A smart, practical, deep learning-based clinical decision support tool for patients in the prostate-specific antigen gray zone: model development and validation

Abstract

Objective: Despite efforts to improve screening and early detection of prostate cancer (PC), no available biomarker has shown acceptable performance in patients with prostate-specific antigen (PSA) gray zones. We aimed to develop a deep learning-based prediction model with minimized parameters and missing value handling algorithms for PC and clinically significant PC (CSPC).

Materials and methods: We retrospectively analyzed data from 18 824 prostate biopsies collected between March 2003 and December 2020 from 2 databases, resulting in 12 739 cases in the PSA gray zone of 2.0-10.0 ng/mL. Dense neural network (DNN) and extreme gradient boosting (XGBoost) models for PC and CSPC were developed with 5-fold cross-validation. The area under the curve of the receiver operating characteristic (AUROC) was compared with that of serum PSA, PSA density, free PSA (fPSA) portion, and prostate health index (PHI).

Results: The AUROC values in the DNN model with the imputation of missing values were 0.739 and 0.708 (PC) and 0.769 and 0.742 (CSPC) in internal and external validation, whereas those of the non-imputed dataset were 0.740 and 0.771 (PC) and 0.807 and 0.771 (CSPC), respectively. The performance of the DNN model was like that of the XGBoost model, but better than all tested clinical biomarkers for both PC and CSPC. The developed DNN model outperformed PHI, serum PSA, and percent-fPSA with or without missing value imputation.

Discussion: DNN models for missing value imputation can be used to predict PC and CSPC. Further validation in real-life scenarios are need to recommend for actual implementation, but the results from our study support the increasing role of deep learning analytics in the clinical setting.

Conclusions: A deep learning model for PC and CSPC in PSA gray zones using minimal, routinely used clinical parameter variables and data imputation of missing values was successfully developed and validated.