Alzheimer's disease prediction algorithm based on hippocampal longitudinal hybrid morphological features.

Background: Alzheimer's disease (AD) is a progressive neurodegenerative disorder characterized by cognitive decline; this decline is closely linked to hippocampal morphological changes observed in structural magnetic resonance imaging (MRI). However, the existing AD prediction models have not fully explored the spatiotemporal correlation of hippocampal morphological features. To address this limitation, this study aims to develop a longitudinal prediction framework that captures both the temporal evolution and spatial distribution of hippocampal morphological alterations. Methods: In this paper, we propose a novel deep learning framework for predicting the clinical progression of AD, which consists of a multi-view feature fusion convolutional network (M-FCN) and a bidirectional gated recurrent unit (Bi-GRU). The proposed M-FCN is based on the three-dimensional (3D) topological structure features of the hippocampus that introduces thickness features and heat kernel signature (HKS) to encode hippocampal morphological atrophy features. We utilize these features to construct a deep 3D hippocampus features description system for capturing the micro and macro structural changes of hippocampus. Hence, the task driven attention mechanism for prediction can effectively identify significant morphological changes caused by AD. The Bi-GRU module identifies inter sequence patterns and studies the temporal correlation between longitudinal features of hippocampus. Results: The proposed method was evaluated using longitudinal T1-weighted MRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) (n=221). Compared with the existing AD prediction models, the correspondence between AD-related structural changes and clinical neurodegeneration indicators can be more accurately captured by our proposed deep learning model. The predictive performance was evaluated using root mean square error (RMSE), correlation coefficient (CC), and 95% confidence interval (CI). For the prediction of Mini-Mental State Examination (MMSE) scores, the model achieved a RMSE of 2.34 (95% CI: 2.27-2.45, CC =0.72) at M18, 2.58 (95% CI: 2.52-2.64, CC =0.77) at M24, and 2.60 (95% CI: 2.54-2.66, CC =0.83) at M36. Conclusions: These results highlight the effectiveness of the proposed model in leveraging the spatiotemporal correlation of hippocampal morphology to provide high accuracy and reliable predictions.