Forecasting knee arthroplasty surgery with deep learning : an integrated approach using routine clinical data and radiographs.

Aims: We aimed to develop and validate a practical deep learning model integrating commonly collected clinical data and knee radiographs to predict the need for knee arthroplasty (total or partial) in patients with, or at risk of, knee osteoarthritis, as well as the time to surgery. Methods: Data from the Multicentre Osteoarthritis Study (MOST) and the Osteoarthritis Initiative (OAI) were used. The MOST dataset, comprising 3,026 patients, was the primary training and testing cohort, while the OAI dataset provided external validation. The final architecture was based on DenseNet-201, with a head that combined outputs from the radiological analysis with commonly collected clinical data. Model evaluation used the area under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC). Results: The integration of clinical and radiological data significantly improved predictive accuracy. The combined model achieved an AUROC of 0.85 and AUPRC of 0.62, outperforming models using either data source alone. External validation with the OAI dataset yielded an AUROC of 0.79, confirming the model's generalizability. The AUROC and AUPRC for surgical interventions within 40 months was 0.83 and 0.26, respectively, on the validation dataset, demonstrating higher predictive accuracy for earlier surgical needs. Conclusion: This study highlights the potential of deep learning models, which integrate clinical and radiological data, to predict the need for knee arthroplasty. The robust performance and generalizability of the developed model could streamline clinical pathways and predict local demand for surgery during the next three years. This will facilitate resource planning for providers and accurate and timely access to surgical interventions for patients.