Classifying Alzheimer’s from fMRI Data using Convolutional Networks

Alzheimer’s disease is a neurodegenerative disorder that causes 60 – 70% of dementia and affects 5.3 million Americans. Symptoms begin mild but progressively worsen, leading many to go undiagnosed until late stages. Alzheimer’s has no cure, but early diagnosis leads to better patient care and planning before full cognitive impairment. Current diagnostic procedures involve a combination of medical records, cognitive tests, and hours of skilled doctors’ time. An automated method of diagnosing Alzheimer’s would be cheaper and faster than current diagnostic methods. Recently, machine learning algorithms have been developed to diagnose Alzheimer’s from fMRI data. However, these algorithms have only achieved high accuracies for binary classification, rather than classifying the stages of cognitive impairment. This paper presents a machine learning model that classifies five stages of cognitive impairment. The proposed algorithm achieves a state-of-the-art subject-level classification accuracy of 85.1%. Additionally, by analyzing the brain regions of interest to the model, the model indicates the hippocampus region, areas of white matter, and straight sinus are most important for diagnosing Alzheimer’s.