Improved Photoplethysmography-Based Four-Stage Sleep Classification with Explainable AI-Driven Machine Learning

Classifies four sleep stages from low-cost PPG signals using explainable ML, making interpretable sleep monitoring viable on wearable-grade sensors.

Single-sensor-based multistage sleep classification is a challenging task to match the accuracy of traditional polysomnography (PSG). Feature engineering is the key factor in improving accuracy for any single-sensor, machine-learning-based measurement — but too many features increase model complexity and training time. To address this, explainable Artificial Intelligence (XAI) is used across several machine learning classifiers to determine the most explainable features of single-sensor PPG data for a lightweight model. This study demonstrated substantial improvements in prediction accuracy — 75.54% for 4-stage sleep classification using 42 PPG data variables. With a simpler AI model and a single low-cost PPG sensor, this adaptive approach shows promise for accurately staging sleep, which may simplify the detection and treatment of sleep disorders and potentially enable at-home solutions.

Ferdous, T., Karim, R. U., et al. (2024). Improved Photoplethysmography-Based Four-Stage Sleep Classification with Explainable AI-Driven Machine Learning. IEEE ICEACE 2024. https://doi.org/10.1109/ICEACE63551.2024.10898853