Personalized Estimation of Fatigue in Endometriosis Using Wrist-Worn Actigraphy Data and Hierarchical Bayesian Modeling

Endometriosis is a common chronic condition associated with debilitating pain and fatigue. Wearable technology has the potential to improve symptom monitoring in endometriosis where symptom severity has been traditionally evaluated through retrospectively-completed questionnaires. We analyzed longitudinal actigraphy data from 56 endometriosis patients across a maximum of three 4-6-week "cycles" (74.0 +/- 27.0 days per patient), of whom 46 had sufficient data for validation (>20 days), to predict fatigue scores (Brief Fatigue Inventory, ranging from 0-10) using hierarchical Bayesian regression. We extracted 193 daily actigraphy features, from which 10 features were chosen through feature selection for use in the final model, and we assessed model performance using a leave-one-participant-out cross-validation approach, in which each persons own baseline data were included in model training. We compared the Bayesian model to classical linear regression models and naïve benchmarks using both an individualized (including baseline data) and a generalized (excluding baseline data) approach. The Bayesian model resulted in the lowest average participant mean absolute error of 1.42, compared to 2.03 using the generalized linear regression model, and was best at capturing variability in fatigue (mean Pearson correlation of 0.33). Therefore, Bayesian modeling based on continuous actigraphy data shows promise in monitoring fatigue in endometriosis, offering estimation that adapts to patients' own data, while also utilizing data from the broader population.Clinical relevance - Passively collected wrist-worn actigraphy data enable personalized continuous monitoring of fatigue in endometriosis, which has clear implications towards facilitating longitudinal, remote, and objective symptom monitoring and rehabilitation assessments.