Urosepsis is the most hazardous complication of percutaneous nephrolithotomy (PNL). Early warning and timely intervention are essential to reduce the mortality associated with urosepsis. We aimed to develop and validate a preoperative machine learning (ML) algorithm to predict the urosepsis after PNL based on electronic medical record (EMR) data.

We retrospectively screened 360 patients with urosepsis and 2636 patients without urosepsis. Multimodal clinical parameters were collected, including demographics, admission vital signs, imaging reports, and laboratory tests. Participants were randomly allocated into the training cohort (80%) and the validation cohort (20%). Five ML algorithms were applied to construct the models, including support vector machines (SVCs), random forest (RF), Logistic Regression (LR), XGBoost and Stacking. The predictive abilities of algorithms were assessed by receiver operating characteristic (ROC) curves.

Out of 83 clinical parameters, 44 variables were statistically significant. In the training cohort, the Random Forest (RF) model exhibited the highest accuracy, with an Area Under the ROC curve (AUC) of 0.987. In the validation cohort, all models achieved an AUC of over 0.7. Among them, the Stacking model showed the best predictive performance with an AUC of 0.754. Neutrophils and white blood cell count were the most important factors included in the 31 parameters of the Stacking model.

We established a ML model with 31 parameters from EMR data, which can help to distinguish patients with high risks of urosepsis before the operation. This will allow clinicians to make appropriate interventions to reduce the mortality from PNL.