The nomogram conundrum: a demonstration of why a prostate cancer risk model in Turkish men underestimates prostate cancer risk in the USA

Abstract

The utility of a nomogram is based on the patient population it is designed for-and their inherent properties and biases. Our aim was to demonstrate the variability in predictive model accuracy and utility between different populations. Our model is based on 761 men who underwent initial TRUS biopsy at a single institution in Turkey. Patients were included if they had at least 10 cores on biopsy and PSA level < 20 ng/ml. Multivariable logistic regression models were used to develop a new nomogram. External validity was tested with two different cohorts one from another institution in Turkey (N = 136) and cohort from USA (N = 2242). Prostate cancer (PCa) and high-grade PCa was diagnosed in 249/761 (32.7 %) and 101/761 (13.3 %) patients from Ankara, Turkey, respectively. Predictors of PCa were age (p < 0.0001, OR 2.11), PSA (p = 0.044, OR 1.44), PV (p < 0.0001, OR 0.38), %fPSA (p = 0.016, OR 0.72), and abnormal DRE (p < 0.0001, OR 2.05). The predictive accuracy (c-index) of our nomogram was 73 %. C-indices of 71 and 70 % were recorded in external validation cohorts from Turkey and the USA, respectively. Virtually ideal calibration was recorded for the internal validated predictive model, and good calibration was recorded when applied to the Istanbul cohort. However, the model/nomogram underestimates PCa risk in the US cohort. This is the first nomogram predicting the risk of PCa at initial biopsy in a Turkish population and provides a good risk estimation tool with good predictive accuracy and calibration in the Turkish populations. However, our study demonstrates the poor transferability of predictive tools to widely different populations.