Distinct clinical patterns of preterm infants vulnerable to intestinal injury identified by unsupervised clustering

Current necrotizing enterocolitis (NEC) diagnostic criteria fail to capture the full spectrum of intestinal injury. We aimed to identify diagnosis-independent clinical patterns of intestinal injury among vulnerable very low birth weight (VLBW) infants using unsupervised hierarchical clustering.

Using prospectively collected data from 20,352 VLBW infants in the Korean Neonatal Network (2013-2022), unsupervised clustering was performed after excluding variables reflecting prematurity and NEC diagnosis.

Eight primary clusters (C1-C8) were identified from the dataset without NEC-related labels. Cluster 8, which exhibited the highest prevalence of previously diagnosed NEC ≥ stage 2, was subdivided into five distinct clinical subclusters (SC23-SC27) based on feature importance: antenatal antibiotic exposure (SC23), assisted reproduction (SC24), pregnancy-induced hypertension (SC25), chorioamnionitis/PPROM (SC26), and severe postnatal morbidities (SC27). In terms of hierarchical structure validation, survival patterns across clusters showed strong concordance between Kaplan-Meier estimates and supervised survival model predictions.

Unsupervised clustering successfully identified distinct clinical features among preterm infants vulnerable to intestinal injury, independent of traditional diagnostic labels. These nuanced clusters provide a valuable framework for developing a novel, more comprehensive taxonomy of intestinal injuries in this high-risk population.

The newly proposed framework of intestinal injury in very low birth weight infants comprises clinically distinct phenotypes that are not adequately captured by diagnosis-based NEC classification. Using unsupervised hierarchical clustering in a large multicenter cohort, this study identifies diagnosis-independent phenotypic structure underlying neonatal intestinal injury. These findings reinforce the heterogeneity of intestinal injuries. They provide more nuanced clusters derived using unsupervised machine learning from a large database of patients and provide a framework for the development of more precise taxonomies in preterm infants that obviate the need for the misnomer currently termed NEC.