To create a machine-learning model for estimating the likelihood of post-retrograde intrarenal surgery (RIRS) sepsis. All consecutive patients with kidney stone(s) only undergoing RIRS in 16 centers were prospectively included (January 2022-August 2023). adult, renal stone(s) only, CT scan (within three months), mid-stream urine culture (within 10 days). concomitant ureteral stone, bilateral procedures. In case of symptomatic infection/asymptomatic bacteriuria, patients were given six days of antibiotics according to susceptibility profiles. All patients had antibiotics prophylaxis. Variables selected for the model: age, gender, age-adjusted Charlson Comorbidity Index, stone volume, indwelling preoperative bladder catheter, urine culture, single/multiple stones, indwelling preoperative stent/nephrostomy, ureteric access sheath, surgical time. Analysis was conducted using Python programming language, with Pandas library and machine learning models implemented using the Scikit-learn library. Machine learning algorithms tested: Decision Tree, Random Forest, Gradient Boosting. Overall performance was accurately estimated by K-Fold cross-validation with three folds. 1552 patients were included. There were 20 (1.3%) sepsis cases, 16 (1.0%) septic shock cases, and three more cases (0.2%) of sepsis-related deaths. Random Forest model showed the best performance (precision = 1.00; recall = 0.86; F1 score = 0.92; accuracy = 0.92). A web-based interface of the predictive model was built and is available at https://emabal.pythonanywhere.com/ CONCLUSIONS: Our model can predict post-RIRS sepsis with high accuracy and might facilitate patient selection for day-surgery procedures and identify patients at higher risk of sepsis who deserve extreme attention for prompt identification and treatment.
Read full abstract