A Spatio-Temporal Coupling Framework for Cognitive Impairment Screening Based on Heterogeneous Graph Neural Networks and Mamba

Abstract

The 2026 George B. Moody PhysioNet Challenge focuses on utilizing complex and heterogeneous signals derived from polysomnography (PSG) to predict the risk of cognitive impairment. Our team, Kust_MeAI, proposed a Heterogeneous Graph Neural Network (HGNN) framework that achieves spatial coupling of multimodal physiological features by explicitly encoding physiological meta-paths. We align the multimodal signals recorded throughout the night and segment them into consecutive 30-second time windows. For each 30-second PSG segment, we construct a heterogeneous graph comprising five types of nodes: brain, eye movement, electromyography (EMG), cardiac, and respiratory. Each node is initialized as a 64-dimensional vector using modality-specific encoders-for instance, employing 1D convolutional networks with small receptive fields to capture the microscopic, transient features of high-frequency neural signals, and those with large receptive fields to extract the long-range, slow-wave rhythmic characteristics of low-frequency cardiopulmonary signals. Based on physiological knowledge, we define four types of directed meta-paths: Eye-to-Brain (REM coupling), Muscle-to-Brain (muscle tone–arousal association), Heart-to-Brain (autonomic–cortical activation), and Lung-to-Heart (respiratory sinus arrhythmia). Each meta-path utilizes a distinct transformation matrix, and edge weights are dynamically computed via node-level attention mechanisms. Finally, the coupled features-comprising thousands of HGNN outputs generated over the entire night-are fed into a Bidirectional Mamba network for temporal modeling, thereby capturing the macro-level sleep structure and long-range state transitions occurring throughout the sleep process. The proposed approach is evaluated on the public training datasets and achieves an average 5-fold cross-validation score of 0.391 using the Challenge's official scoring metric. Our team, Kust_MeAI, achieved a challenge score of 0.387 on the hidden validation dataset, ranking 89th out of more than 92 participating teams.