(042) Use of Machine Learning to Evaluate Incremental Value of Actigraphy Data for Classifying Treatment Response in Patients with Rheumatoid Arthritis

Background: Digital health technology to collect electronic patient reported outcomes (ePRO) and biosensor data are increasingly used to generate real-world data in pharmacoepidemiology. However, the utility of passive biosensor data from actigraphy devices incremental to collecting ePRO data to classify treatment response in rheumatoid arthritis (RA) or other musculoskeletal conditions is not clear.

Objectives: To determine the incremental value of actigraphy data in classifying treatment response among RA patients.

Methods: We conducted a prospective observational study at 28 US community rheumatology clinics among RA patients initiating upadacitinib or adalimumab. Patients contributed multiple ePROs at home on a smartphone app (PatientSpot, formerly ArthritisPower) on a daily, weekly and monthly basis (mean duration 2 min/day) from baseline to a follow-up visit occurring about 3-4 months later. We incorporated ePRO data, with and without actigraphy data, from a study-provided device (Fitbit Versa3) into machine learning (ML) models to classify whether patients achieved low disease activity (LDA) or remission (REM; Clinical Disease Activity Index < 10) as assessed by a rheumatologist at the follow-up visit. Feature engineering was performed on both ePRO and actigraphy raw data to derive within-person slope and variance of longitudinal change. A 2/3:1/3 random split sample approach was used to create separate training and testing samples; precision (positive predictive value) and recall (sensitivity) from the testing dataset were used to compare model accuracy.

Results: A total of 150 RA patients completed the study, and 96 contributed actigraphy data. Mean(SD) age was 52(12) years and 79% were women. A total of 62% (test sample) achieved LDA or REM. With only ePRO data, the best-performing ML generalized linear model achieved 86% precision and 75% recall to classify LDA and required 6 ePROs longitudinally to obtain adequate model performance. Ensemble ML methods attained higher recall (97%; 67% precision). The most important ePROs for classifying patients in LDA or REM were PROMIS fatigue, pain intensity, RADAI5, PROMIS physical function, PROMIS anxiety, and PROMIS pain interference. With the addition of passive actigraphy data, performance was similar or better (86% precision, 83% recall), and required fewer actively collected ePROs (pt global, RADAI5, PROMIS anxiety).

Conclusions: Longitudinal ePRO data collected at home accurately classified treatment response to new RA therapies. Passively contributed actigraphy data increased model performance slightly and reduced participant burden for data collection. Remote actigraphy data capture may be useful to develop a better understanding of real-world patient experience and treatment response.