Artificial Intelligence–Enabled Dementia Risk Prediction for Smart and Sustainable Healthcare: An Interpretable Machine Learning Study Using NHATS

Abstract

Dementia is an increasing public health challenge, yet scalable methods for early risk detection using non-clinical data remain limited. This study develops and evaluates interpretable machine learning models to predict dementia risk among older adults using nationally representative longitudinal data. Data were sourced from the National Health and Aging Trends Study (NHATS, 2011–2022) and included 5984 community-dwelling U.S. adults aged 65 and older who were dementia-free at baseline. Dementia onset was identified using the validated NHATS classification algorithm based on cognitive assessments, proxy reports, and physician diagnoses. After data preprocessing and feature engineering, missing values in continuous variables were imputed with k-nearest neighbors, while categorical variables were handled via one-hot encoding and mode-based imputation. Five supervised machine learning algorithms were trained and evaluated through stratified cross-validation, using performance metrics that account for class imbalance. Among these models, XGBoost showed the strongest overall performance, achieving the highest classification accuracy (0.881 ± 0.004), the lowest Brier score (0.094 ± 0.002), and the highest ROC–AUC (0.823 ± 0.005), with RF showing comparable results. Explainable AI analyses with SHapley Additive exPlanations (SHAP) consistently identified digital technology use, outdoor activity frequency, and social network size as the most influential predictors across models. These findings indicate that interpretable machine learning based on non-clinical, modifiable behavioral and social factors can support scalable, prevention-focused dementia risk assessment and inform prevention-oriented strategies that promote digital inclusion and social engagement among older adults.

Ähnliche Arbeiten