Threshold-Based Stratification of Predicted Preterm Birth Risk in a National Natality Cohort: An Empirical...

Clinical prediction becomes actionable only when estimated risk is connected to a decision boundary. The prevalence threshold is a likelihood-ratio-derived landmark that may complement conventional risk stratification, but its empirical behaviour in obstetric populations is incompletely described.

To evaluate how fixed probability cutoffs and a formally derived prevalence threshold partition predicted preterm birth risk in a national cohort.

We analysed singleton births to U.S. residents in the 2024 natality public-use file. Preterm birth was delivery before 37 completed weeks using the obstetric estimate. A logistic model generated predicted probabilities, and adjusted risk ratios were estimated by modified Poisson regression with robust variance. We assessed discrimination, calibration across risk tenths, fixed cutoffs of 5%, 10%, 15% and 20%, and a cervical-length prevalence threshold derived from a sensitivity of 23.3% and a specificity of 93.6%.

Among 3,004,719 complete-case pregnancies, 263,034 (8.8%) were preterm. The c-statistic was 0.634 (95% CI 0.633, 0.636), with close agreement between predicted and observed risk across tenths. Predicted risk had a mean of 8.75%, a median of 7.40%, and a range of 4.83%-84.22%. At the 10%, 15% and 20% cutoffs, 17.7%, 7.1% and 4.5% of pregnancies were above threshold, with observed preterm birth rates of 16.9%, 24.4% and 27.4%, respectively. The cervical-length likelihood ratio was 3.64, giving a prevalence threshold of 34.4%; 21,212 pregnancies (0.71%) were above this landmark, with an observed preterm birth rate of 38.1%.

Higher fixed cutoffs selected smaller, more outcome-enriched groups while capturing fewer preterm births. The derived prevalence threshold supplied a transparent evidentiary landmark for a specified downstream test, but it was not a utility-derived treatment threshold or a screening recommendation.