Abstract 4370378: Camera is All You Need: Low-cost Atrial Fibrillation Detection using Facial Remote Photoplethysmography

Abstract

Introduction: Atrial fibrillation (AF) is one of the most prevalent cardiac arrhythmias, yet current screening and monitoring methods require contact-based electrocardiographic (ECG) equipment, leading to higher costs and potential diagnostic delays. Facial remote photoplethysmography (rPPG) uses video to capture subtle light absorption changes in the skin and is a promising avenue for accessible AF monitoring. While algorithms exist for detecting AF from R-R intervals extracted from ECG, the inherent noisiness of rPPG signals makes direct application difficult. Deep learning models have been proposed; however, they require high computational costs and offer limited explainability. Aims: This study aimed to develop and validate a low-cost, explainable approach to detect AF from beat-to-beat (B-B) intervals extracted from facial rPPG. Our goal was to avoid complex model architectures and to instead train a simple model using only R-R intervals extracted from ECG data. Methods: We first trained a model using 1172 single-lead recordings from the 2017 CinC Challenge dataset, of which 567 were labeled as AF. From these recordings, R-R intervals were extracted and then used for computing seven handcrafted features. We then trained a support vector machine (SVM) model solely on this dataset. For external validation, we conducted a prospective clinical study in which facial video (320×240 pixels, 150 Hz) and 12-lead ECG were recorded from 15 AF and 52 normal sinus rhythm (NSR) patients. ECGs were annotated by board-certified physicians. Videos were segmented into 30-second clips. The green channel was isolated, bandpass filtered and then temporally smoothed. Signals from the forehead and cheek regions were averaged to produce a representative rPPG signal. A heuristic algorithm was used to assess signal quality. Beat onsets were detected using an adaptive thresholding algorithm and then used to compute the B-B intervals for classification. Results: The model has achieved a sensitivity of 100.00%, a specificity of 97.73% and an area under the curve (AUC) of 0.9773. The indeterminate rate due to low signal quality was 11.94%. Conclusion: We haved successfully demonstrated that an interpretable, low-cost machine learning model trained solely on ECG data can accurately detect AF from facial rPPG. This shows significant promise for developing accessible, low-cost, non-contact rPPG-based systems for AF screening and monitoring, with possible future applications to other arrhythmias.