Predictive model for PSA persistence after radical prostatectomy using machine learning algorithms

Abstract

Objective: To evaluate the efficacy of a machine learning model for predicting prostate-specific antigen (PSA) persistence after radical prostatectomy (RP).

Methods: Data from 470 patients who underwent RP at the Affiliated Hospital of Qingdao University from January 2018 to June 2021 were retrospectively analyzed. Ten risk factors, including age, body mass index (BMI), preoperative PSA, biopsy Gleason score, total prostate specific antigen density (PSAD), clinical tumor stage, clinical lymph node status, seminal vesicle invasion, capsular invasion and positive surgical margin, were included in the analysis. The data were randomly divided into a training set and a test set at a ratio of 7:3, and seven different machine learning algorithms were compared. The confusion matrix, receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) were used to evaluate the diagnostic performance of the model, and the random forest algorithm found to be the optimal prediction model.

Results: In the entire cohort, 142 (30.21%) patients developed PSA persistence. Based on all included risk factors, the random forest model had the best effect among the seven models, with an AUC of 0.8607 in the training set and 0.8011 in the test set. The feature importance results showed that capsular invasion, positive surgical margin, preoperative PSA and biopsy Gleason score were the four most important risk factors for PSA persistence after RP.

Conclusion: The Random Forest algorithm performed excellently in this study and can be used to construct a predictive model for PSA persistence. By incorporating clinical data from the Asian region and exploring the risk factors for PSA persistence, this study contributes to the existing research and aids clinicians in assessing the risk of PSA persistence occurrence, enabling timely treatment planning and improving patient prognosis.

Keywords: PSA persistence; machine learning; prediction model; radical prostatectomy; random forest algorithm.