Raw photoplethysmogram waveforms versus peak-to-peak intervals for machine learning detection of atrial fibrillation: Does waveform matter?

Abstract

Background: Machine learning-based analysis can accurately detect atrial fibrillation (AF) from photoplethysmograms (PPGs), however the computational requirements for analyzing raw PPG waveforms can be significant. The analysis of PPG-derived peak-to-peak intervals may offer a more feasible solution for smartphone deployment, provided the diagnostic utility is comparable.

Aims: To compare raw PPG waveforms and PPG-derived peak-to-peak intervals as input signals for machine learning detection of AF.

Methods: We developed specialized neural networks for raw waveform and peak-to-peak interval analyses and trained them on 7,704 PPGs from 106 patients from the TeleCheck-AF project. We evaluated the neural networks on 48,912 PPGs from 416 patients from the VIRTUAL-SAFARI project. We recorded computational requirements, sensitivity, positive predictive value (PPV), and F1 score.

Results: With 1.6 million trainable parameters, the waveform model was more than 100 times as complex as the interval model (15,513 parameters) and required 19 times more computational power. In external validation, metrics were comparable between the interval and waveform models. For the interval model vs. the waveform model, sensitivity was 91.7 % vs. 81.9 % (p=0.4), PPV was 80.5 % vs. 84.5 % (p=0.3), and F1 score was 85.6 % vs. 81.3 % (p=0.5), respectively.

Conclusion: PPG-derived peak-to-peak intervals and PPG waveforms were equivalent as input signals to neural networks in terms of accurate AF detection. The reduced computational requirements of the interval model make it a more suitable option for deployment on digital end-user devices such as smartphones.