IEEE Access (21693536)13pp. 44607-44619
Alzheimer's disease (AD) presents a significant global health challenge, necessitating accurate and early prediction methods for effective intervention and treatment planning. In this work, a novel approach to meta-learning for the prediction of AD is proposed, which leverages the combined power of neural processes (NPs) and transformer architectures. We introduce a framework that integrates NPs with a transformer encoder to model the complex temporal dependencies inherent in longitudinal health data, where our model learns to capture subtle patterns and variations indicative of disease progression. The novelty of our approach lies in the fusion of NPs, renowned for their ability to model stochastic processes, with transformer architectures, known for their ability to capture long-range dependencies. This combination enables our model to effectively adapt to individual patient trajectories and generalize across diverse populations, enhancing its predictive performance and robustness. We trained our proposed model with the Alzheimer's Disease Prediction Of Longitudinal Evolution dataset (TADPOLE), which contains three classes: Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and AD. The experimental results demonstrate that the proposed model enhances the prediction of these models in terms of mAUC, Recall, and Precision by 0.937 - 0.014, 0.920 - 0.010, and 0.923 - 0.009, respectively. These findings prove the efficacy of the proposed framework in accurately predicting the progression of AD, highlighting its potential for early detection and personalized treatment strategies. © 2013 IEEE.
