Predicting clinical dementia rating scores from neuropsychological testing: a machine learning study in a large clinical cohort

Abstract

To reduce variability of Clinical Dementia Rating (CDR) assessments taken from raters and informants, quantitative and reproducible approaches to support consistent interpretation are needed. We developed multiclass classification models of CDR global and regression models of CDR sum of boxes (CDR-SB) using neuropsychological test results from the Korean version of the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD-K). A total of 6,374 visit-level assessments from 3,460 participants older than 60 in the Catholic Aging Brain Imaging (CABI) database are included. Multiple machine learning models were used to predict CDR scores. Shapley additive explanations (SHAP) were utilized to enhance interpretability. XGBoost and TabPFN models showed comparatively high predictive performance for both CDR global and CDR-SB prediction tasks, achieving accuracies of up to 89% for CDR global and R 2 scores of 0.92 for CDR-SB. Our machine learning models offer quantitative and reproducible estimates of CDR staging that may support more consistent staging across clinical settings by complementing clinical judgment.