Wearable multi-wavelength photoplethysmography deep learning heart rate estimation

Ray, Daniel (2024) Wearable multi-wavelength photoplethysmography deep learning heart rate estimation. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (8MB) | Preview

Abstract

Wrist-worn photoplethysmography (PPG) has become a popular method for continuous and remote heart rate monitoring, but single-wavelength PPG faces limitations in accuracy, robustness, and generalisability. This study explores multi-wavelength PPG sensing to enhance heart rate estimation accuracy, robustness, and fairness across diverse populations, particularly for healthcare applications. A novel dataset comprising 26,442 samples from 20 participants with diverse skin types (Fitzpatrick I-VI) and varying heart rates and motion types was introduced, including blue, green, red, and infrared PPG wavelengths. Additionally, an uncertaintyaware deep learning method was developed for wrist-worn PPG heart rate estimation, optimised for single- and multi-wavelength PPG, using sensor fusion and LOSO crossvalidation. The pilot study analysed the impact of skin melanin, biological sex, and wavelength on PPG heart rate estimation. The blue-green-red-IR combination proved most effective. Significant differences in error distributions across wavelengths were observed for skin melanin and biological sex. High melanin content was associated with higher MAE (8.4 ± 2.1 BPM) compared to low melanin (6.1 ± 2.2 BPM). An uncertainty-aware postprocessing method demonstrated competitive performance, mitigating the effects of skin melanin content by equalising the MAE to 3.3 ± 0.9 BPM for high melanin and 3.3 ± 1.3 BPM for low melanin. The method recorded lowest MAE values on three existing single-wavelength datasets—1.3 ± 0.6 BPM on IEEE Train, 1.2 ± 0.4 BPM on BAMI 2, and 2.5 ± 0.9 BPM on PPG DaLiA-compared to existing deep learning methods. For the newly collected multi-wavelength dataset, the method achieved a MAE of 3.3 ± 1.1 BPM. The pilot study improved reliability through selective rejection of uncertain samples, despite lower retention rates. By investigating multi-wavelength PPG and introducing reliability indicators, this research aims to enhance accuracy and reliability of wristworn PPG heart rate monitoring across diverse populations, addressing disparities and improving healthcare applicability. These findings lay groundwork for further research advancing more inclusive and reliable wrist-worn PPG heart rate estimation methods.