Comparison of Two Models Predicting IVF Success
Comparison of Two Models Predicting IVF Success
Of the 9060 IVF cycles of the NL couples, 63% were from couples with primary infertility. In the NL cohort fewer women were over 35 years of age than in the Nelson and Templeton cohorts (Table I). Moreover, fewer NL women had infertility lasting >3 years. In the Templeton study the cause of infertility was more frequently coded as 'tubal'. Fewer NL women had three or more cycles that may be explained by the fact that in the NL cohort only cycles performed during the first year following the start of treatment were taken into account. The success rate in the NL data of 21.2% per IVF cycle was closer to the Nelson success rate of 23.4% than to the Templeton rate of 13.9%.
The discriminative performance of the Nelson and Templeton models on the NL data was similar (Table II). The predicted-to-observed ratios of the original models revealed overestimation of the Nelson model (values above 1) and underestimation of the Templeton model (values smaller than 1). After adjustment for changing success rates, the Nelson model still showed slight overestimation, while the Templeton calibration became almost perfect. These results are further illustrated in Fig. 2. The Nelson model showed systematic overestimation of success rates with line and triangles below the dotted line of perfect calibration (e.g. a predicted probability of 20% corresponds to an observed proportion of ~15%). The Templeton model systematically underestimated the success rates, with the line and triangles above the dotted line (e.g. a predicted probability of 20% corresponds to an observed proportion of nearly 30%). The results indicate that the adjustments for the success rate of IVF considerably improved the calibration of both models, with excellent calibration for the Templeton model.
(Enlarge Image)
Figure 2.
Calibration plots for the Templeton and Nelson models. The left column shows the Nelson model applied in the Netherlands (NL) cohort and the right column shows the Templeton model applied in the NL cohort. Calibration is shown for the original models (first row), and for the models after adjusting for the IVF success rates (second row). Triangles indicate quintiles of cycles, ordered according to predicted success rate.
Results
Of the 9060 IVF cycles of the NL couples, 63% were from couples with primary infertility. In the NL cohort fewer women were over 35 years of age than in the Nelson and Templeton cohorts (Table I). Moreover, fewer NL women had infertility lasting >3 years. In the Templeton study the cause of infertility was more frequently coded as 'tubal'. Fewer NL women had three or more cycles that may be explained by the fact that in the NL cohort only cycles performed during the first year following the start of treatment were taken into account. The success rate in the NL data of 21.2% per IVF cycle was closer to the Nelson success rate of 23.4% than to the Templeton rate of 13.9%.
Discrimination and Calibration
The discriminative performance of the Nelson and Templeton models on the NL data was similar (Table II). The predicted-to-observed ratios of the original models revealed overestimation of the Nelson model (values above 1) and underestimation of the Templeton model (values smaller than 1). After adjustment for changing success rates, the Nelson model still showed slight overestimation, while the Templeton calibration became almost perfect. These results are further illustrated in Fig. 2. The Nelson model showed systematic overestimation of success rates with line and triangles below the dotted line of perfect calibration (e.g. a predicted probability of 20% corresponds to an observed proportion of ~15%). The Templeton model systematically underestimated the success rates, with the line and triangles above the dotted line (e.g. a predicted probability of 20% corresponds to an observed proportion of nearly 30%). The results indicate that the adjustments for the success rate of IVF considerably improved the calibration of both models, with excellent calibration for the Templeton model.
(Enlarge Image)
Figure 2.
Calibration plots for the Templeton and Nelson models. The left column shows the Nelson model applied in the Netherlands (NL) cohort and the right column shows the Templeton model applied in the NL cohort. Calibration is shown for the original models (first row), and for the models after adjusting for the IVF success rates (second row). Triangles indicate quintiles of cycles, ordered according to predicted success rate.
