Although lymphovascular invasion is a major predictor of bladder urothelial carcinoma, postoperative pathological biopsy-an intrusive procedure-remains the only reliable diagnostic technique for this condition. By combining logistic regression with various machine learning techniques, this work seeks to assess its risk factors and create a non-invasive preoperative prediction model. 520 patients who had never received treatment were included in the retrospective data collection from two centers, lymphovascular invasion was the main result. Based on the originating center, the cohort was split into an cross-center validation set (176 patients) and a training set (344 patients). Six machine learning techniques and univariate/multivariate regression were used to examine risk factors for lymphovascular invasion and build models. Multi-criteria evaluation was used to choose the best model, while SHapley Additive exPlanations(SHAP) was used for interpretability analysis. Ultimately, the best model was used to create an interactive online calculator. The area under the receiver operating characteristic curve(AUC) for the regression model based on traditional techniques was 0.968 on the training set and 0.972 on the cross-center validation set. With AUC values of 0.986 and 0.982, respectively, the random forest model outperformed all other machine learning models. In contrast to conventional analytical techniques, our model highlighted the relative influence of seven different risk factors on lymphovascular invasion and uncovered additional relevant factors, such as the neutrophil-to-lymphocyte ratio and alcohol consumption. A successful online lymphovascular invasion risk calculator was created.
Zhou et al. (Sat,) studied this question.