Predicting and Understanding Disorders of Consciousness of Sleep Patients Through Multimodal Data Fusion and Temporal Attention Mechanisms

This study introduces MTAMA-DoC (Multimodal Temporal Attention Multitask Analyzer for Disorders of Consciousness), a novel deep learning approach that leverages sleep-related physiological data to predict consciousness states in patients with Disorders of Consciousness (DoC). Employing a transformer-based architecture with task-aware guided masking and joint-optimization training, MTAMA-DoC integrates time series data (heart rate and breath rate) with structured clinical information, addressing the limitations of traditional assessment methods. In a comparative analysis involving 170 patients, our model significantly outperformed existing methods in consciousness state classification (binary classification accuracy: 0.87, F1-score: 0.82; ternary classification accuracy: 0.72, F1-score: 0.64) and demonstrated competitive performance in Glasgow Coma Scale score prediction. The model’s interpretability, achieved through attention mechanisms and feature importance analysis, provides valuable insights into the temporal dynamics of physiological signals and the relative importance of clinical factors in consciousness assessment, highlighting the critical role of cardiovascular and respiratory metrics, as well as sleep architecture features. By demonstrating the efficacy of sleep-related data in consciousness evaluation, MTAMA-DoC not only enhances diagnostic capabilities but also contributes to the theoretical understanding of the sleep-consciousness relationship, offering a non-invasive, continuous monitoring approach that complements traditional clinical methods. This advancement opens new avenues for AI-driven approaches in neurology and sleep medicine, potentially improving diagnosis and care for patients with disorders of consciousness, while paving the way for future investigations into the complex interplay between sleep, consciousness, and brain function.