A Leakage-Proof Benchmark and Conformal Selective Triage for Electrohysterogram-Based Preterm Birth Prediction

Abstract

Preterm birth remains a major cause of neonatal morbidity and mortality worldwide. Electrohysterography (EHG), a noninvasive measure of uterine myoelectrical activity, has been studied for preterm-birth prediction, but performance estimates may be biased when segments from the same maternal record are split across training and validation data. We formalized the distinction between segment-level and patient-independent record-grouped validation, established a patient-independent benchmark on the Term-Preterm Electrohysterogram Database, and evaluated class-conditional conformal selective prediction. All 300 records (38 preterm) were analyzed across three prespecified regimes. A 92-feature elastic-net logistic model was evaluated with record-grouped nested cross-validation; preprocessing, model fitting, Platt calibration, and conformal estimation were confined to training data, and performance was calculated from strictly out-of-fold record-level predictions with 1,000 bootstrap resamples. Under patient-independent evaluation, AUROC was 0.493 (95% CI 0.467-0.520), AUPRC 0.122 (0.095-0.152), and Brier score 0.115 (0.094-0.137). AUROC was 0.514 at or before 26 weeks and 0.469 thereafter. At miscoverage alpha=0.10, marginal coverage was 0.897, abstention 72.7%, and singleton-prediction accuracy 0.624. These findings establish a record-separated reference benchmark for EHG prediction and a broader validation principle for segmented physiological data: the unit of resampling should correspond to the unit at which predictive performance is intended to generalize. Conformal prediction further quantifies when the available information supports a singleton classification and when uncertainty warrants deferral.

Accepted for publication in the Proceedings of the 48th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC 2026), Toronto, Canada. Paper ID 4188