Fatigue Detection with Machine Learning Approaches using Data from Wearable Devices

Severe and chronic fatigue is one of the top symptoms in patients with non-communicable chronic immune-mediated inflammatory diseases like Systemic Lupus Erythematosus (SLE) and Sjögren's disease (SjD). The majority of fatigue assessment approaches rely on self-reported questionnaires. While these subjective measures are important for capturing how patients feel, function and survive; they could potentially be complemented by objective, passive measures, which can provide further insights on fatigue and its impact on daily life. In this paper, we aim to objectively estimate fatigue using accelerometer-derived measures and a machine learning-based framework in individuals with SLE and SjD and compared to demographically matched healthy volunteers (HNV). Accelerometer data were collected from 96 study participants in a free-living environment over 24 weeks using the ActiGraph Centrepoint Insight Watch. The raw data were processed to extract several activity-based, sleep-based, and circadian rhythm-based features. The extracted features were used to train a classifier to determine an individual's daily fatigue status (fatigued or not fatigued). Our methods, evaluated using stratified k-fold and leave-one-participant-out cross-validation, effectively distinguished fatigue status with better than random performance and performed similarly with accelerometer features alone as with baseline participant characteristics alone. Model performances varied across cohorts: ROC-AUC ranges from 0.44 to 0.70 in the individual cohorts, improving to 0.76 to 0.83 when cohorts were combined. The results presented in this paper demonstrate the potential of wearable devices for developing digital biomarkers of fatigue, which could be used to assess response to therapy across therapeutic areas.