Improving Photoplethysmography-Based Multiclass Cardiac Arrhythmia Detection via Pulse Wave Analysis

No Thumbnail Available
Jeanningros, Loïc
Braun, Fabian
Van Zaen, Jérôme
Le Bloa, Mathieu
Porretta, Alessandra
Herrera, Claudia
Teres, Cheryl
Domenichini, Giulia
Carroz, Patrice
Graf, Denis
Methods PPG signals were acquired from the wrist simultaneously with 12-lead ECG (used for gold-standard annotation of CAs) from 42 patients referred for catheter ablation at the Lausanne University Hospital. PPG segments of 30 s were automatically classified as either SR, AF or non-AF based on spectral and temporal features extracted from raw PPG time series and from inter-pulse interval series. In addition, novel PWA features extracted by detecting specific points in the PPG waveform and its derivatives provided insights into the morphology of individual pulses. Their discriminative power was evaluated based on the Relief algorithm for feature selection. Finally, we compared performance metrics for CA classification with and without PWA features. Results The classification accuracy using ridge regression was significantly increased by 2.8%, from 73.5% to 76.3% (p = 0.009), when using PWA features on top of temporal and spectral features. Likewise, the sensitivity in detecting AF increased by 3.4%, from 77.3% to 80.7% (p = 0.09). The most discriminative PWA features were the acceleration magnitude of the pulse upstroke and the timing of the dicrotic notch. Conclusion PWA features showed some potential for the detection and classification of AF and non-AF CAs. Although further work and extensive data is required to classify each type of CA individually, these results show the potential for improving our understanding of the peripheral hemodynamic signature of CAs.
Publication Reference
BMT 2022, Innsbruck (Austria), pp. 226