Early postnatal risk stratification for severe adverse outcomes in twin neonates admitted to the neonatal intensive care unit: development and temporal validation of an interpretable machine learning model.

BACKGROUND: Twin neonates face disproportionately higher risks of severe composite adverse outcomes, such as intraventricular hemorrhage (IVH), periventricular leukomalacia (PVL), and bronchopulmonary dysplasia (BPD), compared to singletons. However, specific predictive tools for this vulnerable population are lacking, as existing scoring systems often fail to account for twin-specific physiological dynamics. This study aimed to develop and validate an interpretable machine learning (ML) model for early risk stratification of adverse outcomes in twin neonates.

METHODS: This single-center retrospective cohort study included twin neonates admitted to the neonatal intensive care unit (NICU) at Shanxi Children's Hospital. A derivation cohort (n=912; July 2022-June 2023) was used for model development, and a temporally separated cohort (n=592; July-December 2023) for temporal validation. Missing data were addressed using multiple imputation. We developed and compared ML prediction models, evaluating discrimination, calibration, and decision-curve analysis. Shapley additive explanations (SHAP) were used to provide clinician-facing global and patient-level explanations of risk estimates.

RESULTS: After comparing four feature selection strategies, the 10-feature subset identified by least absolute shrinkage and selection operator (LASSO) was utilized for model development. In temporal validation, random forest (RF) and gradient boosting (GB) models showed comparable discrimination [area under the curve (AUC): 0.851 vs. 0.844]. GB demonstrated a more balanced classification performance across clinically relevant thresholds, with acceptable calibration and positive net benefit on decision-curve analysis, and was selected as the final model. A web-based risk calculator was implemented using the GB model.

CONCLUSIONS: We developed and temporally validated an interpretable GB-based ML model using routinely available NICU variables to support early risk stratification for severe adverse outcomes in twin neonates. Combined with a web-based risk calculator and SHAP-based interpretability, this model may assist NICU clinicians in identifying higher-risk twin neonates and prioritizing closer monitoring or early intervention. Multicenter external validation is warranted before broader clinical implementation.