Multimodal Sleep-Wake Detection System: A Machine Learning Pipeline for Physiological Signals from Wearable Devices

This project addresses the pain points of traditional sleep monitoring (e.g., PSG) such as high cost and inconvenience for daily use, and designs a multimodal sleep-wake detection pipeline specifically for leg-worn wearable devices. By integrating data from LED/PPG, accelerometer, gyroscope, and temperature sensors, it uses machine learning models to automatically classify sleep and wake states, providing a practical solution for health monitoring.

Sleep quality monitoring is crucial for health management, but traditional PSG devices rely on laboratory environments, are costly, and inconvenient for daily use. With the popularization of wearable devices, sleep monitoring based on wearing positions like the leg has become a research hotspot. This project aims to provide a complete multimodal detection pipeline to address the limitations of single sensors.

Sensor Types and Features

• LED/PPG: Extract features such as heart rate, heart rate variability, signal amplitude, etc.
• Motion Sensors: Calculate SMA, direction-independent metrics, activity burst detection, etc.
• Temperature Sensor: Analyze baseline temperature, rate of change, variability, etc.

Preprocessing and Feature Engineering

• Sensor-specific preprocessing (filtering, peak detection, direction normalization, etc.)
• Extract statistical time-domain (mean, standard deviation, etc.), frequency-domain (FFT, PSD, etc.), and PPG-specific features

Model Selection

Supports models like logistic regression, random forest, XGBoost, etc., for sleep-wake classification.

Evaluation Metrics

Use accuracy, F1 score, ROC-AUC, and confusion matrix for model evaluation

Feature Importance

Identify key sensors and features through gain analysis of random forest and XGBoost

Visualization

• Sensor feature scatter plots: Distinguish sleep (blue) and wake (red) states
• Predictive probability time-series plots: Compare the consistency between model predictions and actual labels

For example, the RMS AUC feature of the accelerometer's X-axis shows high values during wakefulness and is close to zero during sleep.

This project demonstrates the potential of multimodal sensor fusion in sleep monitoring, improving detection accuracy and robustness by integrating multiple signal sources. It provides a complete reference implementation for developers, assists researchers in advancing sleep algorithm development, and promotes the practical application of wearable health monitoring.

The project organizes the workflow using Jupyter Notebooks:

• 01_preprocess.ipynb: Data loading, cleaning, and windowing
• 02_model.ipynb: Model training and evaluation

Utility functions are encapsulated in the utils/ package, including filters, feature extraction modules, etc., supporting modular adaptation to different datasets and device specifications.