Deep Learning AD Detection Model based on a Two-Layer Ensemble Module with Data Augmentation and Contrastive Learning

Abstract

Abstract—Alzheimer's Disease (AD) is a long-term disease that gradually decreases cognitive functioning, such as thinking, memory and behavior. In 2015, 29.8 million AD cases were recorded and 1.9 million AD-related deaths were reported worldwide. Early detection and intervention are critical for such a deadly and costly disease. I present to tackle the detection of AD and its severity using a deep-learning architecture that consists of a two-stage ensemble system with contrastive learning and data augmentation. I evaluated it on the ADReSS Challenge's dataset, which is subject-independent and balanced in terms of age and gender. When compared against a one-stage ensemble baseline approach, my two-stage ensemble system was able to achieve better results, with a F1-score of 95.7% in the AD classification task, and an RMSE score of 5.432. Moreover, I found that the data augmentation can effectively improve the robustness of the AD detection performance, particularly when there are sensor noises in the training and test data. Besides data augmentation, I also explored whether contrastive loss can further boost the robustness, and the results showed that contrastive learning might not be necessary when we have data augmentation.