Timo Soeterik

112 CHAPTER 6 Recalibration Insertion of the original linear predictor as the only variable in the original model resulted in the following coefficients; 1.42 for the intercept and 0.8258 for the beta coefficients. Adding 1.42 to the original intercept and multiplying all separate beta coefficients with 0.8258 corrected the original linear predictor, leading to a new model: Ln (odds EPE) = -2.77 + 0.049*PSA + 1.18*ISUP2 + 1.44*ISUP3 + 2.12*ISUP4 + 0.31*Tumour involvement >50% + 1.21*EPE present on MRI Recalculation of the AUC using the update formula showed no change in discriminative ability, since the AUC remained 0.78 (95% CI 0.75 − 0.82). As shown in the new calibration plot, agreement between predicted probability and prevalence of the outcome improved (Figure 3). Also, calibration-in-the-large showed similar mean values of predicted and observed probabilities (respectively, 32% and 32%). Although calibration improved graphically after the model adjustment, the Hosmer-Lemeshow test was still statistically significant ( p = 0.0007). As shown in Figure 3 and Table 4, agreement between predicted and observed probability was excellent in group 1 and 2 and fair in groups 6 and 8. In the other groups we observed substantial miscalibration; both under- and overestimation of the predicted risk (Table 4). Clinical usefulness Sensitivity and specificity per risk threshold were presented in Table 5. As shown in this table, sensitivity was excellent if a threshold of 7% was used (prostate lobes with predicted risk ≥7% were regarded as EPE positive). However, this would lead to high false positive rates as specificity was low (2%). Highest cumulative optimum (respectively 140%) would be reached if risk thresholds of either 25% or 40% were used. Decision curve analysis showed that use of the updated version of the nomogram would be superior to relying on MRI, with regard to prediction of side-specific EPE, for all clinically relevant thresholds (Figure 4). Highest net benefit would be achieved if risk thresholds between 20% and 30% would be used (Figure 4).