Multitask learning approach for PPG applications: Case studies on signal quality assessment and physiological...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Wearable photoplethysmography (PPG) enables remote health monitoring of physiological parameters like heart rate (HR) and respiration rate (RR).
• Current PPG analysis often treats physiological parameters independently, neglecting their inherent interdependencies.
• Multitask learning (MTL) is a machine learning paradigm that can leverage these interdependencies by training a single model for multiple related tasks.

Purpose of the Study:

• To investigate the potential of MTL for enhancing PPG-based remote health monitoring applications.
• To develop and evaluate novel multitask deep learning models for PPG quality assessment and simultaneous HR/RR estimation.
• To demonstrate the superiority of MTL approaches over traditional single-task learning methods in PPG analysis.

Main Methods:

• Development of customized multitask deep learning architectures tailored for PPG data.
• Application of MTL for two distinct tasks: (1) PPG signal quality assessment for HR/HRV and (2) simultaneous estimation of HR and RR.
• Evaluation of proposed models using a PPG dataset collected from 46 subjects during daily activities using smartwatches.

Main Results:

• The proposed MTL models significantly outperformed baseline single-task models across all evaluated tasks.
• MTL approaches achieved higher accuracy in PPG quality assessment compared to single-task methods.
• Reduced error rates were observed for simultaneous HR and RR estimation using MTL.

Conclusions:

• Multitask learning effectively exploits shared characteristics among PPG-derived physiological parameters.
• MTL offers a robust framework for improving the performance and accuracy of wearable PPG-based health monitoring.
• The developed MTL models demonstrate significant potential for real-world applications in remote patient monitoring.