Missingness-Aware Knowledge Distillation for Low Birthweight and Preterm Birth Prediction Under State-Specific Item Missingness in U.S. Birth Certificates

Birth-certificate items that carry much of the antenatal signal for perinatal risk prediction (month prenatal care began, pre-pregnancy body mass index, WIC participation, smoking before pregnancy, maternal education and payer) are also the items most often left unknown, and their completeness differs across U.S. states. Models developed on complete records are therefore deployed under missingness they never saw. We evaluate missingness-aware knowledge distillation (MAKD): a gradient-boosted teacher is trained on records complete for all six items, and students are distilled from its cross-fitted soft predictions to operate under the observed missingness of each state. Using NCHS public-use natality files (2016, 2017, 2023, 2024) and state item-completeness rates from CDC WONDER, we imposed the missingness of 51 jurisdictions on national records under completely-at-random and covariate-dependent mechanisms and compared students with imputation, missing-indicator and non-distilled baselines on sensitivity and positive predictive value (PPV) at a 10% alert rate, calibration, and subgroup performance in 2024 births. Test-year prevalence was 6.9% for low birthweight (LBW) and 8.5% for preterm birth (PTB). At the missingness rates states actually report, all tree-based methods stayed within a fraction of a percentage point of the PPV of the full-information teacher, and the best method differed by outcome. On 200,000 real 2024 records with naturally missing items, the pooled distilled student had the highest discrimination for both outcomes, with paired bootstrap intervals for its AUROC advantage over mean imputation (LBW +0.004 to +0.006; PTB +0.009 to +0.011) and iterative imputation (LBW +0.013 to +0.016; PTB +0.016 to +0.018) excluding zero. Evaluating models on naturally incomplete records, not only on simulated masks, changed the conclusion about which missing-data strategy to prefer.