SomniTrack: A Multimodal Attention Framework with Neuro-Topographic Sleep Staging and Disorder Classification

Abstract

Sleep staging and obstructive sleep apnea (OSA) detection are central to sleep medicine, yet remain con-strained by the complexity of full polysomnography (PSG) and the subjectivity of manual scoring. We present SomniTrack, a multimodal framework that integrates EEG and ECG to automate sleep assessment without relying on respiratory channels. The framework comprises two complementary pipelines. CORTEX-EEG extracts oscillatory and temporal descriptors via denoising and band decomposition, then applies spatio–temporal attention to highlight informative channels and time windows, enabling neuro-topographic im-portance analysis that identifies spatially and temporally significant EEG regions associated with sleep stage transitions. CARDIA-ECG employs R-centered sub-epochs to align features to cardiac dynamics, applies variational mode decomposition on cardiorespiratory signals to separate morphology and rhythm compo-nents, uses adaptive feature weighting for adjusting contributions of handcrafted descriptors, and introduces temporal attention for giving the most weight to those sub-epochs that bear greatest entropy of apnea-related irregularities and stage transitions. The two pipelines generate features grounded in physiology, which are processed by light neural modules (a multi-layer perceptron for EEG; a recurrent neural network for ECG), followed by fusion for joint decision-making. DOD-H/O, ISRUC, and MASS show strong performance with thorough evaluations. The accuracies for sleep staging are respectively 94.3%, 91.2% and 93.9% with stable training–validation dynamics and good class separability in ROC analyses. The framework achieves an accuracy of 94.2%, 95.3%, and 96.2% for OSA risk assessment, respectively, across all datasets. The combi-nation of EEG and ECG is a scalable and interpretable solution for the multimodal analysis of sleep.