Implementing the use of nomograms by choosing threshold points in predictive models: 2012 updated Partin Tables vs a European predictive nomogram for organ-confined disease in prostate cancer

Abstract

Objectives: To implement the use of nomograms in clinical practice showing how to choose thresholds in nomograms' predictions to select risk groups. To validate and compare the predictive ability and clinical utility of the Hospital Universitario 'Miguel Servet' (HUMS) and the updated Partin Tables 2012 (PT-2012) nomograms to predict organ-confined disease (OCD) after radical prostatectomy (RP).

Patients and methods: Cohort of 1285 patients with prostate cancer treated with RP at Instituto Valenciano de Oncología (IVO) between 1986 and 2011. The predictive value of the nomograms was assessed by means of calibration curves, discrimination ability (area under the receiver operating characteristic (ROC) curve (AUC) and probability density functions). The clinical utility was evaluated through Vickers' decision curves and thresholds were chosen through probability density functions.

Results: The calibration curves showed a minimal underestimation in low probabilities (<20%), a minimal overestimation in high probabilities (>50%) in the HUMS nomogram and a regular minimal overestimation in the PT-2012. Their AUC of 0.7285 (95% confidence interval [CI] 0.7010-0.7559) and 0.7288 (95%CI 0.7013-0.7562) respectively, show an adequate discrimination ability for both predictive models in the IVO cohort. The decision curves show similar net benefits for both models. In this study we advocate for a threshold of 53% for the identification of OCD.

Conclusions: The HUMS-nomogram and the PT-2012 predictions of OCD confirm their utility in a contemporary cohort of patients. Patients with a probability of OCD >53% should be classified as OCD, helping physicians to better counsel their patients. A selection of adequate thresholds, as presented in this paper, makes nomograms more accessible tools.

Keywords: nomograms; organ-confined disease; prostate cancer; thresholds; validation.