Identifying Mild Cognitive Impairment Using Decision Tree–Based Machine Learning with Physical, Functional, and Psychosocial Measures in Community-Dwelling Older Adults: Evidence from the Northern Japanese ORANGE Registry

Background Mild cognitive impairment (MCI) is common in later life and represents a key target for early identification and prevention. Scalable, non-imaging approaches using routinely collected community health data may support risk stratification and guide follow-up assessments. Methods We analyzed community health check-up data from Akita Prefecture, Japan. The outcome was binary MCI classification (0 = non-MCI; 1 = MCI). Candidate predictors included demographics, medical history, physical function, and psychosocial measures. Data were split into training (70%) and test (30%) sets using stratification. We trained a decision tree, random forest, and gradient-boosted trees with five-fold cross-validation and hyperparameter tuning. Model discrimination and classification metrics were evaluated on the independent test set. Permutation importance was computed for the best-performing model, and a shallow decision tree was derived using the top-ranked predictors for interpretability. Results The analytic sample included 2,650 participants (non-MCI: n = 1,893; MCI: n = 757). On the test set, the random forest model achieved the highest ROC AUC (0.719). At a 0.5 threshold, accuracy was 0.753, with sensitivity 0.254 and specificity 0.952. Using the Youden threshold (~ 0.256) increased sensitivity to 0.794 while reducing specificity to 0.537. Permutation importance ranked GDS-15 score, osteoporosis, social frailty, and living alone among the top predictors. Conclusions A random forest model demonstrated moderate discrimination for classifying MCI using routinely collected community health variables. The choice of operating threshold had a substantial impact on the sensitivity–specificity trade-off, underscoring the importance of clearly defining intended use and decision thresholds. External validation and prospective evaluation are required before clinical deployment.