A lightweight multi-modal CNN-LSTM based sleep apnea severity detection using ECG and respiratory effort signals

Sleep Apnea (SA) is widespread sleep disorder with intermittent breathing interruption during sleep, typically undiagnosed and with severe health consequences. In this work, we present a better deep learning paradigm for automatic SA severity classification based on a blend of respiratory signals (thoracic, abdominal, airflow) and cardiac signals (ECG). Contrary to conventional methods based on peripheral oxygen saturation (SpO2), our method uses more physiology-relevant signals to improve detection accuracy. We present a hybrid deep architecture that combines Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, supplemented with cross-signal attention mechanisms. CNN sub-networks learn morphological and spatial features from raw bio signals, while LSTM networks learn temporal interactions critical to apnea event detection. The merged architecture allows multi-class severity classification (Normal, Mild, Moderate, Severe) via AHI computations. The model was evaluated on a publicly available polysomnography dataset and recorded 90.35% classification accuracy in severity estimation, with precision, recall, and F1-scores of 0.91 across all severity classes, much better than conventional machine learning and previous deep learning approaches. These findings demonstrate the potential of our multi-modal technique for efficient SA severity detection in clinical and home settings.