Predictors and associations of complications in ureteroscopy for stone disease using AI: outcomes from the FLEXOR registry

We aimed to develop machine learning(ML) algorithms to evaluate complications of flexible ureteroscopy and laser lithotripsy(fURSL), providing a valid predictive model. 15 ML algorithms were trained on a large number fURSL data from > 6500 patients from the international FLEXOR database. fURSL complications included pelvicalyceal system(PCS) bleeding, ureteric/PCS injury, fever and sepsis. Pre-treatment characteristics served as input for ML training and testing. Correlation and logistic regression analysis were carried out by a multi-task neural network, while explainable AI was used for the predictive model. ML algorithms performed excellently. For intraoperative PCS bleeding, Extra Tree Classifier achieved the best accuracy at 95.03% (precision 80.99%), and greatest correlation with stone diameter(0.21) and residual fragments(0.26). PCS injury was best predicted by RandomForest (accuracy 97.72%, precision 63.50%). XGBoost performed best for ureteric injury (accuracy 96.88%, precision 60.67%). Both demonstrated moderate correlation with preoperative characteristics. Postoperative fever was predicted by Extra Tree Classifier with 91.34% accuracy (precision 58.20%). Cat Boost Classifier predicted postoperative sepsis with 99.15% accuracy (precision 66.38%), and the best overall performance. At logistic regression, postoperative fever/sepsis positively correlated with preoperative urine culture(p = 0.001). ML represents a powerful tool for automatic prediction of outcomes. Our study showed promises in algorithms training and validation on a very large database of patients treated for urolithiasis, with excellent accuracy for prediction of complications. With further research, reliable predictive nomograms could be created based on ML analysis, to serve as aid to urologists and patients in the decision making and treatment planning process.