A total of 1027 community-dwelling Hong Kong residents aged 60 years or older were recruited for the study. Eligible participants were those living independently in housing estates, able to walk independently for at least 8 meters, and capable of providing informed consent. Exclusion criteria included (1) severe cognitive impairment preventing informed consent; (2) inability to communicate in Cantonese or read Traditional Chinese; (3) history of hip surgery, knee surgery, or ankle surgery within the past 12 months; (4) recent knee injury (within the past 6 mo); and (5) diagnosed neurological diseases including Parkinson disease, stroke, multiple sclerosis, or other conditions. Of the 1027 participants initially screened, 852 participants (83%) met all inclusion criteria and completed the comprehensive baseline assessment protocol. Study flowchart and participant characteristics are presented in Figure 1 and Table 1, respectively. Reporting follows the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + AI statement for prediction model studies using AI and ML [17], consistent with the Journal of Medical Internet Research–recommended guidelines for ML predictive models in biomedical research [18].

This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of The Hong Kong Polytechnic University (HSEARS20220628001). All participants provided written informed consent before data collection, after the nature and possible consequences of the study were explained. Video recordings captured during the smartphone-based assessments were transmitted to institutional servers for automated pose-estimation analysis; all data were deidentified before analysis and stored in a password-protected database accessible only to the research team. No identifiable participant information is presented in this paper or its supplementary materials. As part of the wider project, participants who completed the exercise program and achieved a predefined fitness criterion at the 12-month follow-up received a health hamper valued at HKD 200 (approximately US $26).

The primary outcome was the presence of self-reported knee pain experienced within the preceding 3 months, ascertained through structured interviewer-administered questionnaires conducted by trained research assistants. Knee pain was defined as any ache, discomfort, or pain sensation experienced in one or both knees during this 3-month period [1]. Participants were classified as having knee pain (Yes) or not (No) based on this question.

All ML analyses were conducted using R statistical software (version 4.5.3; R Core Team). Baseline characteristics of the analytic cohort were compared between participants with and without self-reported knee pain using independent 2-tailed t tests (continuous variables) or χ² tests (categorical variables).

Of the 852 participants, 314 (36.9%) reported knee pain in the preceding 3 months and 538 (63.1%) did not. Baseline characteristics by knee pain status are presented in Table 1. The two groups were comparable in age (2-tailed t₆₉₃=1.27; P=.21), with a small difference in the proportion of female participants (χ²₁=5.08; P=.02). Participants with knee pain had higher BMI (t₆₁₃=−3.22; P=.001), were more frequently rental estate residents (χ²₁=35.57; P<.001), and showed slower walking speed (t₇₀₂=2.41; P=.02) and fewer sit-to-stand repetitions (t₇₁₃=3.85; P<.001). They reported more sitting time (t₆₃₅=−3.81; P<.001), less walking-related (t₈₁₆=4.61; P<.001), and vigorous physical activity (all P<.001), poorer health-related quality of life across multiple EQ-5D dimensions, lower mental health scores (t₆₄₀=2.74; P=.006), and a higher count of sleep disorder categories (t₅₅₃=−4.88; P<.001). The largest between-group difference was observed for EQ-5D Pain/Discomfort (Cohen d=−0.73), followed by EQ-5D Utility (Cohen d=0.57) and EQ-5D Mobility (Cohen d=−0.39).

Pooled performance metrics for the six classification algorithms are presented in Table 2 and Figure 2. AUC values clustered within a narrow range from 0.692 (95% CI 0.661-0.724) for LightGBM to 0.723 (95% CI 0.686-0.760) for RF. The LR baseline showed a pooled AUC of 0.721 (95% CI 0.691-0.751), placing it within 0.002 of RF and 0.001 of SVM. F₁-score, recall, and MCC were modest across all algorithms (MCC range 0.252-0.328), and LR produced the lowest Brier score (0.200), indicating the best calibration. Performance was stable across the 5 imputations, with within-imputation variation in mean AUC less than 0.02 for every algorithm (Figure 2D). The discrimination performance of all six algorithms was effectively comparable; RF was retained as the best-performing algorithm by pooled AUC for SHAP analysis.

Values are pooled estimates (95% CI) across five multiply-imputed datasets, obtained via Rubin’s rules from nested 5×5 cross-validation. Algorithms are ordered by pooled AUC.

Pooled performance of six classification algorithms (LR with elastic-net penalty, k-nearest neighbors, RF, XGB, LightGBM, and SVM with radial basis function kernel) estimated using nested 5×5 cross-validation on five multiply-imputed datasets and pooled via Rubin rules.

Pooled SHAP-based feature importance for the RF model is presented in Figure 3, with values for all 21 variables in Table S5 in Multimedia Appendix 1. The three highest-ranked features were EQ-5D Pain/Discomfort (mean |SHAP|=0.064), EQ-5D Utility (0.038), and house estate (0.034). Subsequent features clustered at substantially lower magnitudes (|SHAP| range 0.013-0.019) and included physical activity, anthropometric, smartphone-based physical performance, and mental and sleep variables. The SHAP value distributions in Figure 3 characterized the direction of each association: higher EQ-5D Pain/Discomfort scores, lower EQ-5D Utility, rental estate residence, lower walking physical activity and sit-to-stand repetitions, higher BMI and sitting time, and higher sleep disorder counts were each associated with increased predicted probability of knee pain, consistent with the group differences observed in Table 1.

PDPs for the top 9 SHAP-ranked features are shown in Figure 4. The model-implied probability of knee pain rose sharply between EQ-5D Pain/Discomfort scores of 1 and 2 (from approximately 0.31 to 0.43), reaching a plateau near 0.45. The relationship for EQ-5D Utility was relatively flat across most of the range and decreased steeply at utility values above approximately 0.85. House estate showed a step function consistent with its binary coding. For continuous features, the model-implied probability of knee pain decreased with higher walking physical activity and sit-to-stand repetitions and increased with higher BMI and sitting time. Gait knee flexion and age showed comparatively flat partial dependence functions. These patterns describe model-implied marginal relationships and are presented as exploratory; they do not establish causal effects or clinical thresholds.

The pooled partial-correlation network for the 11 nodes is shown in Figure 5A, with the full pooled partial-correlation matrix in Table S6 in Multimedia Appendix 1. After EBICglasso regularization and edge thresholding, 11 conditional dependencies survived. The highest strength centrality values were observed for sit-to-stand (0.679), EQ-5D Pain/Discomfort (0.666), and mental health (0.526), with the knee pain node ranking 5th of 11 (0.418; Figure 5). The case-dropping bootstrap CS-coefficient was 0.36 for strength (acceptable stability), and 0.00 and 0.05 for closeness and betweenness, respectively.

Two direct edges connected the knee pain node to the predictor set: a positive edge with EQ-5D Pain/Discomfort (partial correlation=0.260) and a positive edge with house estate (0.158). The remaining nine network edges represented conditional dependencies among the predictors after controlling for all others. The prespecified Spearman rank correlation between SHAP importance and network strength centrality across the 10 feature nodes was ρ=0.018 (P=.97), indicating no monotonic association between the two paradigms (Figure 5).

The decomposition of each variable’s contribution into direct and indirect pathways (Figure 5D) showed a clear binary pattern. EQ-5D Pain/Discomfort and house estate carried their contributions entirely through direct conditional dependencies, with edge magnitudes (0.260 and 0.158) larger than their corresponding mean absolute SHAP values (0.064 and 0.034). The remaining eight predictors had no direct edge to knee pain after regularization (Figure 5), indicating that their contributions to the model output were accounted for by their conditional dependencies with other variables in the network. This pattern of two variables with direct dependencies and eight contributing only through indirect pathways is consistent with the multidimensional structure described by the ICF framework.

Excluding EQ-5D Pain/Discomfort (n=20 variables; EPV=15.7) preserved the qualitative pattern of findings (Table S4 in Multimedia Appendix 1, Figures S1 and S2 in Multimedia Appendix 1). The pooled best AUC was 0.724 for SVM (ΔAUC =+0.001 vs main analysis; Figure S1 in Multimedia Appendix 1). The top SHAP feature shifted from EQ-5D Pain/Discomfort to EQ-5D Utility, but the next-ranked variables (house estate, walking physical activity, and BMI) remained the same (Figure S2 in Multimedia Appendix 1). House estate retained a direct conditional dependency with knee pain (partial correlation=0.20). The Spearman correlation between SHAP importance and network strength remained close to zero (ρ=−0.10; P=.78), and the CS-coefficient for strength was 0.44. The principal findings, including the modest discrimination performance, the predominance of indirect pathways from physical performance and mental health variables to knee pain, and the divergence between SHAP and network rankings, did not depend on inclusion of EQ-5D Pain/Discomfort.

This study examined self-reported knee pain in 852 community-dwelling older adults in Hong Kong using ML and partial-correlation network analysis. Based on the ICF model, knee pain in older age is associated with body function and structure, activity, participation, environmental, and personal factors [6,9,34]. These findings support this multidimensional view. Across the 21 predictors covering seven ICF domains, no single variable was dominant, and the model output reflected a distributed contribution from the health-related quality of life, environmental, physical performance, physical activity, mental health, and sleep domains. These results agree with the biopsychosocial view of knee pain in older adults [35], and with the projected rise of knee osteoarthritis to 595 million prevalent cases globally by 2050 [4,36]. Combining SHAP and network analysis allowed the assessment of which variables relate to knee pain and how those variables relate to one another.

The six classification algorithms showed similar discrimination, with pooled AUC ranging from 0.692 to 0.723. LR (AUC 0.721, 95% CI, 0.691–.751) was within 0.002 of RF and 0.001 of SVM, and showed the lowest Brier score (0.200) among the six algorithms. These results agree with prior reports that ML did not outperform LR in clinical prediction studies [37], and with recent recommendations that adequate sample size and prespecified variable selection, rather than algorithmic complexity, are the main determinants of stable AI-based prediction [38]. The discrimination level (best AUC 0.723) was comparable with prior ML models for knee pain or knee osteoarthritis in older adults, in which reported AUCs ranged from 0.70 to 0.80 [39-41]. With mild class imbalance (36.9% vs 63.1%), algorithmic resampling can degrade calibration without improving discrimination [42]; imbalance-aware metrics (AUC, F₁-score, MCC, and Brier score) were therefore used [43]. The convergence of LR and ML performance suggests that, when EPV is 14.95 and the predictor set is conceptually prespecified, the benefit of more flexible algorithms over a regularized LR is small in cross-sectional knee pain classification.

SHAP-based importance and PDPs showed the model-implied association of each variable with the predicted probability of knee pain. The three highest-ranked features were EQ-5D Pain/Discomfort (mean |SHAP|=0.064), EQ-5D Utility (0.038), and house estate (0.034). The remaining features clustered at lower magnitudes (mean |SHAP| range, 0.013‐0.019) and included variables from the physical activity, sociodemographic, smartphone-based physical performance, mental health, and sleep domains. The SHAP-derived directions of association agreed with the baseline comparisons: lower walking-related physical activity and sit-to-stand repetitions, higher BMI and sitting time, lower mental health scores, and a higher count of sleep disorder categories were each associated with an increased predicted probability of knee pain. These patterns agree with prior reports linking reduced gait speed and lower-extremity capacity to musculoskeletal pain in older adults [44], and with the bidirectional association between sleep disturbance and chronic musculoskeletal pain [45]. Because PDPs show model-implied marginal associations rather than causal effects, the inflection points described below are reported as hypothesis-generating and require prospective validation. The predicted probability of knee pain rose sharply between EQ-5D Pain/Discomfort scores of 1 and 2 (0.31 to 0.43) and declined above an EQ-5D Utility of 0.85, while for the modifiable variables it declined steeply with walking physical activity up to 3000 MET·minutes/week, with sit-to-stand performance up to 15 repetitions, and rose above a BMI of 22 kg/m² and above a sitting time of 150 minutes/day. These patterns are consistent with prior reports linking sedentary behavior, low physical activity, higher BMI, and reduced lower-extremity capacity to musculoskeletal pain in older adults, but the exact values should not be treated as clinical cutoffs.

Network analysis showed how variables relate to one another after adjusting for the rest of the network. In the pooled 11-node network, only two predictors had direct edges with the knee pain node after EBICglasso regularization: EQ-5D Pain/Discomfort (partial correlation=0.260) and house estate (0.158). The remaining nine edges were among predictors. The Spearman correlation between SHAP importance and network strength centrality across the 10 feature nodes was ρ=0.018, indicating that the two analyses ranked variables differently and captured distinct facets of the multidimensional structure. Sit-to-stand showed this difference clearly. Although its marginal SHAP magnitude was modest (mean |SHAP|=0.0168; rank 7 of 21) and no direct edge to knee pain survived regularization, sit-to-stand showed the highest strength centrality (0.679). Sit-to-stand therefore acted as a central physical performance hub whose conditional dependencies extended across multiple ICF domains, including body structure and function (BMI, partial correlation = −0.146; age, −0.260), mental and cognitive function (mental health, +0.128), and activity (walking physical activity, +0.150). This pattern is consistent with the role of the sit-to-stand test as an integrative measure of lower-extremity power, balance, and global physical capacity in older adults [46,47]. Sit-to-stand therefore appears not as a direct correlate of knee pain, but as a multidimensional hub that links body function, body structure, activity, and mental health variables, which agrees with the biopsychosocial pathways described in the ICF model [15,48,49]. The direct edge between house estate and knee pain reflects rental estate residency as an indicator of lower socioeconomic position in Hong Kong, and is consistent with the well-documented inverse association between socioeconomic status and chronic pain [50], and with reports that public rental housing in Hong Kong is associated with poorer health-related behaviors and chronic disease outcomes [51]. Beyond reflecting socioeconomic position, the housing categories in this cohort were defined primarily by income level rather than by physical or care-related features, so the mechanisms underlying the association with knee pain remain uncertain and warrant further study. The direct edge between EQ-5D Pain/Discomfort and knee pain partly reflects conceptual overlap, as both capture self-reported pain; however, the two are distinct constructs (general pain or discomfort versus knee-specific pain), and a sensitivity analysis excluding EQ-5D Pain/Discomfort preserved the qualitative pattern of findings, indicating that our conclusions do not depend on this predictor.

Several implications for clinical practice and future research are noted. First, the multidimensional pattern across ICF domains supports multidisciplinary, biopsychosocial assessment for older adults with knee pain, rather than the use of isolated impairment-based measures [34,35]. Second, the central network position of sit-to-stand, despite its weak direct association with knee pain, indicates that this measure was conditionally connected to body structure, mental health, and activity domains within the observed cross-sectional network, in which sit-to-stand was simultaneously connected to BMI and age, mental health, and walking physical activity. Sit-to-stand performance was therefore associated with multiple ICF domains in this network, a pattern that would be missed by considering its marginal SHAP rank alone and that prospective studies should test directly. This relationship is consistent with reports linking sit-to-stand power to functional outcomes in advanced knee osteoarthritis [26], and with the general role of sit-to-stand as a functional vital sign in geriatric assessment [46]. Third, the smartphone-based digital health system used in this study, validated for walking speed (r=0.975), knee flexion (r=0.907‐0.923), and test-retest reliability (intraclass correlation coefficient=0.86‐0.94) [16], provides a feasible method for measuring ICF body function and activity domain variables in community settings. Because these inferences are hypothesis-generating, independent external validation in separate cohorts is a necessary next step before any clinical implementation, alongside prospective longitudinal studies.

This study had several limitations. First, the cross-sectional design did not allow causal inference; the direct and indirect pathways describe conditional dependencies and model-implied associations, not mechanistic effects. Second, knee pain was measured by a single self-report item covering the preceding 3 months and did not distinguish etiology (osteoarthritis, rheumatoid arthritis, meniscal injury, or referred pain), severity, or laterality; medical diagnoses and medication use for knee pain (such as oral or topical nonsteroidal anti-inflammatory drugs, duloxetine, or intra-articular injections) were not ascertained and could not be included in the modeling. These omissions limit the clinical interpretability of the outcome and may contribute to the modest discrimination observed. Third, participants were recruited from eleven Hong Kong Housing Society estates by convenience sampling and were predominantly female (77.5%); generalizability to other regions, ethnicities, sex distributions, and socioeconomic strata is therefore limited. Fourth, no external validation was performed; nested 5×5 cross-validation provided unbiased internal estimates but not external transportability. Fifth, strength centrality showed acceptable bootstrap stability (CS-coefficient=0.36), but closeness and betweenness were unstable; these two measures were therefore not interpreted further. Sixth, the interpretive tools have inherent limits: PDPs show model-implied marginal associations, not clinical thresholds; SHAP values are model-dependent and may split importance among correlated predictors; and centrality metrics, especially closeness and betweenness, can be unstable in modest cross-sectional networks, which is why strength was pre-specified as the primary measure. Finally, the smartphone-based assessment used a research-prototype application that is not commercially available, and broader translation will require additional regulatory and engineering work. Future prospective and longitudinal studies with larger and more diverse cohorts, severity-graded outcomes, and independent external validation are required to test the hypotheses generated by this work.

This study combined ML and partial-correlation network analysis to examine self-reported knee pain in community-dwelling older adults within the ICF framework. The six classification algorithms showed pooled AUC values ranging from 0.692 to 0.723, indicating modest discrimination across the predictor set. Network analysis showed that only EQ-5D Pain/Discomfort and house estate had direct edges with knee pain, while sit-to-stand acted as a central hub linking body function, body structure, mental health, and activity variables. These cross-sectional, hypothesis-generating findings may inform multidisciplinary biopsychosocial assessment, but not causal or intervention conclusions, pending prospective external validation.