Presentation Authors: Jungyo Suh*, Young Cheol Hwang, Jae Hyun Jung, Junghoon Lee, Won Hoon Song, Yu Jin Kang, Ja Hyeon Ku, Cheol Kwak, Hyeon Hoe Kim, Chang Wook Jeong, Seoul, Korea, Republic of
Introduction: Although machine learning (ML) methods show promising results in diagnostic fields, well-designed logistic regression model provides still more powerful results for detecting prostate cancer than a single machine learning model. We want to assess the feasibility of the machine learning based approach for developing prostate cancer risk calculator in prostate biopsy patients.
Methods: We use a multicenter database of 4954 patients prostate biopsy which was used for developing and validation of Seoul national university hospital prostate cancer risk calculator (SNUPC-RC). Development dataset included 3482 patients information from May. 2003 to Nov. 2010. Validation dataset consisted of 1161 patients information from Dec. 2010 to Jun. 2012. Age, PSA level, prostate volume, nodule in the digital rectal exam, nodule in transrectal sonography, and obtained biopsy core was used for variables. Gradient boosting model (GBM), Random forest (RF) and single layered deep learning model (DL) were developed with 10-fold cross-validation. After the ensemble of each developed model, hyperparameter tuning by grid methods. The predictive accuracy of the final ensemble model's area under the curve (AUC) to SNUPC-RC, ERSPC, and PCPT by DeLong methods. All machine learning and statistic analysis were conducted by R and H2O package.
Results: The predictive accuracy of each developed ML model's AUC was 0.740, 0.771 and 0.762 for each of RF, GBM and DL model. After stacking of three ML models, the ensemble model was developed by weighting for each model with 0.21, 0.23, and 0.73 to each of RF, GMB, and DL models. After 10 fold cross validations best of best learning model's AUC was 0.810 and average of developed ensemble model's AUC was 0.810. Compared with each of logistic regression models, the ML-based prediction model was not inferior to SNUPC-RC (AUC=0.811, p=0.733) but more accurate than ERSPC (AUC=0.768, p < 0.001) and PCPT (AUC=0.704, p < 0.001).
Conclusions: In this pilot development study for the ML-based prostate cancer risk calculator, a single ML model was not effective than nomogram, which developed by logistic regression. By using the ensemble methods, the predictive accuracy of the ML-based model can be increased as the most accurate regression model. External validation of the developed ML-model using updated prostate biopsy data is needed.