Multi-cohort modeling strategies for scalable globally accessible prostate cancer risk tools

Online clinical risk prediction tools built on data from multiple cohorts are increasingly being utilized for contemporary doctor-patient decision-making and validation. This report outlines a comprehensive data science strategy for building such tools with application to the Prostate Biopsy Collaborative Group prostate cancer risk prediction tool. We created models for high-grade prostate cancer risk using six established risk factors. The data comprised 8492 prostate biopsies collected from ten institutions, 2 in Europe and 8 across North America. We calculated area under the receiver operating characteristic curve (AUC) for discrimination, the Hosmer-Lemeshow test statistic (HLS) for calibration and the clinical net benefit at risk threshold 15%. We implemented several internal cross-validation schemes to assess the influence of modeling method and individual cohort on validation performance. High-grade disease prevalence ranged from 18% in Zurich (1863 biopsies) to 39% in UT Health San Antonio (899 biopsies). Visualization revealed outliers in terms of risk factors, including San Juan VA (51% abnormal digital rectal exam), Durham VA (63% African American), and Zurich (2.8% family history). Exclusion of any cohort did not significantly affect the AUC or HLS, nor did the choice of prediction model (pooled, random-effects, meta-analysis). Excluding the lowest-prevalence Zurich cohort from training sets did not statistically significantly change the validation metrics for any of the individual cohorts, except for Sunnybrook, where the effect on the AUC was minimal. Therefore the final multivariable logistic model was built by pooling the data from all cohorts using logistic regression. Higher prostate-specific antigen and age, abnormal digital rectal exam, African ancestry and a family history of prostate cancer increased risk of high-grade prostate cancer, while a history of a prior negative prostate biopsy decreased risk (all p-values < 0.004). We have outlined a multi-cohort model-building internal validation strategy for developing globally accessible and scalable risk prediction tools.