AI-Driven Prediction of Cognitive Decline Using XGBoost: Insights from a cohort

Background:
Predicting cognitive decline is crucial for early intervention in aging populations. CHARLS provides a large-scale dataset for developing machine learning models using health, behavioral, and socioeconomic data.
Methods:
We used CHARLS (2011-2020) data to train an XGBoost model, a gradient boosting algorithm known for handling structured data effectively. The outcome was defined as a ≥1.5 SD decrease in memory recall over four years. Features included demographics (age, sex, education), health conditions (hypertension, diabetes, cardiovascular disease), mental health (CES-D depression score), lifestyle factors (smoking, alcohol use, physical activity), and social factors (living alone, social support, marital status).

For comparison, we referenced OASIS, where cognitive decline prediction used neuroimaging data. Unlike OASIS, CHARLS-based models rely on behavioral and health predictors rather than MRI biomarkers. We applied Recursive Feature Elimination (RFE) for feature selection and SHAP analysis to interpret feature importance. Model performance was evaluated using AUC-ROC, F1-score, and precision-recall metrics.

Results:
The XGBoost model achieved an AUC-ROC of 0.81, outperforming logistic regression (AUC = 0.72). Top predictors included age, depression (CES-D score), low education, hypertension, and social isolation. SHAP analysis highlighted depression and social isolation as stronger predictors than some physical health conditions. The model maintained high accuracy across subgroups (urban vs. rural, male vs. female).

Conclusion:
XGBoost effectively predicts cognitive decline using CHARLS data, emphasizing the role of mental health and social factors. This AI-driven approach supports early risk assessment and intervention strategies in public health.