697 - Clustering of VLBW Infants Reveals Distinct Clinical Patterns Associated with Intestinal Injury Pathophysiology

Publication Number: 1674.697

Yong-Sung Choi, Kyung Hee Unsiversity Hospital, Seoul, Seoul-t'ukpyolsi, Republic of Korea; Jehyeuk Ahn, Kyung Hee University, Yongin-si, Kyonggi-do, Republic of Korea; Sung-Hoon Chung, Kyung Hee University College of Medicine, Seoul, Seoul-t'ukpyolsi, Republic of Korea; Seong Hee Oh, University of Ulsan College of Medicine, Gangneung Asan Hospital, Gangneung, Kangwon-do, Republic of Korea; Ji Yoo Kim, Kyung Hee university medical center, Dondaemungu, Seoul-t'ukpyolsi, Republic of Korea; Jinseok Lee, Kyung Hee University, Yongin, Kyonggi-do, Republic of Korea; Nima Aghaeepour, Stanford University School of Medicine, Palo Alto, CA, United States; David K. Stevenson, Stanford University School of Medicine, Stanford, CA, United States; Josef Neu, University of Florida, Gainesville, FL, United States

Background: Very low birth weight infants (VLBWIs) exhibit marked heterogeneity in necrotizing enterocolitis (NEC) risk and mortality, poorly captured by conventional factors.

Objective: To discover data-driven clinical subgroups via unsupervised hierarchical clustering (blinded to NEC, gestational age, birth weight, death) and quantitatively validate prognostic utility using supervised survival models.

Design/Methods: From the Korean Neonatal Network (2013-2024, n=20,352), 78 variables underwent principal coordinate analysis (PCoA; 70% variance: 12 PCs). Hierarchical clustering generated 8 clusters (C1-C8); C8 (n=2,778) was subclustered into 27 groups (D1-D27). PC loading (|≥0.15|) defined features. Kaplan-Meier (KM) and Random Survival Forest (RSF) assessed survival.

Results: Among C1-C8 (n=353/ 3,701/ 1,812/ 3,992/ 1,454/ 4,726/1,536/2,778), C8 showed the highest NEC (17.85% vs. 6.2% overall). Subclusters D23-D27 (n=202/1,156/760/426/234) had features based on the PC loading values of importance: D23 (antenatal antibiotics/bronchopulmonary dysplasia), D24 (multiple gestation/IVF), D25 (maternal hypertension/small-for-gestational-age), D26 (chorioamnionitis/preterm premature rupture of membranes/vaginal delivery), D27 (outborn/intraventricular hemorrhage ≥grade 2/seizure/air-leak/pulmonary hemorrhage). The NEC incidence of D23-D27 were 18.81, 15.92, 17.37, 19.01, and 26.07% respectively. RSF using cluster labels predicted C8 lowest on test set (n=4,071). Among D23-D27, KM revealed extreme divergence (p=0.001). RSF exactly replicated KM ranking (Spearman ρ=1.000), with restricted mean survival time (RMST) mean absolute error (MAE) 1.07 days (1.88%) and normalized absolute Brier curve (nABC) 0.025(mean).

Conclusion(s): Unsupervised hierarchical subclustering identifies NEC-enriched (D27) and clinically distinct patterns (D23-D27), quantitatively validated (ρ=1.000) by supervised models, enabling precision risk stratification in VLBWIs.

Figure 1. Principle component analysis plotting of 8 clusters in 2D.
Unsupervised machine learning using hierarchical clustering method generated 8 distinct clusters according to 6 PCoA axes.

Figure 2. Dendrogram of the hierarchical clustering.
Eight (C1-8) clusters structure was further categorized into 27 (D1-27) subclusters.

Figure 3. Validation for the unsupervised machine learning by supervised learning using survival curve.
A is the Kaplan-Meyer's survival curve of the unsupervised clusters D23-27. B is random survival forest(RSF) prediction curve which matches to each other.