Predicting Outcomes of Preterm Neonates Post Intraventricular Hemorrhage

Abstract

Intraventricular hemorrhage (IVH) in preterm neonates presents a high risk for developing posthemorrhagic ventricular dilatation (PHVD), a severe complication that can impact survival and long-term outcomes. Early detection of PHVD before clinical onset is crucial for optimizing therapeutic interventions and providing accurate parental counseling. This study explores the potential of explainable machine learning models based on targeted liquid biopsy proteomics data to predict outcomes in preterm neonates with IVH. In recent years, research has focused on leveraging advanced proteomic technologies and machine learning to improve prediction of neonatal complications, particularly in relation to neurological outcomes. Machine learning (ML) approaches, combined with proteomics, offer a powerful tool to identify biomarkers and predict patient-specific risks. However, challenges remain in integrating large-scale, multiomic datasets and translating these findings into actionable clinical tools. Identifying reliable, disease-specific biomarkers and developing explainable ML models that clinicians can trust and understand are key barriers to widespread clinical adoption. In this prospective longitudinal cohort study, we analyzed 1109 liquid biopsy samples from 99 preterm neonates with IVH, collected at up to six timepoints over 13 years. Various explainable ML techniques-including statistical, regularization, deep learning, decision trees, and Bayesian methods-were employed to predict PHVD development and survival and to discover disease-specific protein biomarkers. Targeted proteomic analyses were conducted using serum and urine samples through a proximity extension assay capable of detecting low-concentration proteins in complex biofluids. The study identified 41 significant independent protein markers in the 1600 calculated ML models that surpassed our rigorous threshold (AUC-ROC of ≥0.7, sensitivity ≥ 0.6, and selectivity ≥ 0.6), alongside gestational age at birth, as predictive of PHVD development and survival. Both known biomarkers, such as neurofilament light chain (NEFL), and novel biomarkers were revealed. These findings underscore the potential of targeted proteomics combined with ML to enhance clinical decision-making and parental counseling, though further validation is required before clinical implementation.

Keywords: biomarker; intensive care; intraventricular hemorrhage; machine learning; neonate; posthemorrhagic hydrocephalus; prediction; prematurity; proteomics; survival.

Figure 1

Graphical representation of defined events in a timeline. Dotted red lines intersecting the timeline mark instances of IVH and the corresponding NSI of EVD/Ommaya reservoir. The IVH event encompasses samples collected from the onset of bleeding up to 2 days afterward. The IVHp event spans samples obtained from 3 to 9 days post-bleeding. The PHVD event is delineated by NSI, covering samples from 2 days before the intervention until the intervention itself. Subsequent PHVDp1, PHVDp2, and PHVDp3 events include samples from 1 day after NSI up to 8 days after, 9 to 39 days after, and 40 days or more after NSI, respectively. Comparable timeframes were established for IVH patients without the development of PHVD/without NSI. The equivalent PHVD event comprises samples from 10 to 18 days after IVH, while the equivalent PHVDp1, PHVDp2, and PHVDp3 events include samples from 10 to 18 days, 19 to 49 days, and 50 days or more after IVH, respectively. Further, we included standard time frames, indicated by blue diamond-shaped rectangles, including 28 days of life (±7 days), 32 weeks after conception (±7 days), and term-equivalent age (GA at sampling time 36.0-41.14).

Figure 2

(A)-visualization of the relevance associations network based on the rCCA between distinct clinical datasets characterizing the samples and the Olink measurements based on the urine data. The applied correlation threshold for inclusion in both networks was set to 0.6. (B)-visualization of the relevance associations network based on the rCCA between distinct clinical data sets characterizing the samples and the Olink measurements based on the serum data. (C)-heatmap of the rCCA between distinct clinical datasets characterizing the samples and the Olink measurements based on the urine data. The dendrogram in the heatmap was computed with the complete linkage method to find similar clusters based on the Euclidean distance. (D)-heatmap of the rCCA between distinct clinical datasets characterizing the samples and the Olink measurements based on the serum data. The dendrogram in the heatmap was computed with the complete linkage method to find similar clusters based on the Euclidean distance. All clinical data beginning with “Event_” indicate the samples belonging to a specific timeframe; these timeframes are highly correlated to features such as “day of life”, “day after IVH” or “GA of the sample”.