Machine Learning for Predicting Progression of Alzheimer’s Disease

Abstract

In Alzheimer’s disease (AD), amyloid- (A ) peptides aggregate in the brain forming plaques. Strangely, these plaques are persistent in both severely cognitively impaired and cognitively normal individuals. Therefore, it is of big value to investigate whether other factors cause some patients with A plaques to have AD dementia and others not. We used data from The Alzheimer’s Disease Neuroimaging Initiative (ADNI) to study the differences in individuals with evidence of A plaques and those without. Furthermore, we tried to predict how the cognitive ability of individuals with plaques would progress in the next four years using machine learning techniques. Random forest and elastic net estimators were created, predicting the decline in cognitive test scores as well as diagnosis change of patients only using data from their first visit. The best regression models, predicting the change in cognitive test scores achieved R2 scores of 0.428 to 0.580 while the classification models, predicting whether a patient will get a worse diagnosis achieved a weighted F1 score of 0.817. Moreover, patients with A plaques seem to decline faster than those without. The most important features for predicting future cognitive decline were cognitive tests indicating that already cognitive impaired individuals would deteriorate more. Other important factors were fluorodeoxyglucose (FDG) obtained from positron emission tomography and proteins measured in cerebrospinal fluid. These models could possibly, with further development, be used in clinical settings as an aid for evaluating how the cognitive function of an individual with A plaques will develop in the near future.