Lightweight ResNet-Based Deep Learning for Photoplethysmography Signal Quality Assessment

TL;DR

This paper introduces a lightweight ResNet-based deep learning model with SE modules for PPG signal quality assessment, achieving high accuracy while significantly reducing computational complexity on two datasets.

Contribution

It presents a novel, efficient ResNet-based framework with SE modules for PPG SQA, optimized with various input configurations, and demonstrates superior performance and efficiency over existing methods.

Findings

Achieved up to 96.52% AUC on M4M dataset

Reduced model parameters by over 99% compared to existing studies

Lowered FLOPs by more than 60% while maintaining high accuracy

Abstract

With the growing application of deep learning in wearable devices, lightweight and efficient models are critical to address the computational constraints in resource-limited platforms. The performance of these approaches can be potentially improved by using various preprocessing methods. This study proposes a lightweight ResNet-based deep learning framework with Squeeze-and-Excitation (SE) modules for photoplethysmography (PPG) signal quality assessment (SQA) and compares different input configurations, including the PPG signal alone, its first derivative (FDP), its second derivative (SDP), the autocorrelation of PPG (ATC), and various combinations of these channels. Experimental evaluations on the Moore4Medical (M4M) and MIMIC-IV datasets demonstrate the model's performance, achieving up to 96.52% AUC on the M4M test dataset and up to 84.43% AUC on the MIMIC-IV dataset. The novel M4M dataset was collected to explore PPG-based monitoring for detecting atrial fibrillation (AF) and AF burden in high-risk patients. Compared to the five reproduced existing studies, our models achieves over 99% reduction in parameters and more than 60% reduction in floating-point operations (FLOPs).