Detecting inflammatory arthritis in hand smartphone photographs: development and validation of a computer vision model in clinical settings.

Abstract

Computer vision can classify inflammatory arthritis on smartphone photographs. We aimed to train, validate and judiciously choose a model for detecting hand synovitis from standardized smartphone photographs in a real-world rheumatology outpatient population. A dataset of 2296 hand photos from 1112 patients attending rheumatology clinics in India was partitioned at the patient level into training (70%), validation (15%), and test (15%) sets. Two deep learning architectures (ConvNeXt-V2 and EfficientNet-V2) and their weighted ensemble were trained against a ground truth of specialist-detected synovitis and compared using Area under Receiver Operating Curve (AUROC). In the chosen model, 95% confidence intervals were obtained via patient-level bootstrap in the independent test set. Prespecified subgroup analyses examined model performance by deformity status, age, and sex. ConvNeXt-v2 outperformed EfficientNet-v2 (validation AUROC 0.856 vs 0.831). The ensemble achieved the highest validation AUROC (0.864, α = 0.24), with modest incremental gain over ConvNeXt. On the independent test set, ConvNeXt achieved an AUROC of 0.852 (95% CI 0.802-0.896). At the fixed operating threshold, test accuracy was 0.79 (95% CI 0.75-0.83), sensitivity 0.76 (0.67-0.85), specificity 0.80 (0.75-0.84). Model performance remained stable across all prespecified subgroups, including patients with hand deformities. A computer vision model trained on standardized smartphone photographs can detect hand synovitis in routine clinical populations including in those with deformities. This validated model on a large, prospectively assembled dataset represent an important step toward scalable decision support in non-specialist settings and reducing diagnostic delay in inflammatory arthritis.