Smartphone-Based Physical Performance and Multidimensional Determinants of Self-Reported Knee Pain in Community-Dwelling Older Adults: Cross-Sectional Machine Learning and Network Analysis Study.

Background: Knee pain affects 22.9% of individuals aged 40 years and over globally and is associated with body function, activity, environmental, and personal factors described in the International Classification of Functioning, Disability and Health (ICF) model. Most prior studies examined isolated risk factors using conventional regression. Objective: This study aimed to examine self-reported knee pain in community-dwelling older adults by combining machine learning and partial-correlation network analysis with smartphone-based physical performance measurement. Methods: This cross-sectional study included 852 adults aged 60 years and older. Baseline assessment of 38 variables across 7 ICF-aligned domains was reduced to 21 predictors using prespecified, rule-based criteria (events per variable=14.95). Walking speed, sit-to-stand, gait knee flexion, and gait asymmetry were measured using a validated vision-based smartphone app. Missing data were addressed by multiple imputation (m=5) with Rubin rules. Six algorithms (logistic regression with elastic-net penalty as baseline, k-nearest neighbors, random forest, Extreme Gradient Boosting, Light Gradient Boosting Machine, and support vector machine) were trained using nested 5×5 cross-validation, and the best model was interpreted using Shapley Additive Explanations and partial dependence plots. Partial-correlation network analysis examined the 10 top-ranked features and the knee-pain node, with bootstrap stability assessment and a sensitivity analysis excluding EuroQol 5-Dimension (EQ-5D) Pain/Discomfort. Results: Knee pain prevalence was 36.9% (314/852). Pooled area under the receiver operating characteristic curve values ranged from 0.692 to 0.723, with random forest highest (area under the receiver operating characteristic curve=0.723, 95% CI 0.689-0.757). The top Shapley Additive Explanations features were EQ-5D Pain/Discomfort, EQ-5D Utility, and house estate. In the 11-node network, only EQ-5D Pain/Discomfort (partial correlation=0.260) and house estate (0.158) had direct edges with knee pain. Sit-to-stand showed the highest strength centrality (0.679) without a direct edge, acting as a hub linking body function, body structure, mental health, and activity variables. Shapley Additive Explanations importance and network strength were weakly correlated (ρ=0.018). The sensitivity analysis preserved the pattern (best area under the receiver operating characteristic curve=0.724). Conclusions: Knee pain in older adults was associated with variables spanning 7 ICF domains, with EQ-5D Pain/Discomfort and housing environment (lower-income rental estate residence) as direct correlates and sit-to-stand as a network hub. Combining machine learning and partial-correlation network analysis may inform multidisciplinary biopsychosocial assessment, pending confirmation in prospective and externally validated studies.