A Predictive Tool Powered by Machine Learning for Evaluating the Status of Surgical Margins After Robot-Assisted Radical Prostatectomy

Objective This research intends to create a predictive model utilizing machine learning(ML) techniques to evaluate the likelihood of positive surgical margins (PSM) in individuals receiving robot-assisted radical prostatectomy (RARP), thus aiding in clinical decision-making. Methods A retrospective study was conducted involving 301 patients who underwent RARP. The subjects were randomly assigned to two groups in a ratio of 7:3, which included 201 individuals in the training cohort and 100 in the validation cohort. A total of twenty-four clinical and oncological characteristics were gathered, initially assessed through univariate logistic regression, and later refined using feature selection facilitated by the Boruta algorithm. Utilizing the chosen features, seven distinct ML models were developed. The effectiveness of these models was comprehensively assessed through a range of metrics, including the area beneath the ROC curve and the F1 score. To conduct a detailed analysis of how features influence the optimal model, the SHAP approach was utilized for evaluating feature contributions. Results In the final analysis, there were 301 patients included, revealing a postoperative incidence of PSM at 42.0% after RARP. Through univariate logistic regression and the Boruta algorithm, five key predictive variables were recognized for the construction of the model. Of the seven ML models assessed, the ANN model demonstrated the best performance, achieving an AUC of 0.808 (95% CI: 0.702–0.899) on the validation dataset, along with superior levels of accuracy (80.11%), sensitivity (78.9%), and F1 score (77.9%). Analysis using SHAP indicated that an advanced clinical stage, increased levels of creatinine, high-risk stratification of prostate cancer, and a greater percentage of positive biopsies were strongly linked to a heightened risk of PSM. In contrast, neoadjuvant therapy demonstrated a protective influence on the occurrence of PSM. Conclusion Machine learning models demonstrate significant utility in predicting positive surgical margins after RARP. Integrating the random forest model with the SHAP interpretation framework enables precise prediction of individual PSM risk and provides intuitive insights into the impact of key features on predictive outcomes. This approach facilitates preoperative risk stratification and the development of early postoperative intervention strategies.