Screening mild cognitive impairment using aspects of personal, social, and functional lifestyle: Machine Learning Approaches

TL;DR

This study uses machine learning to predict mild cognitive impairment using lifestyle and social factors, finding that simpler models work best with age, stress, and social factors as key indicators.

Contribution

The study introduces a practical ML approach for MCI screening using routinely collected lifestyle data and identifies key predictors with high interpretability.

Findings

Most ML models achieved good discrimination with AUROC > 0.8 for predicting MCI.

Logistic regression outperformed complex models in predictive accuracy and interpretability.

Age, ethnicity, functional difficulties, social disconnectedness, and stress were consistently key predictors.

Abstract

Mild cognitive impairment (MCI) signals cognitive decline beyond normal aging and increases dementia risk. Early identification enables preventative interventions, yet many patients in primary care go undetected. This study examines whether machine learning (ML) models can predict MCI using routinely collected personal, social, and functional lifestyle factors and identifies the most important predictors.

Data from round 2 and 3 of the National Social Life, Health, and Aging Project was used, including 4,586 older adults with complete Montreal Cognitive Assessment (MoCA) scores. Predictors included demographics, childhood experiences, health behaviors, psychosocial measures, and functional difficulties. Eight ML models—including elastic net, multivariate adaptive regression splines, random forest, oblique random forest, boosted trees, decision trees, and a stacked ensemble—were trained and tuned using repeated cross-validation, with 20% of the dataset withheld for final testing. Model performance was assessed using area under the receiver operator curve (AUROC), accuracy, sensitivity, specificity, and Matthew’s correlation coefficient (MCC).

Most models achieved good discrimination (AUROC > 0.8), with the stacked ensemble performing best (AUROC = 0.823; MCC = 0.462). The best individual model was logistic regression (AUROC = 0.818). Across models, key predictors of MCI included age, ethnicity, functional difficulties, social disconnectedness, and perceived stress.

Logistic regression outperformed more complex machine learning models, providing the best combination of predictive accuracy and interpretability for identifying MCI. Across models, age, ethnicity, functional difficulties, social disconnectedness, and stress consistently emerged as key predictors, highlighting their central role in cognitive health. These findings suggest that psychosocial and functional measures can serve as practical indicators for those who need to be screened early for MCI, offering an opportunity for timely intervention and support. However, future work should include longitudinal data and clinical diagnoses to validate and refine these predictive tools.

Figures