Machine learning vs human experts: sacroiliitis analysis from the RAPID-axSpA and C-OPTIMISE phase 3 axSpA trials.

Objective: Diagnosis of axial spondyloarthritis (axSpA) is primarily established through the identification of the presence or absence of radiographic sacroiliitis. However, the reliability of conventional radiographs (X-rays) is undermined by significant interreader variability. A machine learning tool could reduce diagnosis time, thereby minimising interreader variability. The present study aimed to evaluate the performance of a deep learning model for detecting radiographic sacroiliitis in axSpA patients from the RAPID-axSpA (NCT01087762) and C-OPTIMISE (NCT02505542) trials.

Methods: Radiographs from the RAPID-axSpA and C-OPTIMISE cohorts were retrospectively used. The deep learning model was previously trained by using a transfer learning approach on non-medical data. The model's agreement with expert readers was tested on baseline X-rays using central reader data. Sensitivity, specificity, Cohen's κ, positive and negative predictive values and the area under the receiver operating characteristics curve were calculated.

Results: The model's performance was evaluated in the RAPID-axSpA (n = 277) and C-OPTIMISE (n = 739) cohorts. In RAPID-axSpA, the model achieved 82% sensitivity, 81% specificity and a Cohen's κ of 0.61, closely matching central reader performance. In C-OPTIMISE, the model demonstrated 90% sensitivity, 56% specificity and a Cohen's κ of 0.48. The agreement between the model and central readers was 82% (RAPID-axSpA) and 75% (C-OPTIMISE).

Conclusions: The tested deep learning model exhibited accurate radiographic sacroiliitis detection in axSpA patients from diverse clinical trials. The proposed deep learning model could expedite diagnosis, reduce healthcare resource usage and improve patient care pathways.