To explore the predictive value of machine learning models based on clinical and radiomics features for the risk of sepsis after percutaneous nephrolithotomy (PCNL), and to provide a scientific basis for early prevention and intervention in clinical practice.

A retrospective analysis was conducted on the data of 339 patients with kidney stones who underwent PCNL from January 2020 to January 2024. Patients were divided into sepsis and non-sepsis groups based on the quick sequential organ failure assessment (qSOFA) score at 24 h postoperatively. Clinical data were collected and radiomics features were extracted from CT images. After a multi-step feature selection strategy, 8 radiomics features with significant predictive value were retained. Patients were randomly divided into a training set (n=237) and a test set (n=102) at a ratio of 7:3. Eight machine learning algorithms, namely logistic regression (LR), support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), extremely randomized trees (ExtraTrees), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and multi-layer perceptron (MLP), were used to construct the clinical model and radiomics model. Subsequently, the two models with the highest performance were integrated to develop a combined clinical-radiomics model for sepsis prediction, and a nomogram was established accordingly. Calibration curve was drawn to evaluate the calibration efficacy of the nomogram, and Hosmer-Lemeshow fitting analysis was used to evaluate the calibration ability of the nomogram. Decision curve analysis (DCA) was drawn to evaluate the clinical utility of the prediction model. Model performance was assessed by the area under the receiver operating characteristic (ROC) curve (AUC).

Among the 339 patients in this study, 24(7.1%) developed sepsis postoperatively. Comparisons between the two groups revealed significant differences in gender, preoperative suspicion of infectious stones, stone burden, hemoglobin, positive urine nitrite, urine white blood cells, total blood protein, globulin, and prealbumin (P<0.05). Multivariate logistic regression further showed that preoperative suspicion of infectious stones (OR=5.589, 95%CI: 1.659-18.834, P=0.005), positive urine nitrite (OR=5.312, 95%CI: 1.802-15.662, P=0.002), and total blood protein <68.1 g/L (OR=0.245, 95%CI: 0.071-0.846, P=0.026) were independent risk factors for postoperative sepsis after PCNL. Consequently, these three indicators were selected as parameters for the construction of a clinical prediction model. In the test set of the clinical model, the ExtraTrees model exhibited superior performance among the eight machine learning models, achieving an AUC of 0.743. In the test set of the radiomics model, the KNN model demonstrated the highest performance among the eight machine learning models, with an AUC of 0.878. By integrating the aforementioned clinical and radiomics model, a combined clinical-radiomics model for sepsis prediction was developed, and a corresponding nomogram was constructed. The results indicated that in the test set, the integrated model incorporating both clinical and radiomics features achieved the best diagnostic performance, with an AUC of 0.898.

The clinical-radiomics combined model based on the ExtraTrees algorithm and the KNN algorithm can effectively predict the risk of postoperative sepsis after PCNL, which is helpful for clinicians to identify high-risk patients early and intervene in a timely manner, thereby reducing the incidence of postoperative sepsis.