Machine Learning Approaches To Diagnose Alzheimer's Disease

Abstract

This study uses machine learning (ML) approaches to model the disease severity of patients with Alzheimer’s Disease (AD). Magnetic Resonance Imaging (MRI) images of AD patients were labeled as non-dementia, very mild dementia, mild dementia, and moderate dementia by a physician and were used to train, validate, and test 3 different models. The models included a convolutional neural network (CNN), VGG-16, and VGG-16 + SMOTE. The CNN model was a basic neural network. The VGG-16 model was pre-trained on image data, and the VGG-16 + SMOTE employed a data preprocessing technique that balanced the number of images in each class before training on the VGG-16 pre-trained CNN. The CNN model was the most effective, with the highest accuracy of the three models: 0.98. The VGG-16 + SMOTE model is the second most effective, with an accuracy of 0.97. The least effective model is the one with only VGG-16, with an accuracy of 0.87. The data suggests that the CNN model accurately diagnoses the level of dementia in AD patients and has the potential to be used in medical practices. While using SMOTE significantly improved the accuracy of the least accurate model, the VGG-16 + SMOTE model is still slightly less effective than a regular CNN and requires much more time to run. Thus, a regular CNN is more accurate and efficient.