Abstract / Summary
Abstract Standard rheumatoid arthritis (RA) flare assessment relies on components that are either subjective (patient global assessment) or non-specific (ESR, CRP), motivating interest in objective, non-invasive adjuncts. This proof-of-concept study evaluated infrared thermal imaging combined with machine learning for detecting RA flare and classifying disease activity. Eleven seropositive RA patients and nine seronegative controls underwent standardized thermal imaging alongside clinical assessment across 2–3 visits. Whole-hand maximum temperature (Tmax) was compared between RA flare visits and controls. Per-joint thermal zone data were used to train logistic regression and random forest classifiers predicting disease activity, using leave-one-patient-out cross-validation to prevent data leakage. RA patients showed significant clinical improvement with treatment (DAS28-ESR 5.23 to 3.32, p = 0.0007). Whole-hand Tmax did not significantly distinguish flare from controls (p = 0.53). A leakage-free analysis (n = 26 visits) found no model exceeded baseline for 3-class severity classification. Reclassifying activity using a binary treat-to-target threshold yielded a non-significant trend toward improved classification with random forest (73.1% vs. 57.7% baseline, p = 0.08). A single whole-hand temperature value does not reliably distinguish flare from non-flare states. Fine-grained classification was not achievable at this sample size once leakage was corrected. A coarser threshold showed a promising but inconclusive trend, warranting validation in a larger cohort.