A Multi-Level Classification Approach for Sleep Stage Prediction With Processed Data Derived From Consumer Wearable...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Sleep Science

Background:

• Consumer wearable activity trackers are widely used for sleep monitoring.
• Existing devices exhibit inaccuracies in sleep stage classification.
• Accurate sleep stage data is crucial for health assessment and research.

Purpose of the Study:

• To develop and validate a novel approach for predicting sleep stages using processed data from consumer activity trackers.
• To improve the accuracy of sleep stage classification compared to proprietary algorithms.
• To explore the efficacy of machine learning models in enhancing wearable sleep tracking.

Main Methods:

• A two-level classifier system was developed, leveraging steps, heart rate, and sleep metrics.
• Level-I classifier identifies potentially misclassified sleep epochs.
• Level-II classifier re-classifies epochs into light, deep, REM sleep, or wakefulness stages using Support Vector Machine (SVM) and XGBoost.

Main Results:

• The best model (SVM and XGBoost with up-sampling) achieved an epoch-wise accuracy of 0.731.
• This model demonstrated a 23.9% reduction in Mean Absolute Bias (MAB) for sleep stage duration compared to Fitbit's algorithm.
• A sub-optimal model (SVM and XGBoost with down-sampling) achieved a 71.0% reduction in MAB.

Conclusions:

• The proposed machine learning approach significantly enhances the accuracy of sleep stage prediction from consumer wearable data.
• The findings highlight the potential of leveraging readily available wearable data for more reliable sleep analysis.
• Further research is needed to address challenges in machine learning-based sleep stage prediction with wearables.