Enhancing discharge location prediction for patients undergoing orthopedic surgery using side-tuning

Background Postoperative care planning is crucial for efficient allocation of scarce hospital resources and improving patient recovery. However, delays in organizing appropriate post-discharge care often result in prolonged hospital stays, leading to increased bed occupancy and limiting capacity for new surgical admissions. Integrating artificial intelligence (AI) in preoperative planning of post-discharge care can enhance hospital workflow efficiency. Side-tuning is a novel form of transfer learning that could help in improving predictive accuracy. This study aims to evaluate the accuracy of AI models for predicting discharge location in patients undergoing elective orthopedic surgery. Method In this cohort study, electronic health record (EHR) data from the surgical department (n = 33,140) was leveraged to improve prediction accuracy for the orthopedic department (n = 14,976) using transfer learning. Logistic regression, random forest, and neural network models were compared, with the neural network further optimized through feature transfer, fine-tuning, and side-tuning. Model performance was evaluated using area under the curve (AUC) and F1 scores, and SHAP-value analysis provided insights into key predictors. Results The neural network model with side-tuning demonstrated the highest predictive performance (AUC = 0.63, F1 = 0.56), outperforming other models (AUC = 0.49–0.61). This study contributes to the growing field of AI-driven healthcare by evaluating the feasibility of discharge prediction. Conclusions While current predictive performance remains limited, our findings highlight opportunities for future research to refine model development, integrate additional data sources, and explore the added value of transfer learning and particularly the novel method of side-tuning to improve real-world applicability.