Improved Prediction of Significant Prostate Cancer Following Repeated Prostate Biopsy by the Random Forest Classifier

Abstract

Purpose: Numerous strategies and diagnostic tests were proposed in patients suspected of clinically significant (cs) prostate cancer (PCa) after an initial negative prostate biopsy. The study aimed to create a Random Forest (RF) classifier for predicting the probability of csPCa in specimens taken by the repeated systematic prostate biopsy (SBx), and to determine its diagnostic accuracy and clinical utility. Methods: This retrospective, single-center study included patients who underwent repeated SBx due to clinical suspicion of cancer. Data on patient age, serum prostate-specific antigen (PSA) levels, prostate volume, digital rectal examination, first-degree family history, and histology findings from the SBx were collected for all patients. The area under the curve (AUC), and secondary metrics of clinical prediction models were used to assess their discriminative abilities. Clinical usefulness of final model was tested by the decision curve analysis (DCA). The explainability and website placement of the ML model were also performed. Results: In total, 204 patients were eligible for analysis. The csPCa was detected in 26% (n = 53) patients. The AUC, accuracy, sensitivity, and specificity for detection of csPCa were 0.94, 0.91, 0.84, and 0.98, respectively. With an optimal threshold of 0.8, about 34% of unnecessary biopsies would be avoided, but correct diagnosis would be delayed in 4.4% csPC cases. PSA level, prostate volume, and age were the top-ranked variables in the RF model. Conclusion: The RF classifier predicts csPCa with good accuracy and may help urologists when deciding whether the repeated biopsy is necessary to avoid being too invasive.