PREDICTING DECLINE IN CLINICAL ASSESSMENTS OF ALZHEIMER'S DISEASE WITH MACHINE LEARNING AND 3D BRAIN MRI

Alzheimer's disease (AD) is a progressive neurodegenerative disorder affecting around 50 million people globally. An important objective of AD research is to predict the likelihood that a subject with Mild Cognitive Impairment (MCI) -an intermediate stage between healthy aging and Alzheimer's disease-will progress to AD over a specific time span, typically 2 years. Such prognostic models are crucial in assessing effects of risk factors or protective factors (such as novel anti-amyloid drugs). Here we benchmarked a range of machine learning models to predict 2-year decline in clinical scores from measures derived from 3D anatomical brain MRI, age, sex, APOE genotype, and baseline standardized test scores- Mini Mental State Exam (MMSE) and Alzheimer's Disease Assessment Score Cognitive (ADAS-cog) - in 2,448 ADNI participants (1,132 with MCI, 883 Healthy Controls (HC), and 433 with Dementia). We began by predicting change in MMSE and ADAS-Cog values across a 2-year period for both control participants and those with MCI, regardless of their conversion status. Then, as a post hoc sensitivity analysis, we examined the prediction accuracy in converters (HC -> MCI, MCI -> AD). For predicting the change in MMSE over a 2-year interval, the kernel-PCA compacted features achieved a correlation coefficient of 0.803 in Controls and 0.784 in the MCI group. For ADAS-cog (Alzheimer's Disease Assessment Scale-Cognitive subscale), the model achieved a correlation coefficient of 0.861 for Controls and 0.826 for the MCI group.