Introduction: We developed and validated ensemble machine-learning based prostate cancer (PC) and clinically significant prostate cancer (csPC) risk calculator and deployed it by web-embedded structure for clinician’s decision support before prostate biopsy.
Methods: We used 3791 patient’s data for development and validation risk calculator. The csPC defined by Gleason grade group 3 to 5. We split the data for development and test set. XGboost algorithm was used for development calculator with 5 folds cross-validation after hyperparameter tuning and feature selection. We assessed area under curve (AUC) of receiver operating characteristic curve by test set for each of calculator.
Results: About 1216 (32.7%) patients and 562 (14.8%) patients were diagnosed as PC and csPC. Average PSA was 31.7±438.0 mg/dL and prostate volume was 44.4±23.2 cc in the database. 2843 patients’ data were used for development and cross-validation and 948 patents data were used for test set. Total 8 variables (number of biopsies, age, prostate specific antigen(PSA), free PSA, testosterone, total prostate volume, prostate transitional volume and hypoechoic lesion on ultrasound) was selected PC risk calculator and 7 variables (age, PSA, free PSA, testosterone, total prostate volume, prostate transitional volume and hypoechoic lesion on ultrasound) were selected for csPC risk calculator by lasso regression. The AUC of final model for PC was 0.869 (95% confidence interval (CI); 0.844-0.893) and for csPC was 0.945 (95% CI; 0.927-0.963).
Conclusions: We successfully develop and validation of ensemble machine learning based decision supporting tool for prostate biopsy. You can access it by https://boramae-pcrc.appspot.com/. Source of