Simplifying Alzheimer's Disease Monitoring: A Novel Machine-Learning Approach to Estimate the Clinical Dementia Rating Sum of Box Changes Using the Mini-Mental State Examination and Functional Activities Questionnaire.

Abstract

Primary outcome measure in the clinical trials of disease modifying therapy (DMT) drugs for Alzheimer's disease (AD) has often been evaluated by Clinical Dementia Rating sum of boxes (CDRSB). However, CDR testing requires specialized training and 30-50 minutes to complete, not being suitable for daily clinical practice. Herein, we proposed a machine-learning method to estimate CDRSB changes using simpler cognitive/functional batteries (Mini-Mental State Examination [MMSE] and Functional Activities Questionnaire [FAQ]), to replace CDR testing. Baseline data from 944 ADNI and 171 J-ADNI amyloid-positive participants were used to build machine-learning models predicting annualized CDRSB changes between visits, based on MMSE and FAQ scores. Prediction performance was evaluated with mean absolute error (MAE) and R2 comparing predicted to actual rmDeltaCDRSB/rmDeltayear. We further assessed whether decline in cognitive function surpassing particular thresholds could be identified using the predicted rmDeltaCDRSB/rmDeltayear. The models achieved the minimum required prediction errors (MAE < 1.0) and satisfactory prediction accuracy (R2>0.5) for mild cognitive impairment (MCI) patients for changes in CDRSB over periods of 18 months or longer. Predictions of annualized CDRSB progression>0.5, >1.0, or >1.5 demonstrated a consistent performance (i.e., Matthews correlation coefficient>0.5). These results were largely replicated in the J-ADNI case predictions. Our method effectively predicted MCI patient deterioration in the CDRSB based solely on MMSE and FAQ scores. It may aid routine practice for disease-modifying therapy drug efficacy evaluation, without necessitating CDR testing at every visit.