13.81)51 (21.98)Marital status, n(%)1.0690.301 Married176 (97.24)221 (95.26) Unmarried (divorced, widowed)5 (2.76)11 (4.74)Average monthly household income, n(%)1.6880.194 < USD 700127 (70.17)176 (75.86) ≥USD 70054 (29.83)56 (24.14)Medical payment type, n(%)0.9650.617 Urban employee medical insurance120 (66.30)143 (61.64) Urban resident medical insurance45 (24.86)65 (28.02) Others16 (8.84)24 (10.34)Affected side, n(%)2.2070.332 Right40 (22.10)53 (22.84) Left30 (16.57)51 (21.98) Bilateral111 (61.33)128 (55.17) Disease duration (years), median(IQR)3.5 (1.00, 10.00)5 (1.00, 10.00)-0.9920.321Comorbid chronic diseases, n(%)0.0010.975 ≥ 3 types38 (20.99)49 (21.12) < 3 types143 (79.01)183 (78.88)NRS, median(IQR)7.00 (6.00, 8.00)5.00 (4.00, 7.00)-8.017< 0.001*PCS, median(IQR)23.00 (17.00, 27.00)13.00 (10.00, 20.00)-8.183< 0.001*HADS-anxiety, median(IQR)7.00 (5.00, 9.00)6.50 (4.00, 9.00)-1.2130.225HADS-depression, median(IQR)3.00 (1.00, 5.00)2.00 (1.00, 5.00)-1.0950.274WOMAC, median(IQR)51.00 (39.00, 57.50)35.50 (25.25, 52.00)-6.883< 0.001*UCLA Activity Score, median(IQR)4.00 (3.00, 5.00)5.00 (4.00, 6.00)-4.578< 0.001*MSPSS, median(IQR)45.00 (33.00, 61.00)47.00 (36.00, 64.75)-1.1340.257*p < 0.05, compared with non-kinesiophobia groupAbbreviations: BMI Body mass index, HADS Hospital Anxiety and Depression Scale, IQR Interquartile range, MSPSS Multidimensional Scale of Perceived Social Support, NRS Numerical Rating Scale, PCS Pain Catastrophizing Scale, UCLA University of California, Los Angeles, WOMAC Western Ontario and McMaster Universities Osteoarthritis Index Comparison of characteristics between KOA patients with and without kinesiophobia in the training set (n = 413) *p < 0.05, compared with non-kinesiophobia group Abbreviations: BMI Body mass index, HADS Hospital Anxiety and Depression Scale, IQR Interquartile range, MSPSS Multidimensional Scale of Perceived Social Support, NRS Numerical Rating Scale, PCS Pain Catastrophizing Scale, UCLA University of California, Los Angeles, WOMAC Western Ontario and McMaster Universities Osteoarthritis Index Table 2Binary LR analysis in the training setVariablesBSEWald χ2 p OR (95%CI)Age (years)0.0940.01635.431< 0.0011.099 (1.065–1.133)Educational level--8.0590.018- Junior high and below0.7910.3465.2190.0222.206 (1.119–4.351) Senior high /technical school0.1310.3670.1280.7211.140 (0.555–2.340) Junior college and above-----NRS0.3360.08714.794< 0.0011.399 (1.179–1.660)PCS0.0840.01920.673< 0.0011.088 (1.049–1.128)WOMAC0.0150.0102.2850.1311.015 (0.995–1.036)UCLA Activity Score-0.3710.08518.896< 0.0010.690 (0.584–0.816)Constant-9.4921.36248.597< 0.001-Abbreviations: B Unstandardized regression coefficient, CI Confidence interval, NRS Numerical Rating Scale, OR Odds ratio, PCS Pain Catastrophizing Scale, SE Standard error, UCLA University of California, Los Angeles, WOMAC Western Ontario and McMaster Universities Osteoarthritis Index Binary LR analysis in the training set Abbreviations: B Unstandardized regression coefficient, CI Confidence interval, NRS Numerical Rating Scale, OR Odds ratio, PCS Pain Catastrophizing Scale, SE Standard error, UCLA University of California, Los Angeles, WOMAC Western Ontario and McMaster Universities Osteoarthritis Index Establishment of the nomogram Based on the five identified predictors, the LR model was formulated as follows: Educational attainment was modeled using two dummy variables: Education1 (1 = junior high and below; 0 = otherwise) and Education2 (1 = senior high/technical school; 0 = otherwise), with junior college and above as reference (Education1 = Education2 = 0). The nomogram is based on this model, where the total score corresponds to the predicted probability of kinesiophobia. For example, the first patient in the dataset (a 67-year-old KOA patient with junior high education, NRS = 5, PCS = 21, and UCLA Activity Score = 2) yields a predicted probability of 71.4% (Fig. 1). Nomogram predicting kinesiophobia risk in KOA. To estimate risk: map variable values to the Points axis, sum all points, and project the total onto the Probability axis. Blue density distributions show variable spreads; the red dot indicates an exemplary patient case. Abbreviations: NRS = Numerical Rating Scale; PCS = Pain Catastrophizing Scale; UCLA = University of California, Los Angeles Validation of the nomogram ROC analysis quantified discriminative capacity (AUC primary metric). Training set AUC was 0.844 (95% CI 0.806-0.882), indicating strong discriminatory power. At the optimal threshold of 0.410, sensitivity was 81.2% and specificity was 74.6%. External validation on the test set yielded an AUC of 0.815 (95% CI 0.750–0.881); at an applied threshold of 0.454, sensitivity was 78.8% and specificity was 80.4%, confirming the model's generalizability to independent cohorts (Fig. 2). ROC curves of the nomogram for kinesiophobia risk. a Training set (AUC 0.844). b Test set (AUC 0.815). Blue dashed line: chance reference line (AUC = 0.5); red solid line: nomogram performance. Abbreviations: AUC = area under the curve Calibration curves showed close alignment with the 45° ideal line. Both training and validation curves adhered to the reference line (Fig. 3), indicating strong internal validity and cross-sample consistency, supporting clinical utility. Calibration curves assessing agreement between predicted and observed kinesiophobia probabilities. a Training set. b Test set. Black dashed line: ideal concordance; red line: apparent calibration; green line: bootstrap-corrected calibration. Close alignment of corrected curves (green) with the ideal line indicates good model fit DCA demonstrated clinical utility across threshold probabilities, showing greater net benefit than the 'all' and 'none' strategies within an approximately 25-75% range. Beyond this, net benefit declined on the test set, indicating limited generalizability for high-risk predictions. Patient heterogeneity contributed to increased variability at high thresholds, necessitating cautious interpretation of these results (Fig. 4). DCA evaluating clinical utility. a Training set. b Test set. Red line: 'All' strategy (all patients screened); green line: 'None' strategy (no screening); blue line: nomogram. Within 25–75% threshold probability, the nomogram shows higher net benefit than reference strategies, confirming diagnostic value. Declining net benefit beyond 75% indicates limited generalizability for high-risk predictions Performance of ML models Evaluation of ML models To further evaluate the robustness of feature selection, we employed a data-driven approach using the random forest algorithm. Variables were ranked by importance based on both the Mean Decrease Gini and the Mean Decrease Accuracy, with those consistently in the top ten for both metrics selected as the final predictors: age, NRS, PCS, WOMAC, and UCLA Activity Score (Table S2, Supplementary Material 2). This selection showed substantial overlap with the predictors identified by LR. Models trained on these data-driven features showed marginally lower predictive performance compared with those based on the LR-selected features (Table S3, Supplementary Material 2). Therefore, we proceeded to train seven ML models using the predictors from the original LR model. All models demonstrated good discrimination, with the SVM, MLP, and ML-LR models achieving the highest AUC values (Fig. 5). Test set AUC values for ML models. SVM showed numerically highest AUC (0.821), followed by MLP (0.820), ML-LR (0.815), LightGBM (0.805), RF (0.802), XGBoost (0.789), and CatBoost (0.771). Abbreviations: AUC = area under the curve; LR = logistic regression; ML = machine learning; MLP = multilayer perceptron; RF = random forest; SVM = support vector machine; XGBoost = extreme gradient boosting Detailed classification performance across all models is summarized in Table 3. In brief, MLP achieved the highest accuracy, LightGBM the highest sensitivity, and ML-LR the highest specificity and positive predictive value. The SVM model demonstrated balanced and competitive performance across multiple metrics, achieving the second-highest sensitivity and AUC. Table 3Performance comparison of ML models in the test setModelAUCAccuracySensitivitySpecificityPPVNPVF1 scoreCatBoost0.771 (0.701–0.842)0.7060.7070.7060.7220.6900.714LightGBM0.805 (0.738–0.872)0.7800.8150.7410.7730.7880.794ML-LR0.815 (0.750–0.881)0.7680.7390.8000.8000.7390.768RF0.802 (0.737–0.867)0.7230.7500.6940.7260.7200.738SVM0.821 (0.757–0.885)0.7740.8040.7410.7710.7780.787XGBoost0.789 (0.721–0.858)0.7230.6960.7530.7530.6960.723MLP0.820 (0.755–0.885)0.7850.7830.7880.8000.7700.791Abbreviations: AUC Area under the curve, LR Logistic regression, ML Machine learning, MLP Multilayer perceptron, NPV Negative predictive value, PPV Positive predictive value, RF Random forest, SVM Support vector machine, XGBoost Extreme gradient boosting Performance comparison of ML models in the test set Abbreviations: AUC Area under the curve, LR Logistic regression, ML Machine learning, MLP Multilayer perceptron, NPV Negative predictive value, PPV Positive predictive value, RF Random forest, SVM Support vector machine, XGBoost Extreme gradient boosting Calibration curves revealed that SVM, ML-LR, LightGBM, and XGBoost exhibited robust calibration, whereas CatBoost and MLP exhibited marked oscillatory deviations, suggesting calibration bias (Fig. 6a). DCA revealed threshold-dependence in model performance. Certain models (e.g., SVM, ML-LR) provided greater net benefit at low risk thresholds, while others (e.g., CatBoost, MLP) showed reduced net benefit for high-risk predictions. Notably, the RF model maintained a positive net benefit across the broadest range of threshold probabilities (Fig. 6b). These threshold-related variations highlight the need to select models based on the intended clinical use and patient population. Test set calibration and DCA. a Calibration curves, proximity to 45° ideal line reflects goodness-of-fit. b DCA: black = 'All' strategy; grey = 'None'. SVM and ML-LR exceed references at thresholds <0.2. MLP and CatBoost show reduced net benefit at thresholds approaching 0.75 versus robust models. Abbreviations: LR = logistic regression; ML = machine learning; MLP = multilayer perceptron; RF = random forest; SVM = support vector machine; XGBoost = extreme gradient boosting Visual interpretation of the SVM model To quantify feature importance in the SVM model, we used mean absolute SHAP values (Fig. 7a). NRS and PCS demonstrated dominant effects, highlighting the central role of pain perception and catastrophic thinking in kinesiophobia. The UCLA Activity Score also made a substantial contribution, aligning with its known inverse relationship with fear of movement. Age exhibited a non-monotonic pattern, suggesting complex physiological pathways. Educational background showed more modest effects than pain-related variables, consistent with its socioeconomic role as an indirect influencer of health beliefs. These trends were visible in the beeswarm plot (Fig 7b), where higher NRS and PCS values correlated with positive SHAP values (increased risk), while higher UCLA Scores corresponded to negative values (protective effect). Lower education was associated with mildly elevated risk. Together, these findings suggest clinical assessments should prioritize pain and catastrophizing, while also considering physical activity and socioeconomic context. SVM model feature interpretation via SHAP. a Feature importance: mean |SHAP| values rank predictors. b Directional impact: point color = feature value; SHAP value sign = risk direction (positive: risk increase). Abbreviations: NRS = Numerical Rating Scale; PCS = Pain Catastrophizing Scale; SHAP = Shapley Additive exPlanations; UCLA = University of California, Los Angeles Discussion Kinesiophobia represents a significant barrier to rehabilitation in patients with KOA, underscoring the imperative for early identification and intervention to improve clinical outcomes and population health. Our study identified five independent predictors and compared an LR‑based nomogram with seven ML models. SVM had a modestly higher accuracy than the nomogram, but the difference calls for cautious interpretation. The prevalence of kinesiophobia in our study was approximately 45%, which is consistent with the wide range (30–80%) reported in the KOA population [24, 25]. This consistency supports the external validity of our findings and indicates that our sample is representative of the broader KOA population. Interpretation and comparison of predictors Our findings support the key role of pain-related factors, specifically pain intensity and catastrophizing, in the development of kinesiophobia among KOA patients, providing mechanistic support for the fear-avoidance model. The association between higher pain levels and increased kinesiophobia incidence aligns with prior reports [26]. Pain experiences reinforce maladaptive beliefs linking movement to harm, promoting avoidance behaviors [27]. Furthermore, pain catastrophizing exacerbates perceived pain threats, heightening helplessness and fear of movement [27–29]. We confirmed an inverse relationship between physical activity level and kinesiophobia. This observed association is consistent with a hypothesized bidirectional interplay, whereby kinesiophobia discourages activity, leading to deconditioning and impaired joint function, while lower fitness levels amplify safety concerns and reduce exercise willingness [30, 31]. This self-perpetuating cycle partly explains why only 13% of KOA patients meet physical activity guidelines [32]. Although moderate activity promotes beneficial cycles of pain reduction and functional gains, kinesiophobia disrupts this process. Therefore, breaking the cycle requires early identification of high-risk profiles through predictive modeling, combined with interventions targeting cognitive restructuring and graded activity exposure [31, 33, 34]. Our study further identified age and educational attainment as significant, multidimensional predictors of kinesiophobia risk in KOA. Advancing age was associated with increased risk. This was potentially attributable to factors such as age-related declines in balance, heightened perception of fall risk, and a greater tendency among older adults to misinterpret benign exercise-related discomfort as indicative of joint damage, thereby promoting avoidance behaviors [35]. Lower educational attainment (junior high school and below) was independently associated with elevated kinesiophobia risk, consistent with prior reports [36, 37]. This association may reflect that higher education facilitates better health literacy and adaptive coping strategies, whereas limited education may compromise health literacy, thereby increasing susceptibility to maladaptive movement beliefs [37]. These findings highlight potential socioeconomic dimensions of kinesiophobia and support stratified interventions, particularly enhanced pain management combined with personalized health education targeting cognitive biases in elderly and less-educated patients. Nevertheless, due to conceptual overlap between key predictors (e.g., pain intensity and pain catastrophizing) and kinesiophobia, our findings should be interpreted with caution: our models may partly reflect correlated constructs rather than fully independent predictive effects. While this overlap supports reasonable clinical face validity (higher pain and catastrophizing are associated with greater kinesiophobia risk), it also complicates the interpretation of each variable’s unique contribution. Future studies may consider including more distal predictors to better disentangle independent mechanisms. Clinical applicability of the predictive models The developed models demonstrate complementary strengths and are applicable across diverse clinical and digital health contexts. The LR-based nomogram excels in clinical practicality. Its visual interface enables rapid point-of-care risk estimation, making it ideal for settings such as primary care, where simplicity, clear patient communication, and rapid interpretable assessment are crucial. Crucially, its high interpretability allows clinicians to readily understand each predictor’s contribution to the overall risk score. This feature supports its potential utility in musculoskeletal and neurological contexts [9, 38]. Rooted in linear relationships, the nomogram maintains inherent simplicity and clarity. In contrast, the ML-LR model may be particularly suited for ruling out kinesiophobia. The other evaluated ML models offer strengths, each catering to distinct clinical needs. SVM achieved one of the highest levels of discriminative performance, likely attributable to its ability to model complex non-linear predictor interactions [39, 40]. For example, SHAP analysis revealed nuanced, non-monotonic relationships such as that between age and kinesiophobia risk, suggesting SVM’s particular suitability for high-precision prediction tasks involving complex variable dependencies. LightGBM may be useful in screening contexts where minimizing false negatives is prioritized. By contrast, MLP could be considered for settings requiring balanced classification under varying predictor profiles. Notably, SVM, ML-LR, LightGBM, and XGBoost generated well-calibrated risk estimates, enhancing their reliability for clinical application. Decision curve analysis further supported the practical value of SVM and ML-LR, which provided sustained net benefit over baseline strategies, particularly in lower-risk subpopulations. Despite their complexity, these models maintained interpretability through SHAP analysis, which highlighted their reliance on established clinical drivers like pain intensity and catastrophizing. This alignment with clinical intuition supports their potential integration into digital health platforms for population-level screening and intervention. Collectively, the choice among these models is not determined by a substantial superiority of one, but by their fit to specific clinical or operational needs. The modest improvement in predictive accuracy, while not transformative, holds clear clinical relevance: it underscores that model utility is inherently context-dependent. The interpretable LR nomogram is optimal for rapid, transparent assessment in primary care. Conversely, the SVM model, whose marginal accuracy gain likely stems from modeling subtle interactions, is better suited for integration into automated digital health platforms, where even small performance advantages can enhance efficiency in large-scale screening. In practice, embedding the chosen tool into a clinical decision support system enables real-time risk scoring, facilitating stratified care that ranges from targeted referrals for high-risk patients to streamlined management for others. It should be emphasized that these predictive tools are designed to complement standard instruments such as the TSK, not to replace them. Strengths and limitations This study advances the prediction of kinesiophobia in KOA through three key contributions. First, we provide a comprehensive systematic comparison of a traditional clinical prediction tool (LR-based nomogram) with seven diverse ML models. This rigorous evaluation revealed performance differences among models, providing valuable evidence for context-specific model selection. Second, we established a KOA-specific predictive framework that integrates psychometric and demographic variables with explainable artificial intelligence, thereby extending the application of interpretable ML to chronic musculoskeletal contexts and facilitating its use in digital health assessment tools. Third, by incorporating SHAP analysis, we directly addressed the “black box” concern typical of complex models, illuminating the decision-making process and enhancing trustworthiness for clinical adoption. This approach quantified the dominant influence of pain intensity and pain catastrophizing and uncovered subtle predictor patterns (e.g., differential educational effects). These findings contribute to a framework for developing interpretable, high-performance predictive models with potential value for healthcare decision-making and stratified care in this domain. Several limitations warrant acknowledgment. First, as a cross-sectional study, the present models classify current kinesiophobia status rather than predicting new onset, which precludes causal inference and longitudinal validation; prospective studies are essential. Second, external generalizability is not established due to our single‑center, internal‑validation design. Recruitment from a single tertiary hospital may limit applicability to community-based populations with milder KOA. Moreover, the participants' levels of psychological distress was below standard clinical thresholds. While this confirms the clinical relevance of our predictors in patients without severe comorbidities, it also suggests that their predictive strength may differ in groups with higher distress. Third, the exclusion of biomarkers and imaging data restricts exploration of biopsychosocial mechanisms; future work could leverage multi-omics data within a digital health framework to deepen mechanistic insights. Finally, the suboptimal discriminative performance of certain ML models (e.g., CatBoost) may reflect sample size limitations or inadequate hyperparameter tuning, potentially addressable by expanding datasets and employing advanced optimization techniques in future studies. Conclusion In conclusion, early identification of kinesiophobia is critical for optimizing rehabilitation outcomes in KOA. No single model was universally superior, and the models instead offer complementary strengths. SVM showed marginally higher predictive accuracy than other ML models, warranting further exploration of its potential advantages and integration into automated digital health platforms. Notably, the LR-based nomogram performed nearly equivalently to SVM while providing favorable interpretability, enabling rapid, transparent bedside risk assessment in primary care and community settings. Model selection should therefore be guided by specific clinical contexts to support personalized interventions. Supplementary Information Supplementary Material 1. Supplementary Material 2. Acknowledgements None. Authors’ contributions J.J. and Y.Y. contributed to conceptualization and provided supervision; X.L. contributed to conceptualization and data curation; X.H., S.M., and T.Z. contributed to investigation; Y.H. and Q.Z. performed formal analysis; X.L. wrote the original draft; all authors participated in writing—review and editing. All authors reviewed and approved the final manuscript. Funding This work was supported by Medical and Health Scientific Research Project of Qingdao (2024-WJKY018). Data availability The data underlying this article cannot be shared publicly for the protection of the privacy of individuals that participated in the study. The data may be shared upon reasonable request to the corresponding author. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of Qingdao Municipal Hospital (2024-KY-083). All procedures were performed in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Yang Yu, Email: tom___0519@163.com. Junjie Jiang, Email: JiangJJ@xjtu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Data Availability Statement The data underlying this article cannot be shared publicly for the protection of the privacy of individuals that participated in the study. The data may be shared upon reasonable request to the corresponding author.