Machine learning reveals hidden subtypes of intestinal injury in preterm infants

Necrotizing enterocolitis remains a major challenge in neonatology because of its substantial clinical burden, yet the diagnostic label itself may be hiding more than it reveals. Since the modified Bell’s staging criteria were introduced in 1986, clinicians and researchers have relied on a system originally designed to grade clinical severity, not to define a disease. Those criteria rest on clinical and radiographic findings that are often nonspecific and open to interpretation, and they have effectively lumped biologically diverse forms of intestinal injury under a single diagnostic umbrella. Now, a study published in Pediatric Research has used unsupervised machine learning to pull that umbrella apart. Analyzing prospectively collected data from 20,352 very low birth weight infants in the Korean Neonatal Network between 2013 and 2022, researchers identified distinct clinical patterns of vulnerability to intestinal injury without ever telling the algorithm what necrotizing enterocolitis looks like.

The motivation for the diagnosis-blind approach comes from a growing recognition that the condition currently termed necrotizing enterocolitis is not a single disease entity. Accumulating evidence points to interacting pathways, including developmental immaturity, impaired enteral feeding tolerance, intestinal dysbiosis, and systemic inflammation. Conventional risk models have struggled to predict early disease, progression to surgery, or mortality, and even supervised machine learning models built on larger sets of clinical variables remain constrained by their reliance on necrotizing enterocolitis as the primary outcome label, a diagnosis that is itself heterogeneous and imprecise. A recent single-center study applied unsupervised hierarchical clustering to infants with suspected intestinal injury and identified five phenotypes without incorporating diagnostic labels, proposing a novel taxonomic framework. But whether such structure could be recovered at scale, across an entire nation, remained an open question the new study set out to answer.

The cohort drew on 77 neonatal intensive care units across Korea. After excluding 61 records with missing birth date or sex, 20,352 eligible infants remained. The dataset contained 78 variables, 65 binary and 13 continuous, and the pipeline deliberately masked clinical anchors such as gestational age, birth weight, head circumference, body length, mortality, and all necrotizing enterocolitis and spontaneous intestinal perforation-specific variables. Probabilistic principal component analysis then projected the 78 variables into six principal components, explaining 32.8 percent of total variance, using an expectation-maximization algorithm to accommodate missing values without discarding any clinical features. One notable exception preserved missingness: for the variable describing days to achieve enteral feeding above 100 mL/kg/day, absent data often reflect early mortality rather than random absence, so it was retained as informative. Hierarchical agglomerative clustering with Ward’s linkage and Euclidean distance was applied to the component scores, and the optimal number of primary clusters was set at eight based on the within-cluster sum of squares elbow point.

That choice involved a trade-off the authors acknowledge openly. The silhouette score at eight clusters was 0.146, far below the peak of 0.893 at two clusters, but eight provided the clinical granularity required for meaningful phenotypic stratification. Further examination of the hierarchical dendrogram revealed 27 subclusters capturing finer detail. To characterize each subcluster, the team used a dual-threshold strategy, selecting features that both contributed strongly to the principal components, with an absolute loading of at least 0.15, and showed significant differences across subclusters in univariate analyses with two-sided P values below 0.01. This ensured that the defining features were mathematically influential in the model and clinically discriminative at the same time.

The overall cohort told a sobering story before any clustering began: in-hospital mortality was 12.73 percent, or 2,591 of 20,352 infants, and the incidence of necrotizing enterocolitis at stage 2 or higher was 6.30 percent, or 1,282 infants. Yet the eight clusters, derived entirely without diagnostic labels, gestational age, birth weight, or mortality data, separated these outcomes strikingly. Clusters 1, 7, and 8 were characterized by relatively lower gestational ages and birth weights, but their profiles diverged sharply. Cluster 8 showed the highest prevalence of stage 2 or greater disease at 17.9 percent, 496 of 2,778 infants, and surgical disease at 13.0 percent, 362 of 2,778, nearly three times the cohort average. Clusters 1 and 7, in contrast, showed markedly higher mortality with lower disease prevalence. Cluster 8 alone harbored 38.7 percent of all infants with stage 2 or greater disease and 52.7 percent of those requiring surgery across the entire cohort, marking it as a phenotypic hallmark of severe intestinal injury.

Within Cluster 8, secondary clustering revealed five subclusters, designated SC23 through SC27, each defined by a dominant clinical context. SC23 had the highest rate of antenatal antibiotic exposure at 64.9 percent and exceptionally favorable survival at 98 percent. SC24 was predominantly associated with assisted reproduction, including in vitro fertilization at 68.5 percent and multiple gestation at 79.8 percent. SC25 showed a high prevalence of maternal pregnancy-induced hypertension at 36.8 percent. SC26 and SC27 represented the most unstable and high-risk phenotypes, with mortality rates of 28.6 percent and 25.2 percent respectively. SC26 was distinguished by antenatal inflammation, including preterm premature rupture of membranes at 78.9 percent and histologic chorioamnionitis at 68.1 percent. SC27 combined out-of-hospital births at 27.4 percent with the heaviest burden of severe comorbidities, including sepsis at 69.7 percent, hypotension at 82.5 percent, and grade 3 or higher intraventricular hemorrhage at 32.9 percent, alongside acute pulmonary complications such as early pulmonary hypertension at 17.5 percent, air-leak syndrome at 20.5 percent, and massive pulmonary hemorrhage at 23.1 percent. SC27 also showed the most protracted recovery, with a median of 69 days to achieve full enteral feeding and 94 days of invasive mechanical ventilation.

Because unsupervised clusters are exploratory by nature, the team validated the structure with a supervised random survival forest trained on survival time and event indicators, using log-rank splitting and Nelson-Aalen survival estimates. The concordance between blinded cluster assignments and predicted mortality was high: Harrell’s C-indices reached 0.965, with a 95 percent confidence interval of 0.949 to 0.980, on the validation set and 0.976, from 0.967 to 0.985, on the test set, with time-dependent areas under the receiver operating characteristic curve of 0.985 and 0.994 respectively. Restricted mean survival time comparisons confirmed strong agreement between observed Kaplan-Meier curves and model-predicted cluster-level survival trajectories. The authors note, however, that predictive concordance was limited in Clusters 1 and 7, because a globally trained model cannot perfectly reflect minority trajectories characterized by exceptionally high early mortality.

The subcluster profiles align with plausible, though not proven, pathophysiological pathways. The high antibiotic exposure in SC23 likely reflects suspected antenatal infection or inflammation, and the authors suggest that timely treatment may have mitigated downstream inflammatory burden, although specific indications for antibiotic use were unavailable. SC26’s prominent antenatal inflammatory exposure with relatively low antibiotic treatment is consistent with prior evidence that antenatal inflammation may prime the immature intestine toward postnatal injury through dysregulated immune activation and impaired barrier function. SC25’s association with pregnancy-induced hypertension, a recognized cause of placental insufficiency, fits a pathway of impaired fetal intestinal development under chronic hypoxic-ischemic conditions; small-for-gestational-age status was significantly more frequent in this subcluster, although it was not included in cluster characterization due to low component loading. SC24 may similarly reflect placental insufficiency and growth discordance. SC27’s hemodynamic instability and systemic hypoxia may predispose the immature intestine to ischemia-reperfusion injury, and antenatal steroid exposure was lowest in this group, a factor that may have contributed to adverse outcomes given steroids’ potential role in preserving intestinal mucosal barrier integrity. The direction of causality for delayed enteral feeding remains uncertain, since it may reflect both underlying vulnerability and a consequence of evolving injury.

Interpretability analysis using SHapley Additive exPlanations highlighted delayed achievement of full enteral feeding as a dominant contributor to mortality risk prediction. The authors are careful about what these findings do and do not mean: the identified features should not be interpreted as causal risk factors or as a reclassification of the disease, and the framework was not developed as a real-time early prediction tool. Still, the clinically defined phenotypes show intriguing alignment with recent transcriptomic stratification of surgically treated cases, which revealed molecular subtypes dominated by inflammatory or hypoxia and vascular development patterns, supporting the potential for an integrated multiomic taxonomy with phenotype-specific biomarkers. Among the limitations, the retrospective nationwide design requires prospective external validation in independent populations; unsupervised clustering remains sensitive to initial feature selection and hyperparameter settings; and a methodological paradox arises because post-hoc disease prevalence was ultimately used to identify the high-risk subpopulation, a step the authors defend as indispensable for confirming external validity while acknowledging its conceptual circularity. The registry also lacks laboratory biomarkers and radiographic findings. Even so, the demonstration that standard registry data alone can recover a five-phenotype taxonomy, linking antenatal antibiotic exposure, assisted reproduction, maternal hypertension, antenatal inflammation, and severe postnatal morbidities to distinct patterns of intestinal vulnerability, marks a decisive shift away from predicting a single vague disease toward recognizing holistic, data-driven risk patterns in the smallest patients.

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

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

Article References: Oh, S. H., Choi, Y.-S., Chung, S.-H., Ahn, J., Kim, J. Y., Lee, E. H., Aghaeepour, N., Stevenson, D., & Neu, J. (2026). Distinct clinical patterns of preterm infants vulnerable to intestinal injury identified by unsupervised clustering. Pediatric Research. https://doi.org/10.1038/s41390-026-05494-3

Image Credits: AI Generated

DOI: 10.1038/s41390-026-05494-3

Keywords: Distinct, clinical, patterns, preterm, infants, vulnerable, intestinal, injury, identified, unsupervised, clustering, scientific research