A Machine Learning Approach to Predict Cognitive Decline in Alzheimer Disease Clinical Trials

Among the participants of Alzheimer disease (AD) treatment trials, 40% do not show cognitive decline over 80 weeks of follow-up. Identifying and excluding these individuals can increase power to detect treatment effects. We aimed to develop machine learning-based predictive models to identify persons unlikely to show decline on placebo treatment over 80 weeks. We used the data from the placebo arm of EXPEDITION3 AD clinical trial and a subpopulation from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Participants in the EXPEDITION3 trial were patients with mild dementia and biomarker evidence of amyloid burden. For this study, participants were identified as those who demonstrated clinically meaningful cognitive decline (CMCD) or cognitively stable (CS) at final visit of the trial (week 80). Machine learning-based classifiers were trained to classify participants into CMCD vs CS groups using combinations of demographics, APOE genotype, neuropsychological tests, and biomarkers (volumetric MRI). The results were developed in 70% of the EXPEDITION3 placebo sample using 5-fold cross-validation. Trained models were then used to classify the participants in an internal validation sample and an external matched sample ADNIAD. Eight hundred ninety-four of the 1,072 participants in the placebo arm of the EXPEDITION3 trial had necessary follow-up data, who were on average aged 72.7 (±7.7) years and 59% female. 55.8% of those participants showed CMCD (∼2 years younger than those without) at the final visit. In the independent validation sample within the EXPEDITION3 data, all the models showed high sensitivity and modest specificity. Positive predictive values (PPVs) of models were at least 11% higher than base prevalence of CMCD observed at the end of the trial. The subset of matched ADNI participants (ADNIAD, N = 105) were aged 74.5 (±6.4) years and 46% female. The models that were validated in ADNIAD also showed high sensitivity, modest specificity, and PPVs of at least 15% higher than the base prevalence in ADNIAD. Our results indicate that predictive models have the potential to improve the design of AD trials through selective inclusion and exclusion criteria based on expected cognitive decline. Such predictive models need further validation across data from different AD clinical trials.