Abstract
Objective The objective of this study was to create three point-of-care predictive models for
very preterm birth using variables available at three different time points: prior
to pregnancy, at the end of the first trimester, and mid-pregnancy.
Study Design This is a retrospective cohort study of 359,396 Ohio Medicaid mothers from 2008 to
2015. The last baby for each mother was included in the final dataset. Births prior
to 22 weeks were excluded. Multivariable logistic regression was used to create three
models. These models were validated on a cohort that was set aside and not part of
the model development. The main outcome measure was birth prior to 32 weeks.
Results The final dataset contained 359,396 live births with 6,516 (1.81%) very preterm births.
All models had excellent calibration. Goodness-of-fit tests suggested strong agreement
between the probabilities estimated by the model and the actual outcome experience
in the data. The mid-pregnancy model had acceptable discrimination with an area under
the receiver operator characteristic curve of approximately 0.75 in both the developmental
and validation datasets.
Conclusion Using data from a large Ohio Medicaid cohort we developed point-of-care predictive
models that could be used before pregnancy, after the first trimester, and in mid-pregnancy
to estimate the probability of very preterm birth. Future work is needed to determine
how the calculator could be used to target interventions to prevent very preterm birth.
Key Points
-
We developed predictive models for very preterm birth.
-
All models showed excellent calibration.
-
The models were integrated into a risk calculator.
Keywords
preterm birth - predictive model - infant mortality
