OP0300 USE OF THERMOGRAPHY OF HANDS AND MACHINE LEARNING TO QUANTIFY JOINT INFLAMMATION AND ESTIMATE DAS28, CDAI, SDAI IN PATIENTS WITH RHEUMATOID ARTHRITIS

Abstract

Background:Disease activity scores such as DAS28, CDAI and SDAI are used in the follow-up of patients with rheumatoid arthritis (RA). These scores include variables obtained on physical examination such as the tender joint count (TJC) and the swollen joint count (SJC). In telematic consultations, it is not possible to determine these variables by physical joint assessment. Therefore, it is necessary to develop new tools that allow detecting joint inflammation in places close to the patient. Thermography is a safe and fast technique that measures heat through infrared imaging. Inflammation of the joints causes an increase in temperature and can therefore be detect by thermography. Machine learning methods are highly accurate in analyzing medical images automatically.Objectives:To develop an algorithm that, based on thermographic images of hands and machine learning, learn to quantify joint inflammation in patients with RA and estimate the DAS28, CDAI, SDAI by including the patient global health (PGH).Methods:Multicenter observational study conducted in the rheumatology and radiology service of two hospitals. Patients with RA, psoriatic arthritis (PA), undifferentiated arthritis (UA) and arthritis of hands secondary to other diseases (SA) that attended the follow-up visits were recruited. Companions of patients and healthcare professionals were also recruited as healthy subjects (HS). In all cases, a thermographic image of the hands was taken using a Flir One Pro or a Thermal Expert TE-Q1 camera connected to a smartphone. Ultrasound (US) of both hands was performed in patients with RA, PA, UA and SA. The degree of synovial hypertrophy (SH) and power doppler (PD) was assessed for each joint (score from 0 to 3). Machine learning was used to quantify joint inflammation (SH+PD) from the thermal images using US as ground truth. RA patients whose thermal image was taken with the Thermal Expert TE-Q1 camera were used to evaluate the performance (test dataset). The other participants were used as training dataset. The TJC, SJC, PGH, C-reactive protein (CRP) and erythrocyte sedimentation rate (ESR) were also assessed in the test dataset. A linear regression was used to estimate the DAS28, CDAI and SDAI with the resultant joint inflammation quantification from the thermal images and the PGH. Performance was evaluated by means of Pearson’s correlation coefficient. The study was approved by the Clinical Ethics and Research Committee of both centers.Results:The total number of recruited subjects was 521 (422 for the training and 99 for the testing dataset). In the training dataset, the thermography of 296 patients was taken with the Flir One Pro (163 RA, 17 PA, 22 UA, 12 SA and 82 HS) and 126 with the Thermal Expert TE-Q1 camera (6 RA without clinical data, 20 PA, 7 UA, 23 SA and 70 HS).We found higher correlations between joint inflammation variables (US and SJC) and thermography (0.48, p<0.01 for US and 0.48, p<0.01 for SJC) than between joint inflammation variables (US and SJC) and the PGH (0.29, p<0.01 for US and 0.35, p<0.01 for SJC). Thermography did not show statistically significant correlation with the PGH (0.14, p=0.164). The linear regression of thermography and the PGH showed strong correlation with the DAS28 (0.73, p<0.01), CDAI (0.84, p<0.01) and SDAI (0.82, p<0.01).Conclusion:Thermography of hands and machine learning can effectively quantify joint inflammation and can be used in combination with the PGH to estimate disease activity scores. These results open an opportunity to develop tools that facilitate telematic consultations in patients with RA.