Lightweight ResNet-Based Deep Learning for Photoplethysmography Signal Quality Assessment

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Deep learning models are increasingly used in wearable devices, but computational constraints require lightweight and efficient designs.
• Photoplethysmography (PPG) signal quality assessment (SQA) is vital for reliable cardiovascular monitoring using wearables.
• Preprocessing methods can enhance the performance of deep learning models for PPG analysis.

Purpose of the Study:

• To develop a lightweight deep learning framework for PPG signal quality assessment (SQA).
• To evaluate the impact of different input configurations (PPG, derivatives, autocorrelation) on SQA performance.
• To compare the proposed model's efficiency and effectiveness against existing studies.

Main Methods:

• A ResNet-based deep learning framework incorporating Squeeze-and-Excitation (SE) modules was proposed.
• The model was trained and tested using the Moore4Medical (M4M) and MIMIC-IV datasets.
• Various input channel combinations, including PPG signal, first derivative (FDP), second derivative (SDP), and autocorrelation (ATC), were explored.

Main Results:

• The model achieved high performance, with up to 96.52% AUC on the M4M dataset and 84.43% AUC on the MIMIC-IV dataset.
• Significant reductions in model parameters (over 99%) and FLOPs (over 60%) were achieved compared to existing methods.
• The M4M dataset, novel in its focus on PPG for atrial fibrillation (AF) detection, was utilized.

Conclusions:

• The proposed lightweight framework offers efficient and accurate PPG SQA for resource-limited wearable devices.
• This technology can enhance the reliability of continuous cardiovascular monitoring, supporting clinical decisions in telemedicine and remote care.
• The model's reduced computational footprint facilitates practical deployment and broader adoption in wearable health technology.