Objective: Neonatal intraventricular hemorrhage (IVH) is a major cause of mortality and morbidity, particularly following premature birth. Even after the acute phase, posthemorrhagic hydrocephalus is a long-term complication, frequently requiring permanent ventriculoperitoneal shunt (VPS) placement. Currently, there are no risk classification methods integrating the constellation of clinical data to predict short- and long-term prognosis in neonatal IVH. To address this need, the authors developed a two-part machine learning approach for predicting short- and long-term outcomes after diagnosis of neonatal IVH. Integrating both maternal and neonatal characteristics, they developed a binary classifier to predict short-term mortality risk and a clinical scale to predict the long-term risk of VPS placement.
Methods: Neonates with IVH were identified from the Optum Clinformatics Data Mart administrative claims database. Matched maternal and childbirth characteristics were obtained for all patients. The primary endpoints of interest were short-term (30 day) mortality and long-term VPS placement. Classification of short-term mortality risk was evaluated using 5 different machine learning approaches and the best-performing method was validated using a withheld validation subset. Prediction of long-term shunt risk was performed using a multivariable Cox regression model with stepwise variable selection, which was subsequently converted to an easily applied integer risk scale.
Results: A total of 5926 neonates with IVH were identified. Most patients were born before 32 weeks' gestation (67.2%) and with low birth weight (81.2%). Empirical 30-day mortality risk was 10.9% across all IVH grades and highest among grade IV IVH (34.3%). Among the neonates who survived > 30 days, actuarial 12-month postdiagnosis risk of shunt placement was 5.4% across all IVH grades and 31.3% for grade IV IVH. The optimal short-term risk classifier was a random forest model achieving an area under the receiver operating characteristic curve of 0.882 with important predictors ranging from gestational age to diverse comorbid medical conditions. Selected features for long-term shunt risk stratification were IVH grade, respiratory distress syndrome, disseminated intravascular coagulation, and maternal preeclampsia or eclampsia. An integer risk scale, termed the Shunt Prediction After IVH in Neonates (SPAIN) scale, was developed from these 4 features, which, evaluated on withheld cases, demonstrated improved risk stratification compared with IVH grade alone (Harrell's concordance index 0.869 vs 0.852).
Conclusions: In a large cohort of neonates with IVH, the authors developed a two-pronged, integrated, risk classification approach to anticipate short-term mortality and long-term shunt risk. The application of such approaches may improve the prognostication of outcomes and identification of higher-risk individuals who warrant careful surveillance and early intervention.
Keywords: IVH; hydrocephalus; intraventricular hemorrhage; machine learning; neonatal; predictive modeling; ventriculoperitoneal shunt.