A deep learning model using photoplethysmography signals detected atrial fibrillation with a ROC-AUC of 0.968, sensitivity of 97.4%, and specificity of 91.3% compared to gold-standard ECG.
Observational (n=10,767)
Does a deep learning model using PPG signals accurately detect atrial fibrillation compared to ECG in patients undergoing sleep studies?
A deep learning model applied to PPG signals from sleep studies demonstrated high accuracy for detecting atrial fibrillation, providing a scalable method for arrhythmia screening without additional hardware.
Estimación del efecto: ROC-AUC 0.968, Sensitivity 97.4%, Specificity 91.3%
Abstract Introduction Atrial fibrillation (AFib) is a prevalent and often underdiagnosed comorbidity in patients with obstructive sleep apnea (OSA). Photoplethysmography (PPG) is widely available in PSG and home sleep apnea tests (HSAT), yet detection of AFib and atrial flutter (AFL) typically relies on electrocardiography (ECG). We developed and evaluated a deep learning model to detect AF events (AFib and AFL combined) from PPG signals, offering a scalable solution to implement high-accuracy, routine detection within large clinical sleep cohorts. Methods A total of N=10,767 studies with simultaneous ECG and PPG were randomly sampled from a large clinical PSG/HSAT database. Sampling was stratified uniformly across AFib burden to construct balanced training (N=7,752), validation (N=1,938), and test (N=1,077) datasets. AFib burden was estimated using a validated open-source ECG model and defined as the percentage of segments classified as AFib. Definitive AF labels were derived from a validated ECG analysis algorithm applied to a single-lead ECG II. A deep learning model was trained exclusively on PPG signals, and performance was evaluated on a per-epoch (30-second) basis on the test set. A total of 67,023 (7.5%) of the 30-second epochs in the test sample group were excluded due to insufficient PPG signal quality or artifacts. Results The PPG-based deep learning model achieved strong per-epoch AF (AFib/AFL) classification performance on the test dataset of 1,077 patients, which comprised 656,651 AF epochs and 173,579 non-AF epochs. The model achieved a ROC-AUC of 0.968, sensitivity of 97.4%, and specificity of 91.3%. The corresponding positive predictive value (PPV) was 97.7%, and negative predictive value (NPV) was 90.3%. Conclusion The deep learning PPG model demonstrated robust detection of AF, achieving performance comparable to the gold-standard of ECG for AFib diagnosis. These findings highlight the potential to integrate automated AF detection into existing HSAT and PSG workflows, enabling scalable, high-accuracy arrhythmia screening in patients undergoing sleep studies without the need for additional hardware. Support (if any)
Wodnicki et al. (Fri,) conducted a observational in Atrial fibrillation and obstructive sleep apnea (n=10,767). Deep learning model using photoplethysmography (PPG) vs. Single-lead ECG II (gold standard) was evaluated on Per-epoch (30-second) atrial fibrillation and flutter classification performance (ROC-AUC 0.968, Sensitivity 97.4%, Specificity 91.3%). A deep learning model using photoplethysmography signals detected atrial fibrillation with a ROC-AUC of 0.968, sensitivity of 97.4%, and specificity of 91.3% compared to gold-standard ECG.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: