Time-step level sleep apnea detection using LSTM with adaptive thresholding in polysomnography

Sleep apnea is a prevalent disorder linked to severe health risks. Current diagnosis relies on manual polysomnographic (PSG) analysis, which is time-consuming and subjective. This study aims to develop an objective, automated framework for time-step level apnea detection. We proposed a Long Short-Term Memory (LSTM) network utilizing six PSG signals: PTAF, Thermistor, Thoracic/Abdominal effort, Snore, and SpO2. Approximately 7 h of PSG data were segmented into fixed windows (35 s, 40 s, 45 s), each sampled into 400 time steps. Apnea was classified by thresholding positive time steps (150, 200, 250). Individual signal and SHAP-based analyses identified PTAF, Thermistor, and SpO2 as the most influential features. The optimal configuration used a 40-s window with a threshold of 150, achieving a balanced performance with 88.51% accuracy, 87.9% sensitivity, and 88.79% specificity, outperforming other window sizes. The proposed LSTM-based model effectively generates a continuous probability-based signal for apnea detection. This automated labeling system has the potential to streamline diagnostic workflows, reduce manual workload for clinicians, and provide a consistent, objective framework for sleep apnea assessment.