Multi-relational mixed perception based short time series prediction of Alzheimer’s disease progression

Alzheimer’s disease (AD), as a neurodegenerative disease, is long-lasting and irreversible. Therefore, disease detection and intervention in the early stage of AD are crucial. However, the existing deep learning methods for analyzing longitudinal data in AD mostly use long series models. This makes it difficult to fit short time series data with long follow-up intervals in AD. Moreover, most studies currently model pairwise subject relationships using pairwise connections in population graphs, without considering high-order relationships between these subjects. This study aimed to propose a Multi-relational Mixed Perception based Short Time Series model (MMPSTS) for predicting AD progression. First, a time series evolutionary generative adversarial network was proposed, which enhanced the dynamic fitting ability of data evolution and improved the generalization of long time series learning models to short time series. Second, a multi-relational mixed perception module with global noise filtering was designed to capture pairwise and high-order interaction relationships. Specifically, the module enabled the joint modeling of pairwise and higher-order relationships by constructing a noise-filtered population graph and a population hypergraph. The mixed representations of multiple relationships were fed to extended long short-term memory to understand the distribution of time series. We compared the method with other state-of-the-art methods for clinical diagnosis and score prediction, thereby demonstrating the superiority of the proposed method. The experimental results were also combined with pathological analysis to reflect the plausibility of the results. Our code is available at https://github.com/nakoyyx/MMPSTS . • Improve the generalization of long time series learning models to short time series. • Efficient learning of pairs and high-order relationships between subjects. • Non-image data guides image data to generate relational networks. • Adaptable generation of high-order relationships between subjects.