High-Resolution Thermography and Artificial Intelligence to Evaluate and Classify Rheumatoid Arthritis

Objectives Thermography, which captures heat signatures, is a novel method of evaluating articular joints in inflammatory arthritis.[1] We sought to evaluate the utility of thermography of the hands and feet, combined with artificial intelligence, in the evaluation of Rheumatoid Arthritis (RA). Methods RA patients (n = 100) were recruited from an academic rheumatology clinic (Winnipeg, Canada). Healthy controls (n = 137) were recruited through advertising. Thermal images of the dorsal aspects of the hands and feet were captured with a Flir A700 camera. Regions of interest (ROI) were pre-identified as joints commonly affected in RA (Figure 1A). ROIs were manually gated, and the maximum/minimum/average temperatures were extracted. Data analysis was performed in R to identify temperature differences between groups, followed by machine learning classification using ROI-restricted data. Thermograms were also analyzed using a deep learning model which deployed unsupervised feature extraction (DINOV2) and classification (ElasticNet) in Python. Results Principal components analysis using ROI thermogram data revealed clear separation between RA and HC (Figure 1B). Differential analysis identified 56 of 114 temperature parameters that were significantly higher in RA patients. For example, the maximum temperatures of the right 5th PIP joint (adjusted p=1.3×10 −8 ) and right 4th PIP joint (adjusted p=7.6×10 −8 ) were higher in RA compared to controls. All thermogram parameters (minimum, maximum, average) were increased in joints that were tender or swollen on clinical examination in RA (Figure 1C, all p-values <0.0001). Using thermogram ROI data, XGboost achieved an AUC of.869 to classify RA from control. Rankings of variable importance by SHAP index included average temperatures of several hand joints including the Wrist and 3rd PIP. A computer vision model achieved high performance in classifying RA from controls using entire thermograms (without ROI information), with an AUC of 0.977 and recall of 0.909 (Figure 1D). Principal components from this model were mapped to red, green, and blue channels to visualize thermographic differences between RA and control subjects (Figure 1E). Conclusion Thermal imaging is a low-cost, accessible method for detecting synovitis in small joints and can distinguish RA patients from controls. It may enable early detection of subtle joint inflammation at the earliest stages of RA onset, such as in individuals with clinically suspect arthralgia. References [1.] Kow J, Tan YK. Joint Bone Spine 2023;90:105496. Best Abstract on Clinical or Epidemiology Research by a Trainee - Phil Rosen Award