Abstract Introduction Sleep apnea imposes significant cardiovascular and neurocognitive burdens through recurrent nocturnal respiratory disruptions. Timely and accurate detection remains imperative to mitigate downstream complications. While polysomnography remains the diagnostic gold standard, its high cost, technical complexity, and limited accessibility pose substantial barriers to widespread screening and monitoring. In this study, we developed and validated a machine learning–based approach to detect sleep apnea using photoplethysmography signals alone, potentially enabling scalable and non-invasive monitoring. Methods Photoplethysmography signals were acquired from 50 individuals at a clinical sleep center, including 25 participants with diagnosed sleep apnea and 25 individuals undergoing positive airway pressure therapy. Signals were recorded via finger probes at a sampling frequency of 100 Hz. From each subject, 120-second epochs were extracted, yielding 50 engineered features per epoch. All features were standardized using Z-score normalization, and extreme values exceeding ±3 standard deviations were excluded. Feature selection based on importance metrics resulted in 18 dominant features. A linear support vector machine model was developed and evaluated in two phases. First, leave-one-subject-out cross-validation was performed within the sleep apnea group to assess baseline classification accuracy. Second, the trained model was tested on the independent positive airway pressure group to evaluate generalizability. Subject-level classification was determined by averaging epoch-wise probabilities, with a threshold of 0.455 applied to delineate sleep apnea. Results The support vector machine model achieved a subject-level classification accuracy of 88.0% in both the sleep apnea and positive airway pressure groups. Prediction probabilities were consistently above the classification threshold in the sleep apnea group and predominantly below threshold in the positive airway pressure group, indicating robust separation and stable model performance across cohorts. Conclusion This study demonstrates the feasibility of detecting sleep apnea with high fidelity using photoplethysmography signals and a machine learning framework. These findings provide a technical foundation for future wearable-based and home-based sleep monitoring systems. Further validation in larger and more diverse populations, as well as in real-world clinical and home environments incorporating multimodal physiological data, will be essential for clinical translation. Support (if any) This work was supported by the National Research Foundation of Korea (RS-2025-02263045 to D.C.). Contact information: E.K. (ky.eo@slowave.io); *Corresponding author: H.K. (hkim17@skku.edu).
Eo et al. (Fri,) studied this question.