The following is a summary of “Development of a scoring model for the Sharp/van der Heijde score using convolutional neural networks and its clinical application,” published in the June 2023 issue of Rheumatology by Honda, et al.

For a study, researchers sought to develop a predictive model for the Sharp/van der Heijde score (SHS) and assess its suitability for clinical research settings.

They constructed the SHS predictive model using convolutional neural networks (CNN) and their in-house rheumatoid arthritis (RA) image database. The process involved three steps: image orientation, joint detection, and damage prediction. Additionally, a predictive model for radiographic progression (ΔSHS >3/year) was developed using a graph convolutional network (GCN). A multiple regression model assessed the association between the predicted SHS using the CNN model and clinical features.

The CNN model achieved 100% accuracy in image orientation correction and accurate detection of joint coordinates, with all predicted coordinates being within 10 pixels of the correct ones. In the damage prediction phase, the model’s performance, as measured by κ values, was high (0.879 for erosion and 0.865 for joint space narrowing) compared to expert evaluations. Minimal overfitting to expert 1 scoring was observed using a dataset scored by experts 1 and 2. The CNN model predicted high-titre rheumatoid factor (RF) as an independent risk factor for ΔSHS per year in biologics users. The GCN model demonstrated superior performance (AUCs of 0.753 and 0.734) in predicting ΔSHS >3/year compared to other models for patients with and without biologics at baseline. The GCN model identified RF titer as the most important feature for predicting ΔSHS >3/year in biologics users.

The study successfully developed a high-performance SHS scoring model suitable for clinical research applications. The model held promise for improving assessments and predictions of radiographic progression in patients with rheumatoid arthritis.

Source: academic.oup.com/rheumatology/article/62/6/2272/6758238