What question did this study set out to answer?

This research aims to enhance sleep stage classification accuracy by combining CPAP and PPG signal modalities.

May 10, 2026SLEEP0 citations

0655 Probabilistic Ensembling of PPG and CPAP Modalities for Sleep Stage Classification

HCHsin‐Yu ChenChi Mei Medical Center HWHau-Tieng WuCourant Institute of Mathematical Sciences CCCheng‐Yao ChenPratt Institute

Key Points

This research aims to enhance sleep stage classification accuracy by combining CPAP and PPG signal modalities.
Trained separate 1D convolutional neural networks for CPAP and PPG signals.
Input sequences segmented into 30-second epochs; optimized using categorical cross-entropy loss.
Combined model predictions through probability-level ensembling to maximize macro-F1 score.
The ensembled model achieved an F1 score of 0.696 and Cohen’s kappa of 0.531.
Best-case performance reached F1 = 0.817 and κ = 0.780.
PPG-based and CPAP-based models had initial scores of 0.669/0.497 and 0.585/0.384, respectively.

Abstract

Abstract Introduction Continuous Positive Airway Pressure (CPAP) therapy is the standard treatment for obstructive sleep apnea-hypopnea syndrome, and photoplethysmography (PPG) sensors are widely used in wearable devices. Both signal sources capture rich physiological patterns during sleep, including heart rate and respiratory dynamics. Combining these modalities may improve automatic sleep stage classification. Improved sleep stage accuracy can support a smoother titration process by supplying more precise information for optimizing pressure settings. Methods Separate 1D convolutional neural network models were trained for CPAP and PPG signals. PPG features included the PPG envelope, instantaneous heart rate (IHR), blood oxygen saturation (SpO₂), and actigraphy (ACTG). CPAP features included airflow, positive airway pressure (PAP), and instantaneous respiratory rate (IRR). Input sequences were segmented into 30-second epochs, and models were optimized using categorical cross-entropy loss with the Adam optimizer and early stopping. Predictions from both models were combined via probability-level ensembling using a weighted sum of softmax outputs, with weights selected on a validation set to maximize the macro-F1 score. Results The Duke University Hospital dataset (hereafter referred to as Duke) includes 36 subjects who underwent split-night polysomnography with CPAP titration using a ResMed S9 VPAP Tx device, while simultaneously wearing a fingertip PPG sensor (PranaQ TipTraQ). On this dataset, the PPG-based and CPAP-based models achieved overall macro-F1/Cohen’s kappa scores of 0.669/0.497 and 0.585/0.384, respectively. The ensembled model improved the overall F1 and kappa to 0.696 and 0.531, with the best-case performance reaching F1 = 0.817 and κ = 0.780. Conclusion Integrating multiple physiological signal sources can complement each other, enhancing sleep stage detection accuracy and enabling more precise sleep monitoring. Support (if any) None.

Ask AI

Helpful

Bookmark

View Full Paper