Background/Purpose: Remote therapeutic monitoring (RTM) and remote physiologic monitoring (RPM) programs have the potential to capture data between clinical visits and provide information back to providers about the health status of their patients. We evaluated the ability of patient reported outcome (PRO) data coupled with passive biosensor data streams to classify RA low disease activity (LDA).

Methods: We recruited 150 RA patients to a prospective study that captured PRO data and among a subgroup (n=96), wear a Fitbit Versa2 device for approximately 3-4 months at the time that they initiated adalimumab or upadacitinib. Clinical Disease Activity Index (CDAI) was collected at baseline and follow-up. A variety of PROs and passive biosensor activity data were captured at various frequencies.Various machine learning methods (Distributed RandomForests [DRF], Gradiant Boosting, Generalized Linear Models, Ensemble) were used to classify LDA using only the baseline clinical data plus longitudinal PROs and biosensor data. Models were trained on a random split sample (training/testing) in a 2/3:1/3 ratio. Model building prioritized parsimonious models that minimized the number of PROs that patients were asked to provide (allowing for 3 PROs, at most), de-prioritizing more frequently collected PROs (e.g. daily) and favoring less frequently collected PROs (e.g. weekly, monthly). A scenario that simulated PRO data collection with at half the frequency was modeled to evaluate ML model performance under these conditions, intended to reduce participant burden. The best models were selected based on the F1 statistic (harmonic mean of sensitivity and positive predictive value), and separately, the F1 statistic, requiring a PPV of ≥80%. Confidence intervals were estimated using 1000 bootstrap samples.

Results: Characteristics of the RA cohort (Table 1) were: mean (SD) age 52.1(12.0) years, 79.3% female, mean (SD) BMI 31.5 (7.7), mean (SD) baseline CDAI 29.9 (16.4), 64.7% seropositive. Concomitant therapies included NSAIDs (40.7%), glucocorticoids (32.7%), opioids (24.7%) opioids. Prior treatment history: 36.7% were biologic naïve, and 27.3% used exactly 1 biologic. Mean (SD) step count was 6299 (3815) steps and mean (SD) sleep was 6.5 (2.9) hours of sleep.The most important longitudinal PRO variables included weekly PROMIS Fatigue, PROMIS Pain Interference, and RADAI5. Model accuracy (predicted vs. observed LDA) is shown (Figure 1) and indicated that model calibration was reasonable. Model accuracy exceeded 80% with 1, 2, or 3 PROs, with or without Fitbit biosensor data (Figure 2). Using only half-frequency PRO data capture (e.g. every 2 weeks) did not substantially impair model accuracy. Performance was modestly better if daily PROs plus Fitbit data were used (not shown), albeit at the expense of greater participant burden.

Conclusion: Machine learning can accurately identify RA patients in LDA or remission measured 3-4 months after starting adalimumab or upadacitinib using as few as 1 longitudinal PRO measures collected as infrequently as every 2 weeks. Only small increases in performance were observed when actigraphy biosensor data was added.

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/artificial-intelligence-applied-to-patient-reported-outcomes-and-passive-physiologic-sensor-data-can-accurately-classify-low-disease-activity-in-rheumatoid-arthritis-patients/