OBJECTIVES

ML classification models of univariate outcomes have been widely discussed for AD; however, due to censoring, ML survival models for predicting AD progression (bivariate outcome) are not extensively studied. The objectives were to 1) predict AD progression using ML survival models and 2) compare the model performance.

METHODS

Data of 527 subjects who were cognitively normal at baseline was extracted from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. The outcome was time-to-first mild cognitive impairment (MCI) or dementia, whichever happened first. Subjects free of MCI/dementia were censored. Model covariates were identified based on publications using ADNI, which ended up with age, gender, education, Apolipoprotein E gene status and longitudinal results from 20 neuropsychological tests. Models were built using two sets of covariates, successively: 1) baseline values for all covariates; 2) baseline values for all covariates plus average changing rates of the neuropsychological test scores during follow-up time. Survival tree and random forest models were trained (R, version 3.4.3) on 80% randomly selected subjects while the rest 20% were used for testing. Integrated Brier Score (IBS) was used to compare model performance.

RESULTS

Survival tree models sub-grouped subjects based on their covariates values and provided the probability of developing AD at each time point, i.e., the Kaplan-Meier curve, for each subgroup; whereas random forest models predicted such probability for each type of subject. Most neuropsychological tests (e.g., Logical memory delayed recall) were found more important than demographic and genetic variables on predicting AD. The changing rates of test scores enhanced model performance according to IBS. Random forest models outperformed survival tree models on both training and test datasets, with 7.09% and 8.50% error rates respectively when including changing rates covariates.

CONCLUSIONS

ML survival models provide new aspects of modeling AD progression and diagnosis, as well as the dependence on individual-level characteristics, e.g., neuropsychological tests.