Performance evaluation of a machine learning-based methodology using dynamical features to detect nonwear intervals...

Area of Science:

• Wearable sensor technology
• Machine learning in healthcare
• Biomedical data analysis

Background:

• Wearable sensors are crucial for health monitoring, but nonwear time can bias data quality.
• Subjective methods for determining nonwear time can confound sleep/wake classification in actigraphy.
• Developing objective methods to identify nonwear episodes is essential for reliable data.

Purpose of the Study:

• To develop and evaluate a machine learning algorithm for discerning wear/nonwear episodes in actigraphy data.
• To improve the accuracy of actigraphy data by objectively identifying nonwear periods.
• To provide an alternative to manual data benchmarking for devices without dedicated nonwear sensors.

Main Methods:

• Utilized Extreme Gradient Boosting (XGBoost), a tree-based classifier, to classify wear/nonwear episodes.
• Supplemented XGBoost with dynamic features calculated over various time windows.
• Collected data from 853 employed adults over one week using wrist actigraphy.

Main Results:

• The XGBoost classifier significantly improved balanced accuracy, sensitivity, and specificity compared to default algorithms.
• Dynamic features were found to be effective in wear/nonwear classification, as indicated by SHapley Additive exPlanations (SHAP) values.
• The algorithm accurately distinguished between valid and invalid days and identified the duration of nonwear periods.

Conclusions:

• XGBoost, incorporating dynamic features of activity levels, offers a robust approach to wear/nonwear classification in actigraphy.
• This methodology enhances data quality for wearable sensors, particularly those lacking built-in nonwear detection.
• The developed algorithm provides a valuable tool for analyzing large datasets from actigraphy devices.