Page 106 - Strategies for non-invasive managementof high-grade cervical intraepithelial neoplasia - prognostic biomarkers and immunotherapy Margot Maria Koeneman

Chapter 5
Table 2. Prediction model for spontaneous regression of CIN 2
Variable
Intercept Smoking
PAP 1 or PAP 2 before CIN 2 diagnosis
Concomitant CIN1 in same biopsy
More than one bio- psy containing CIN 2
Regression coefficient (initial model)
0.928 -0.713 1.302
0.928 -0.792
Regression coefficient (inter- nally validated model)*
0.928 -0.677 1.237
0.881 -0.752
Odds ratio (initial model) (95% CI)
-
0.49 (0.18-1.32) 3.68 (0.78 - 17.43)
2.53 (1.05 - 6.12) 0.45 (0.10 - 2.06)
To calculate the absolute risk of regression on CIN 2:
P(regression) = 1/(1+e-Linear part) × 100%.
Linear part = 0.928 - (0.677 * smoking) + (1.237 * PAP 1 or PAP 2 before CIN 2 diagnosis) + (0.881 * concomi- tant CIN 1 in same biopsy) - (0.752 * more than one biopsy containing CIN 2).
*Regression coefficients after adjustment for over-fitting by shrinkage (shrinkage factor = 0.95).
CI, confidence interval.
Performance of the model
The overall performance of the model showed a Nagelkerke’s R2 of 14% and a Brier score of 0.19. Figure 2 shows the ROC curve with the AUC for assessment of the discriminative performance of the prediction model. The AUC was 69.2% [95% confidence interval (CI), 58.5–79.9%], which indicates a reasonable discriminative ability. A non-significant H-L statistic (0.715 with p = 0.982) indicated good model fit. Figure 3 shows the corresponding calibration curve, representing the accuracy of the model. Predicted probabilities ranged from 38% to 95%, with a mean of 70.4% [standard deviation (SD), 14%]. The developed prediction model has a good fit to the reference curve. The calibration plot shows that the predicted probabilities are especially well calibrated from about 70% upwards. When using an outcome of 70% as the cut-off value (with a predicted outcome of >70% indicating disease regression and a predicted outcome of <70% indicating non-regression), the sensitivity is 73.6% and the specificity is 55.3%. This represents a positive predictive value of 79.8% and a negative predictive value of 46.7% in a population with a similar frequency of outcomes. In this cohort, 84 women had a predicted probability of disease regression > 70%. Of these, 17 patients (20%) experienced disease persistence.
104