Interpretable Machine Learning for Cognitive Aging: Handling Missing Data and Uncovering Social Determinants

TL;DR

This study uses machine learning to predict cognitive aging from social factors, addressing missing data and identifying key predictors of cognitive decline.

Contribution

A novel SVD-based imputation pipeline and interpretable model for cognitive aging with social determinants in data with substantial missingness.

Findings

The SVD-based imputation pipeline effectively handles missing data in both continuous and categorical variables.

Key social determinants were identified as strong predictors of cognitive performance across different age groups.

The framework shows robustness and interpretability for cognitive aging analysis in datasets with missing values.

Abstract

Early detection of Alzheimer’s disease (AD) is critical, as neuropathologic change and modifiable social behavioral risks accumulate years before diagnosis. Identifying higher-risk individuals earlier enables prevention, timely care, and more equitable resource allocation. We study prediction of cognitive performance from social determinants of health (SDOH) using the NIH NIA supported PREPARE Challenge Phase 2 dataset, derived from the nationally representative Mex-Cog cohort within the 2003 and 2012 Mexican Health and Aging Study (MHAS). The target is a validated composite cognitive score covering orientation, immediate and delayed memory, attention, language, constructional praxis, and executive function, which derived from 2021 and 2016 MHAS. We curated features across demographic, socioeconomic, health, lifestyle, psychosocial, and healthcare access domains to capture multidimensional social and behavioral influences on cognitive aging. Substantial missingness was addressed with a singular value decomposition (SVD)-based imputation pipeline, treating continuous and categorical variables separately. This leverages latent feature correlations to recover missing values while balancing reliability and scalability. To better understand the relationship between input features and the composite cognitive score, we conducted a thorough post hoc analysis of the top contributing features, examining the mechanism of these features are associated with cognitive scores. The study further stratified the analysis by age group to explore whether the most predictive features differ across life stages. The proposed framework demonstrates robustness, interpretability, and computational efficiency, underscoring its potential as a practical modeling strategy for MHAS data in AD and MCI detection with substantial missingness across both continuous and categorical features.