Enhanced pre-biopsy risk stratification for aggressive prostate cancer using prostate density in an EHR-derived cohort.

Abstract

e17127 Background: Total prostate-specific antigen (PSA) lacks specificity for distinguishing aggressive prostate cancer (Gleason score ≥7) from indolent/benign disease, leading to unnecessary biopsies. Prostate-specific antigen density (PSAD), incorporating prostate volume, may improve diagnostic prediction. We developed and internally validated prediction models comparing PSAD-based versus PSA-based approaches using electronic health record (EHR) data. Methods: We conducted a retrospective cohort study using EHR data from Rush University Medical Center. Men aged > 30 years who underwent prostate biopsy between 2017-2024 were included if they had PSA ≤1 year pre-biopsy, prostate volume from mpMRI/TRUS, and pathologic confirmation. The outcome was aggressive prostate cancer (Gleason ≥7) versus indolent/benign disease. Candidate predictors included demographics, comorbidities, medications, and PSA or PSAD (log 2 -transformed). Logistic regression and random forest models were developed, evaluating PSA and PSAD alone and in combination with clinical predictors. Data were randomly split 80/20 for training/testing with 5-fold cross-validation for internal validation. Results: 673 men total were included, of whom 333 (49%) had aggressive prostate cancer, with a median age of 65 years. PSAD consistently outperformed total PSA across all models. In multivariable logistic regression, PSAD showed a significant association with aggressive cancer (OR per doubling 2.87; 95% CI 2.32–3.61). Age was independently associated with aggressive cancer, whereas race and ethnicity were not significant. Overall, PSAD-based models improved area under the curve (AUC) by an average of 0.10 in training and 0.12 in testing compared with PSA-based models. Conclusions: PSAD significantly improves prediction of aggressive prostate cancer compared with total PSA alone. Incorporating PSAD into pre-biopsy risk stratification algorithms may reduce unnecessary biopsies while maintaining detection of clinically significant disease. EHR-derived clinical notes represent a valuable data source for algorithm development and highlight the importance of integrating structured and unstructured data sources in oncology decision support. Comparison of model performance for aggressive prostate cancer discrimination. Model Predictors Train AUC (95% CI) Test AUC (95% CI) Logistic Regression Total PSA 0.67 (0.63–0.72) 0.56 (0.46–0.66) Logistic Regression Total PSA + Demographics 0.70 (0.66–0.74) 0.57 (0.47–0.67) Logistic Regression PSAD 0.78 (0.74–0.82) 0.67 (0.57–0.76) Logistic Regression PSAD + Demographics 0.79 (0.76–0.83) 0.67 (0.58–0.77) Random Forest Total PSA + Demographics 0.67 (0.63–0.72) 0.50 (0.39–0.60) Random Forest PSAD + Demographics 0.78 (0.74–0.81) 0.65 (0.55–0.75) Demographics included age, race, and ethnicity, based on a p-value < 0.10 in univariable modeling.