Digital health strategies for endometriosis symptom monitoring using wrist-worn wearables and statistical machine learning

Endometriosis is a chronic condition affecting an estimated 10% of reproductive-aged women. Debilitating symptoms including pain and fatigue have a major impact on wellbeing and quality of life. Endometriosis, which is defined as the presence of endometrial-like tissue found outside of the uterus, is generally diagnosed at surgery, but as surgical presentation often does not align with symptoms, there are few objective outcome measures that can be used to monitor disease state. Advances in digital technologies including smartphone- and wearable-based assessments have enabled the longitudinal tracking of symptoms and behaviours non-invasively and remotely, yet they have seldom been used to track symptoms in endometriosis. In this thesis, my aim was to explore the feasibility and utility of wrist-worn actigraphy for the monitoring of endometriosis symptoms by comparing objective tracking data with that recorded by patients using standardised questionnaires (patient-reported outcome measures: PROMs). Together with the clinical research team, I conducted the first world-wide longitudinal observational study of endometriosis patients deploying both PROMs and wearable technology. In this longitudinal study, 68 patients with surgically confirmed endometriosis were recruited to (a) wear smartwatches, which collected three-dimensional acceleration (actigraphy), ambient light, and temperature, and (b) submit daily and retrospective PROMs (recalling the previous 4 weeks) over a maximum of three 4-6-week smartwatch cycles. The PROMs used were the Brief Fatigue Inventory (BFI, range 0-10) and pain scores (average and worst pain, range 1-10). Throughout the study, I assessed feedback through a patient group and adjusted the study design to adapt to this patient input to carry forward to the rest of the study. Using the data collected from the longitudinal study, I developed a reusable data pipeline to process the raw actigraphy data, clean the PROMs into useable outputs, and visualise these outputs in a new user-friendly dashboard tool. To detect non-wear periods in the dataset, I improved upon previous non-wear detection algorithms, outperforming multiple other methods in the literature. From the actigraphy data, I extracted 201 actigraphy measures relating to physical activity, sleep, and diurnal rhythms based on multiple approaches established in the literature, as well as newly developed methods. To assess the utility of the actigraphy data for providing new insights into endometriosis symptom trajectories, I used visualisation, correlation analysis, and linear mixed-effects modelling to assess associations between actigraphy measures and PROMs, examining both within- and between-person associations and considering factors such as medical history, current medication, menstruation, and demographics. Correlation analysis revealed that daily physical activity was strongly negatively correlated with self-reported fatigue (repeated-measures correlations R<-0.3). Linear mixed-effects models indicated weak (standardised effect sizes ranging from |0.02-0.13|) but nonetheless statistically significant (p<0.05) associations between actigraphy measures and fatigue when controlling for pain. However, there were fewer associations between actigraphy measures and pain when controlling for fatigue. Generally, participants with more severe or variable pain and fatigue symptoms displayed lower physical activity, increased sleep disturbance, and more disrupted sleep and activity rhythms (Spearman’s |R|>0.3). In a pilot study, comprising a subset of 13 patients with deep endometriosis who received surgery and completed follow-up smartwatch cycles immediately after surgery and 4-6 months later, I found evidence of actigraphy changes immediately following surgery that coincided with changes in PROMs, including decreased physical activity and increased sleep disturbance. At the 4-6-month follow-up, patients’ self-reported symptoms on pain and fatigue had improved, while only few actigraphy-derived measures such as sleep regularity similarly improved. Collectively, this small-scale study demonstrates the potential for the use of wearable technology as an objective approach to assess surgical rehabilitation outcomes short- and long-term. Finally, to examine the utility of actigraphy for monitoring endometriosis symptoms, I compared the use of a personalised statistical learning setup, where a patient could provide several weeks of training data, to a standard generalised setup, where no individual data was available for training models. I found that hierarchical Bayesian regression models, where the first smartwatch cycle from each participant was used to train individualised models, performed best in estimating daily pain and fatigue. When evaluated using a leave-one-participant-out validation method, the hierarchical Bayesian model estimating daily self-reported fatigue resulted in a mean participant Mean Absolute Error of 1.42 (scale from 0-10) compared to 2.03 using classical linear regression with no individualised training. I found that estimating pain using actigraphy data resulted in lower performance relative to fatigue (considering the different rating scales), but again Bayesian models performed best (MAE of 1.51 compared to 1.85 using classical linear regression). The work presented in my thesis lays a foundation for understanding and monitoring endometriosis symptoms using wearable technology and was the first longitudinal study collecting raw actigraphy data alongside daily PROMs in endometriosis. Using a combination of statistical methods to gain insight into associations as well as statistical learning models to inform monitoring potential, I have demonstrated that actigraphy can be used to monitor fatigue symptoms more robustly than pain, and additionally can help shed light on the relationships between physical activity, sleep, diurnal rhythms, and endometriosis symptoms, which could potentially guide symptom self-management approaches and be used for non-invasive monitoring of interventions.