A multicenter study on the application of artificial intelligence radiological characteristics to predict prognosis after percutaneous nephrolithotomy.

Background: A model to predict preoperative outcomes after percutaneous nephrolithotomy (PCNL) with renal staghorn stones is developed to be an essential preoperative consultation tool. Objective: In this study, we constructed a predictive model for one-time stone clearance after PCNL for renal staghorn calculi, so as to predict the stone clearance rate of patients in one operation, and provide a reference direction for patients and clinicians. Methods: According to the 175 patients with renal staghorn stones undergoing PCNL at two centers, preoperative/postoperative variables were collected. After identifying characteristic variables using PCA analysis to avoid overfitting. A predictive model was developed for preoperative outcomes after PCNL in patients with renal staghorn stones. In addition, we repeatedly cross-validated their model's predictive efficacy and clinical application using data from two different centers. Results: The study included 175 patients from two centers treated with PCNL. We used a training set and an external validation set. Radionics characteristics, deep migration learning, clinical characteristics, and DTL+Rad-signature were successfully constructed using machine learning based on patients' pre/postoperative imaging characteristics and clinical variables using minimum absolute shrinkage and selection operator algorithms. In this study, DTL-Rad signal was found to be the outstanding predictor of stone clearance in patients with renal deer antler-like stones treated by PCNL. The DTL+Rad signature showed good discriminatory ability in both the training and external validation groups with AUC values of 0.871 (95% CI, 0.800-0.942) and 0.744 (95% CI, 0.617-0.871). The decision curve demonstrated the radiographic model's clinical utility and illustrated specificities of 0.935 and 0.806, respectively. Conclusion: We found a prediction model combining imaging characteristics, neural networks, and clinical characteristics can be used as an effective preoperative prediction method.