Performance evaluation of a machine learning-based methodology using dynamical features to detect nonwear intervals in actigraphy data in a free-living setting

Goal and aims: One challenge using wearable sensors is nonwear time. Without a nonwear (e.g., capacitive) sensor, actigraphy data quality can be biased by subjective determinations confounding sleep/wake classification. We developed and evaluated a machine learning algorithm supplemented by dynamic features to discern wear/nonwear episodes. Focus technology: Actigraphy data from wrist actigraph (Spectrum, Philips-Respironics). Reference technology: The built-in nonwear sensor as “ground truth” to classify nonwear periods using other data, mimicking features of Actiwatch 2. Sample: Data were collected over 1 week from employed adults (n = 853). Design: Extreme gradient boosting (XGBoost), a tree-based classifier algorithm, was used to classify wear/nonwear, supplemented by dynamic features calculated over various time windows. Core analytics: The performance of the proposed algorithm was tested over 30-second epochs. Additional analytics and exploratory analyses: Evaluation of the SHapley Additive exPlanations (SHAP) values to find the effectiveness of the dynamic features. Core outcomes: The XGBoost classifier yielded substantial improvements in balanced accuracy, sensitivity, and specificity, including dynamic features and comparison to default actiwatch classification algorithms. Important supplemental outcomes: The proposed classifier effectively distinguished between valid and invalid days, and the duration of contiguous periods of nonwear correctly identified. Core conclusion: Our findings highlight the potential of XGBoost using dynamic features of varying activity levels across the time series to provide insights on wear/nonwear classification using a large dataset. The methodology provides an alternative to laborious manual benchmarking of the data for similar devices that do not have a nonwear sensor.