Machine Learning Reveals Five Distinct Clinical Phenotypes Hidden Within a Single Neonatal Diagnosis

What neonatologists have treated as a single intestinal disease in preterm infants may actually represent a collection of at least five distinct clinical phenotypes with survival rates ranging from 98% to less than 75%, according to new research using machine learning to analyze nearly a decade of clinical data.

The discovery emerged from an analysis of 20,352 very low birth weight infants across 77 Korean neonatal intensive care units between 2013 and 2022, according to research published in Pediatric Research. Rather than training algorithms to predict necrotizing enterocolitis (NEC) as currently defined, researchers took a diagnosis-agnostic approach—letting unsupervised machine learning identify natural patient groupings based on 78 clinical variables without reference to existing disease labels.

When One Diagnosis Obscures Multiple Clinical Patterns

Since Bell's staging criteria were introduced in 1986, clinicians 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, effectively lumping biologically diverse forms of intestinal injury under a single diagnostic umbrella.

The new analysis identified eight primary clinical clusters and 27 subclusters among vulnerable infants. One cluster—Cluster 8—contained 38.7% of all stage 2+ NEC cases and 52.7% of surgical cases despite representing just 13.6% of the total cohort. Within this high-risk group, five subclusters emerged with strikingly different characteristics.

Subcluster SC23, defined by 64.9% antenatal antibiotic exposure, achieved a 98% survival rate. In contrast, SC27—characterized by 27.4% out-of-hospital births, 69.7% sepsis rates, and severe hemodynamic instability—faced 25.2% mortality and required a median 94 days of mechanical ventilation. SC26, marked by 78.9% preterm premature rupture of membranes and 68.1% histologic chorioamnionitis, showed 28.6% mortality, consistent with inflammatory priming as a potential pathophysiological pathway.

SC24 and SC25 represented yet different profiles: the former linked to assisted reproduction (68.5% in vitro fertilization, 79.8% multiple gestation), the latter to maternal pregnancy-induced hypertension (36.8% prevalence).

Beyond Imprecise Labels

The research challenges a fundamental assumption in medical artificial intelligence: that existing diagnostic categories represent meaningful biological entities suitable for training predictive models. Conventional risk models have struggled to predict early disease, progression to surgery, or mortality. Even supervised machine learning models built on larger sets of clinical variables remain constrained by their reliance on NEC as the primary outcome label—a diagnosis that is itself heterogeneous and imprecise.

By ignoring these potentially flawed labels, the unsupervised approach achieved remarkable predictive performance. Random survival forest validation yielded concordance indices of 0.965 to 0.976—substantially higher than typical supervised models constrained by imprecise outcome definitions. Time-dependent area under the curve reached 0.985 and 0.994 in validation and test sets. 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.

Clinical Implications and Limitations

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. SC26's prominent antenatal inflammatory exposure is consistent with prior evidence that antenatal inflammation may prime the immature intestine toward postnatal injury. SC25's association with pregnancy-induced hypertension fits a pathway of impaired fetal intestinal development under chronic hypoxic-ischemic conditions.

The authors emphasize important caveats: 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. The phenotype-specific risk profiles may enable more precise resource allocation, with high-risk subclusters like SC26 and SC27 potentially requiring intensive monitoring protocols, though the direction of causality for many associations remains uncertain.

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.

Broader Implications

The methodology extends beyond neonatal care. Any complex syndrome currently diagnosed through imprecise clinical criteria—from sepsis to acute respiratory distress syndrome—may similarly obscure biological heterogeneity that diagnosis-agnostic machine learning could reveal. The demonstration that standard registry data alone can recover a five-phenotype taxonomy marks a shift away from predicting a single vague disease toward recognizing holistic, data-driven risk patterns in vulnerable populations.

Sources

1. Machine learning reveals hidden subtypes of intestinal injury in preterm infants — scienmag.com