Machine-Learning Prediction of Cognitive Impairment and Dementia From the Health and Retirement Study: A Phased Evaluation of Preprocessing and Its Limits (Preprint)
Preprint
- 13 October 2026
- preprint
- Published by JMIR Publications Inc. in JMIR Preprints
Abstract
BACKGROUND Population surveys such as the Health and Retirement Study (HRS) provide scalable, low-cost tabular predictors of cognitive status. However, pervasive missingness, imperfect administrative labels, and extreme class imbalance constrain what predictive models can recover. OBJECTIVE This study evaluated how far structured preprocessing and cohort construction can improve machine-learning prediction of cognitive impairment and dementia from HRS data, while characterizing the remaining limits imposed by data quality and construct validity. METHODS We used curator-guided variable mapping, iterative imputation, composite indices, explicit missingness indicators, and reproducible dataset tiers. A phased pipeline evaluated multiclass and binary endpoints, tiered predictor sets, subgroup performance, model explanations, calibration, decision curves, and model-label concordance. RESULTS Relative to exploratory precuration baselines, the curated pipeline improved class-sensitive multiclass performance, with macro one-versus-rest receiver operating characteristic area under the curve reaching approximately 0.77 for LightGBM on the iteratively imputed HRS cohort. Missingness indicators produced modest additional discrimination. For the dementia-versus-residual binary endpoint, the area under the curve was approximately 0.81; sensitivity at a probability cutoff of 0.5 was approximately 0.13. Hierarchical refinement increased dementia sensitivity to approximately 0.33 under the archived decision rules. High rates of dementia false negatives persisted across analyses. CONCLUSIONS Structured preprocessing supported more interpretable evaluation, but frequent dementia false negatives remained. The models require further methodological and external validation before a clinical screening role can be considered.Keywords
This publication has 11 references indexed in Scilit:
- Improving accuracy in the estimation of probable dementia in racially and ethnically diverse groups with penalized regression and transfer learningAmerican Journal of Epidemiology, 2025
- Early Identification of Cognitive Impairment in Community Environments Through Modeling Subtle Inconsistencies in Questionnaire Responses: Machine Learning Model Development and ValidationJMIR Formative Research, 2024
- TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methodsBMJ, 2024
- Improving clinical efficiency in screening for cognitive impairment due to Alzheimer'sAlzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring, 2023
- Machine learning methods for predicting progression from mild cognitive impairment to Alzheimer’s disease dementia: a systematic reviewAlzheimer's Research & Therapy, 2021
- Predicting Cognitive Impairment and Dementia: A Machine Learning ApproachJournal of Alzheimer’s Disease, 2020
- A guide to deep learning in healthcareNature Medicine, 2019
- Unsupervised Machine Learning to Identify High Likelihood of Dementia in Population-Based Surveys: Development and Validation StudyJournal of Medical Internet Research, 2018
- The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced DatasetsPLOS ONE, 2015
- DementiaAlzheimer Disease & Associated Disorders, 1996