Conversion-aware forecasting of Alzheimer's disease via featurewise attention

Alzheimer's disease (AD) is a neurodegenerative disorder that causes cerebral atrophy, affecting memory and cognitive functions. An earlier stage, where the symptoms do not overwhelm the patients' daily activities, is known as mild cognitive impairment (MCI). When the underlying cause is AD, MCI patients convert to AD in future stages of the disease. Early detection of the conversion is a vital step in preventative treatment planning. However, identifying this transition is challenging due to the rarity of such events in public datasets. This thesis introduces a deep learning framework to improve conversion detection performance. For this purpose, the proposed architecture is intended to provide insights into the impact of the input biomarkers on conversion detection. It comprises neural blocks to encode attributes and time into a shared space where cross-attention between time and attributes is computed to aggregate the embeddings. While capturing feature importances learned by the model, the temporal information is incorporated into the network with time embeddings. The entire architecture is trained end-to-end to forecast the risk of developing AD. Experiments with two publicly available datasets show promising performance in the early detection of MCI to AD conversion compared to traditional machine learning models and similar deep models. The proposed framework enhances early detection of MCI to AD conversion by effectively integrating temporal dynamics and feature importance. This approach holds promise for improving sensitivity in prediction and supporting more targeted preventative treatment strategies for at-risk individuals.