Predictive modeling of Alzheimer's disease progression: Integrating temporal clinical factors and outcomes in time series forecasting

Abstract

Alzheimer's disease is a complex neurodegenerative disorder that profoundly impacts millions of individuals worldwide, presenting significant challenges in both diagnosis and treatment. Recent advances in deep learning-based methods have shown promising potential for predicting disease progression using multimodal data. However, the majority of studies in this domain have predominantly focused on cross-sectional data, neglecting the crucial temporal dimension of the disease's progression. In this study, we propose a novel approach to predict the progression of Alzheimer's disease by leveraging a multimodal time-series forecasting system based on graph representation learning. Our approach incorporates a Temporal Graph Network encoder, employing k-nearest neighbors and Cumulative Bayesian Ridge with high correlation imputation to generate graph node embeddings at each time step. Furthermore, we employ an Encoder-Decoder architecture, where a Graph Attention Network translates a dynamic graph into node embeddings, and a decoder estimates future edge probabilities. When utilizing all available patient features in the ADNI dataset, our proposed method achieved an Area Under the Curve (AUC) of 0.8090 for dynamic edge prediction. Furthermore, for neuroimaging data, the AUC improved significantly to 0.8807.