Multi-task neural networks for the estimation of cardiorespiratory biomarkers from raw photoplethysmogram signals
No Thumbnail Available
URL
Journal Title
Journal ISSN
Volume Title
School of Science |
Master's thesis
Authors
Date
2025-01-16
Department
Major/Subject
Systems and Operations Research
Mcode
Degree programme
Master's Programme in Mathematics and Operations Research
Language
en
Pages
49
Series
Abstract
Photoplethysmography (PPG) is a non-invasive optical measurement technique commonly integrated into wearable devices. Critical cardiorespiratory biomarkers such as heart rate (HR), heart rate variability (HRV) and respiratory rate (RR) can be derived from PPG signals. As PPG signals are susceptible to noise and motion artifacts, conventional algorithms integrate complex signal pre-processing and error handling procedures to accurately estimate these biomarkers. End-to-end deep learning models, which automate feature extraction and learn robust representations of the PPG signal, have shown excellent performance for PPG-based HR estimation, although they are under-explored for HRV and RR. Multi-task learning (MTL) combines multiple tasks into a single neural network, thereby leveraging shared representations of the input signal. Compared to training separate models for each task, MTL can often reduce memory and runtime costs without compromising accuracy. However, in MTL, the key challenge lies in balancing task-specific gradients during training, as unbalanced contributions can lead to task interference or dominance. This thesis presents an end-to-end multi-task neural network for the estimation of HR, HRV and RR from raw PPG. The network architecture integrates a convolutional neural network (CNN) and a long-short term memory (LSTM) module, combined with task-specific multi-layer perceptron (MLP) heads. Single-task baseline models and multi-task models are trained, validated and tested on 30-second PPG segments with ground-truth labels from medical-grade reference devices, using data from a single-night sleep study with 908 participants. To address the challenge of balancing task-specific gradients, several MTL methods are implemented and compared, and the final multi-task model is chosen by best validation performance from a total of 36 training runs. It achieves state-of-the-art accuracy across the 230-participant test set, with coefficients of determination of 0.999 (HR), 0.983 (HRV), and 0.906 (RR) for overnight averages and 0.992 (HR), 0.905 (HRV), and 0.675 (RR) across 30-second segments. These results are on par with single-task baselines and improve upon the accuracy of conventional methods in comparable studies.Description
Supervisor
Oliveira, FabricioThesis advisor
Vallat, RaphaelKeywords
heart rate, heart rate variability, respiratory rate, photoplethysmography, deep learning, multi-task learning, gradient balancing