Optimization of clinical risk-factor interpretation and radiological findings with machine learning for PIRADS category 3 patients

Background Due to the low cancer-detection rate in patients with PIRADS category 3 lesions, we created machine learning (ML) models to facilitate decision-making about whether to perform prostate biopsies or monitor clinical information without biopsy results. Methods In our retrospective, single-center study, 101 eligible patients with at least one PIRADS category 3 lesion but no higher PIRADS lesions underwent MRI/US fusion biopsies between September 2017 and June 2020. Thirty additional patients were included as the validation cohort from the next chronological period from June 2020 to October 2020. Our ML research was a supervised classification problem, with a binary output based on pathological reports of cancerous or benign tissue. The clinical inputs were age, prostate-specific antigen (PSA), prostate volume, prostate-specific antigen density (PSAD), and the number of previous biopsies. The radiology-report inputs were the number of lesions, maximum lesion diameter, lesion location, and lesion zone. We subsequently removed the inputs with low importance. Logistic Regression, Support Vector Machine, Naive Bayes, Decision Tree, Random Forest, and eXtreme Gradient Boosting Tree (XGBoost) were employed. From receiver operating characteristic (ROC) curves, we determined Area Under the ROC Curve (AUC), the cut-off point, and sensitivity score (recall score) to evaluate the ML-model performance. Results Twenty-four adenocarcinoma patients had a mean age of 70 +/- 5.79 years, a mean PSA of 12.42 +/- 6.67 ng/ml, a mean prostate volume of 46.49 +/- 23.13 ml, and a mean PSAD of 0.31 +/- 0.22 ng/ml(2). Seventy-seven patients with benign tissue reports had a mean age of 66.39 +/- 6.66 years, a mean PSA of 11.31 +/- 7.50 ng/ml, a mean prostate volume of 65.25 +/- 35.88 ml, and a mean PSAD of 0.19 +/- 0.13 ng/ml(2). On the validation cohort, XGBoost had the best AUC of 0.76, which considered 80% sensitivity and 72% specificity at a probability cutoff of 57%. The remaining possible ML models performed worse with lesser AUC. The worst was Naive Bayes, with AUC of 0.50. Conclusions ML models facilitate PIRADS 3 patient selection for MRI/US fusion biopsies. ML could optimize how we use previously known clinical risk factors to their full potential.