Accelerometry and dual-scale neighborhood indicators for screening of MoCA-defined cognitive impairment: an interpretable machine-learning study.

Background: Community-level identification of older adults who may benefit from confirmatory cognitive assessment is constrained by limited resources and the modest accuracy of single-domain screeners. Objective: This study aimed to develop and internally validate an interpretable prediction model for MoCA-defined cognitive impairment that integrates accelerometry with dual-scale (500 m and 800 m) built-environment indicators, and to evaluate calibration, clinical utility, and subgroup performance. The aim was to support community screening prioritization rather than etiologic inference or clinical diagnosis. Methods: We analyzed a cross-sectional sample of community-dwelling older adults from Nanjing, China ( Results: Tree-based models achieved the best overall performance. In the held-out test set, random forest showed high discrimination with acceptable calibration (AUC 0.95, 95% CI 0.91-1.00; Brier 0.088), while gradient boosting and support vector machine achieved AUCs approximately 0.90 with lower Brier scores for boosting (0.071). Decision curve analysis indicated a positive net benefit relative to the treat-all and treat-none strategies across clinically plausible risk thresholds. Explainability analyses consistently highlighted MVPA, sedentary time, age, central adiposity, and neighborhood transit accessibility as influential predictors. Subgroup analyses indicated broadly comparable discrimination, with small-to-moderate equalized-odds gaps. Conclusion: Combining accelerometry with neighborhood indicators may support calibrated, decision-oriented triage for MoCA-defined cognitive impairment in community settings. This model is intended for screening support rather than causal interpretation or diagnostic replacement, and requires external validation before implementation.