Machine learning model integrating radiomics and clinical features for predicting postoperative bleeding after percutaneous nephrolithotomy.

BACKGROUND AND

objectiveRadiomics and machine learning play significant roles in clinical medical research, particularly in developing prediction models. This study aims to combine radiomic features with clinical variables and apply machine learning to predict the risk of postoperative bleeding after percutaneous nephrolithotomy (PCNL). MATERIALS AND

methodsA retrospective study analyzed 151 patients who had PCNL at the Second Affiliated Hospital of Nanchang University. Clinical variables linked to postoperative bleeding were identified using univariate analysis, and radiomic features were screened using the least absolute shrinkage and selection operator algorithm (lasso regression). Logistic regression, Random Forest (RF), and Support Vector Machine (SVM) were then used to develop prediction models based on the correlated clinical variables and radiomic features. The predictive accuracy of these models was assessed through identification and calibration. In addition, the clinical utility of the optimal model was evaluated using decision curve analysis (DCA), and the significance of each radiomic and clinical feature was elucidated through SHapley Additive exPlanations (SHAP) to interpret the model's predictions.

resultsPostoperative statistics revealed that the postoperative bleeding rate was 31.1% (n = 47), the blood transfusion rate was 1.99% (n = 3), and the final probability of requiring vascular embolization was 1.32% (n = 2). The accuracy rates for predicting postoperative bleeding in patients with PCNL using logistic regression, RF and SVM were 75.6%, 75.6%, and 71.1%, respectively. The corresponding area under the curve AUC (95% CI) were 0.760 (0.719-0.802), 0.740 (0.691-0.800) and 0.631 (0.558-0.693). The top four prediction importance scores in logistic regression and RF were wavelet-HLH_glrlm_ShortRunLowGrayLevelEmphasis, wavelet-HLH_glrlm_LowGrayLevelRunEmphasis, stone shape, operation time and stone shape, stone diameter, operation time, wavelet-HLH_glrlm_ShortRunLowGrayLevelEmphasis. The DCA decision curve demonstrates that the logistic regression model's prediction of postoperative bleeding offers patients enhanced clinical benefits. The SHAP analysis further clarifies the contribution of each variable in the prediction process

conclusionThe machine learning model has demonstrated excellent performance in predicting the bleeding situation after percutaneous nephrolithotomy. This model may provide a basis for surgical risk stratification, thereby facilitating the development of more personalized postoperative monitoring plans.