A deep learning–radiomics model for predicting flexible ureteroscope failure: Toward clinical decision support in endourology

A predictive model was constructed combining clinical and radiomics-associated characteristics to predict flexible ureteroscope damage. Preoperative/postoperative variables were collected from 457 patients with stones obstructing upper urinary tract that underwent flexible ureteroscopy (fURS) at our center. Principal component analysis (PCA) was applied for identifying characteristic variables and avoid overfitting. A model was established for prediction on fURS outcomes in patients with upper urinary tract stones. Additionally, we validated the model’s predictive accuracy and clinical relevance through cross-validation with data from our center. Data was extracted from a total of 457 patients classified into training and external validation sets. We utilized machine learning techniques, including minimum absolute shrinkage and selection operator (LASSO) algorithms, to develop DTL+Rad-signature, radiomics features, and deep migration learning models based on patients' pre- and postoperative imaging data and clinical variables. The parameter DTL-Rad signal yielded the best predictive performance for flexible ureteroscope repair. DTL+Rad model exhibited strong discriminative power in both the training and external validation cohorts, yielding area under the curve (AUC) of 0.871 (95% CI, 0.800-0.942), as well as 0.744 (95% CI, 0.617-0.871), respectively. Decision curve analysis illustrated that model provided a clinical net benefit with specificities of 0.935 and 0.806 in training and validation groups, respectively. A prediction model was established combinedly applying imaging and clinical features and neural networks for flexible ureteroscope repair in patients with upper urinary calculi.