An approach to arousal disorder classification using deformable convolution and adaptive multiscale features in EEG signals

Diagnosing sleep phases, arousal problems, and apnea episodes using Polysomnography (PSG) signals is often time-consuming. However, automated approaches have demonstrated promising results. Early detection of sleep disturbances can facilitate the diagnosis of neuropathologies before they progress. Given the significance of sleep events in diagnosing and treating sleep disorders, automated arousal disorder classification is increasingly crucial. Timely intervention for arousal disorders, if detected early, can potentially slow the progression of neuropathological illnesses such as Multiple System Atrophy (MSA), Parkinson's, and Alzheimer's disease. While PSG signals are sometimes necessary for clinical diagnoses, Electroencephalography (EEG) is often underutilized due to its labor-intensive nature. Automated methods for detecting, analyzing, and classifying arousal disorders offer significant benefits. In this research, we propose a novel method to classify arousal disorders from EEG data and extract post-classification diagnostic features. To our knowledge, this is the first instance of such categorization achieved using a deformable convergence network. Our proposed model, a hierarchical multiscale deformable attention module, excels at detecting complex and abnormal patterns in EEG data. We apply this model after segmenting EEG data into 30-second windows and generating spectrogram images. This study aims to evaluate our model's effectiveness in handling imbalanced classification and reducing false positive rates in arousal detection. We analyzed data from 994 participants in the 2018 PhysioNet Challenge study who experienced sleep-related micro- and macro-arousal events. Our method achieved an accuracy rate exceeding 96 %, outperforming other multi-scale channel attention modules. This approach enables future studies to objectively, efficiently, and precisely examine various arousal disorders. Additionally, we investigated the effect of multimodal signal fusion and observed that integrating EEG with ECG significantly enhances classification performance, highlighting the value of combining cortical and autonomic information in arousal disorder detection.