Evaluation of Deep Learning Models for Early-Stage Alzheimer's Disease Screening

Abstract

Alzheimer's disease, a progressive neurodegenerative disorder, poses significant challenges in its early-stage identification and accurate prediction of disease progression. In the early stages of Alzheimer's, patients are usually undetected because symptoms are difficult to identify and previously, no treatments for Alzheimer’s were available. In 2022 and 2023 the FDA approved three drugs for treating Alzheimer’s disease (Donanemab, Aducanumab, and Lecanemab), a breakthrough in treating the disease. The approved drugs have only been shown to be effective in patients with early cognitive decline. As a result, there is now an urgent need to screen millions of patients for Alzheimer’s disease which presents significant financial challenges and substantial time allocation. Machine learning models have shown accuracy in detecting Alzheimer’s and other disease outcomes from imaging data including MRI scans. However, previous work on Alzheimer’s disease has often focused on identifying patients with late cognitive decline and often using older model architectures and methods. In this study, we train multiple vision-based machine learning models to detect patients in the early stages of Alzheimer’s from MRI imaging data. We overcome technical challenges including class imbalance which diminishes model performance. Furthermore, we apply transformer-based models that utilize an attention-based mechanism that has led to breakthroughs in other areas including LLMs (ChatGPT). Our experiments show that ResNet-18 with class weighting, a batch size of 32, and a learning rate of 1E-06 was the optimal model for identifying early cognitive disease. This model produced F-scores as high as 0.984 for the MCI and EMCI stages, and 0.982 for cognitive normal patients.