Predicting Miscarriage in Pregnancies of Uncertain Viability
Predicting Miscarriage in Pregnancies of Uncertain Viability
During the study period 575 consecutive women with a PUV were recruited. The outcome was known in 513 (89.2%) pregnancies, and 260/513 (50.7%) pregnancies were viable and 253/513 (49.3%) non-viable at the end of the first trimester. Analysis was performed on the 400 patients for whom all covariates and the outcome were known. Most exclusions (n = 103) were due to unknown LMP. The PBAC and presence of the yolk sac was unknown for nine and two pregnancies, respectively, whereas one patient had both variables missing. Of the 400 pregnancies, 200 (50%) were viable and the rest were non-viable at the end of first trimester. The data collection method is shown in Fig. 1. Patients who miscarried had heavy bleeding in the interim with no pregnancy visualized on the follow-up scan at 7–14 days, a non-viable pregnancy diagnosed at a 7–14 day follow-up scan or had a viable pregnancy at the follow-up scan but subsequently miscarried before the end of the first trimester.
(Enlarge Image)
Figure 1.
Flowchart showing data collection for external validation of models and simple scoring systems to predict miscarriage in intrauterine pregnancies of uncertain viability.
The model (Mo) was defined as
where maternal age is taken to be 35 years when the actual age is >35 years. The estimated chance of a viable pregnancy is then obtained as . On the new external validation set, Mo obtained an area under the ROC curve (AUC) of 0.845 (95% confidence interval (CI) 0.806–0.884). The SSo obtained an AUC of 0.832 (95% CI 0.792–0.872) in the same set. The ROC curves of both the model and the scoring system are illustrated in Fig. 2.
(Enlarge Image)
Figure 2.
ROC curves for the original model (Mo) and scoring system (SSo) on the test set of the original paper (Bottomley et al., 2011) and the validation set reported in this study. These models are validated on the set of 400 pregnancies with known outcome and all covariates (including known LMP). Both the Mo and SSo perform well on the independent validation (val) data.
As described above, the original model is not able to obtain a risk estimate for the significant number of women who are unable to report the LMP date with accuracy. The new model (Mn) and scoring system (SSn) that did not include information on LMP, were thus built using the training set from the original study data set (Bottomley et al., 2011). The resulting model is defined as
where maternal age was again taken to be 35 years when actual age is >35 years. This model was validated on the 503 patients with known outcome in the external validation study. The model obtained an AUC of 0.801 (95% CI 0.765–0.841).
Table I summarizes the new score system SSn. The score system obtained an AUC of 0.773 (95% CI 0.733–0.812) on the new external validation data. The ROC curves for the new model and scoring system are given in Fig. 3. Calibration plots (not shown) illustrate that the model and scoring systems are well calibrated, indicating that the chance predicted by the model corresponds to the chance observed in the data set.
(Enlarge Image)
Figure 3.
ROC curves for the new model (Mn) and scoring system (SSn) (not including LMP) on the original test set and the validation set reported here. These models are validated on the 503 pregnancies with known outcome and all covariates, except LMP. Both Mn and SSn perform well on the independent validation data. However, due to the lack of data for LMP, the performance is lower than for the original model (Mo).
For clinical practice, we propose to use the original model Mo or scoring system SSo for patients who can report an accurate LMP, and to use the new model Mn or scoring system SSn otherwise. The colour-based representation in Fig. 4 can be used for this purpose.
(Enlarge Image)
Figure 4.
Graphical representation of the models. Whenever a patient reports her LMP, the original scoring system (above, left) can be used to estimate the probability of a viable pregnancy at the end of the first trimester. Whenever a patient is unable to report an LMP, the new scoring system (above, right) can be used.
Results
During the study period 575 consecutive women with a PUV were recruited. The outcome was known in 513 (89.2%) pregnancies, and 260/513 (50.7%) pregnancies were viable and 253/513 (49.3%) non-viable at the end of the first trimester. Analysis was performed on the 400 patients for whom all covariates and the outcome were known. Most exclusions (n = 103) were due to unknown LMP. The PBAC and presence of the yolk sac was unknown for nine and two pregnancies, respectively, whereas one patient had both variables missing. Of the 400 pregnancies, 200 (50%) were viable and the rest were non-viable at the end of first trimester. The data collection method is shown in Fig. 1. Patients who miscarried had heavy bleeding in the interim with no pregnancy visualized on the follow-up scan at 7–14 days, a non-viable pregnancy diagnosed at a 7–14 day follow-up scan or had a viable pregnancy at the follow-up scan but subsequently miscarried before the end of the first trimester.
(Enlarge Image)
Figure 1.
Flowchart showing data collection for external validation of models and simple scoring systems to predict miscarriage in intrauterine pregnancies of uncertain viability.
The model (Mo) was defined as
where maternal age is taken to be 35 years when the actual age is >35 years. The estimated chance of a viable pregnancy is then obtained as . On the new external validation set, Mo obtained an area under the ROC curve (AUC) of 0.845 (95% confidence interval (CI) 0.806–0.884). The SSo obtained an AUC of 0.832 (95% CI 0.792–0.872) in the same set. The ROC curves of both the model and the scoring system are illustrated in Fig. 2.
(Enlarge Image)
Figure 2.
ROC curves for the original model (Mo) and scoring system (SSo) on the test set of the original paper (Bottomley et al., 2011) and the validation set reported in this study. These models are validated on the set of 400 pregnancies with known outcome and all covariates (including known LMP). Both the Mo and SSo perform well on the independent validation (val) data.
As described above, the original model is not able to obtain a risk estimate for the significant number of women who are unable to report the LMP date with accuracy. The new model (Mn) and scoring system (SSn) that did not include information on LMP, were thus built using the training set from the original study data set (Bottomley et al., 2011). The resulting model is defined as
where maternal age was again taken to be 35 years when actual age is >35 years. This model was validated on the 503 patients with known outcome in the external validation study. The model obtained an AUC of 0.801 (95% CI 0.765–0.841).
Table I summarizes the new score system SSn. The score system obtained an AUC of 0.773 (95% CI 0.733–0.812) on the new external validation data. The ROC curves for the new model and scoring system are given in Fig. 3. Calibration plots (not shown) illustrate that the model and scoring systems are well calibrated, indicating that the chance predicted by the model corresponds to the chance observed in the data set.
(Enlarge Image)
Figure 3.
ROC curves for the new model (Mn) and scoring system (SSn) (not including LMP) on the original test set and the validation set reported here. These models are validated on the 503 pregnancies with known outcome and all covariates, except LMP. Both Mn and SSn perform well on the independent validation data. However, due to the lack of data for LMP, the performance is lower than for the original model (Mo).
For clinical practice, we propose to use the original model Mo or scoring system SSo for patients who can report an accurate LMP, and to use the new model Mn or scoring system SSn otherwise. The colour-based representation in Fig. 4 can be used for this purpose.
(Enlarge Image)
Figure 4.
Graphical representation of the models. Whenever a patient reports her LMP, the original scoring system (above, left) can be used to estimate the probability of a viable pregnancy at the end of the first trimester. Whenever a patient is unable to report an LMP, the new scoring system (above, right) can be used.
Source...