RespNet: A Dual-Network Approach for Automated OSA Severity Classification Utilizing PSG Type III Signals

Obstructive Sleep Apnea (OSA) is one of sleep-disordered breathing characterized by repetitive episodes of partial or complete obstruction of upper airway during sleep, leading to significant health risks. Polysomnography (PSG) is gold standard for diagnosing OSA, but the process is labor-intensive and time-consuming which is often inaccessible. This study proposes a novel deep learning-based framework for severity classification of OSA using physiological signals from PSG type III devices. The proposed method comprises two key models: ApneaDetectNet for detecting apnea/hypopnea events called respiratory events and SleepDetectNet for classifying sleep stages. ApneaDetectNet utilizes abdominal respiratory effort (Abdo), nasal pressure (AIRFLOW), and oxygen saturation (SpO2) signals, while SleepDetectNet uses only Abdo signal. Both models are built using convolutional neural networks (CNNs) and bidirectional long short-term memory (Bi-LSTM) networks. An AHI Estimator consequently calibrates the predicted Apnea-Hypopnea Index (AHI) to resemble clinical observations. Three public datasets were employed: Multi-Ethnic Study of Atherosclerosis (MESA), Wisconsin Sleep Cohort (WSC), and University College Dublin Sleep Apnea (UCDDB). These datasets were split into training, validation, and test sets. Results indicated that the proposed model achieved superior or on par performance compared with state-of-the-art models. The MESA model achieved an accuracy of 71.84%, a Kappa of 0.6057, a macro-averaged F1-score of 70.31%, and an R-square of 0.7739, while the WSC model reached an accuracy of 62.16%, a Kappa of 0.4843, a macro-averaged F1-score of 61.90%, and an R-square of 0.6092. The combined MESA & WSC models showed improved overall metrics. External validation with the UCDDB dataset demonstrated 84.00% accuracy with 0.7743 Kappa, a macro-averaged F1-score of 86.69%, and an R-square of 0.7755, showing the model’s robustness. The accuracy, macro-averaged F1-score, and generalizability of the model suggest potential for real-world application.