Evaluating the performance and acceptability of an artificial intelligence digital health exercise platform in females with and without axial spondyloarthritis: a phase I study.

backgroundAxial spondyloarthritis (axSpA) affects equal numbers of females and males, yet females report lower exercise participation and greater functional limitations. Scalable, home-based exercise interventions may help address these differences. The ExerciseRx™ app is a personalized digital health platform that delivers and monitors exercises. This Phase 1 exploratory study evaluated the performance and acceptability in preparation for a future randomized control trial in axSpA.

methodsBaseline demographic and clinical data were collected. Participants completed a one-time supervised in-lab session using the ExerciseRx app to perform recommended exercises, recorded via the smartphone camera. We first evaluated the performance of the machine learning (ML) models built into the ExerciseRx app on two tasks: (1) exercise identification, reported with accuracy (i.e., number of exercises correctly identified); and (2) repetition counting, reported with mean absolute error (i.e., number of repetitions correctly identified). We then investigated the relationship between movement patterns and clinical measures. Finally, we examined acceptability using the Mobile Health App Usability Questionnaire (MAUQ).

resultsWe enrolled 21 females (11 axSpA, 10 controls), mean age 48.1 ± 14.3 (axSpA) and 47.0 ± 13.3 (controls). Most were non-Hispanic White (85.7%). Mean Bath Ankylosing Spondylitis Disease Activity Index (BASDAI) was 4.03 ± 1.68 for axSpA and 1.76 ± 1.30 for controls; mean Bath Ankylosing Spondylitis Functional Index (BASFI) was 2.75 ± 1.84 for axSpA and 0.88 ± 1.31 for controls. Overall mean ML model accuracy in exercise identification was 81.7%, lower in axSpA (72.3%) than controls (92.3%). Higher BASDAI and BASFI scores were not associated with lower accuracy, whereas any abnormal metrology value was. MAUQ scores were high on a Likert scale out of 7 (6.35 ± 1.23 axSpA; 6.45 ± 1.27 controls). Three participants (14%) reported mild, acceptable soreness the following day; no adverse events occurred.

conclusionsThe ExerciseRx ML models showed adequate performance, and the app was accepted by participants, though findings may not be broadly generalizable. Lower accuracy in axSpA highlights the need for ML model refinement prior to larger studies. High acceptability and absence of adverse events suggest that the ExerciseRx app has strong potential to be an effective home-based exercise tool for these patients.