Transcriptomics-guided serum podoplanin and interpretable machine learning for early risk stratification of necrotizing enterocolitis in preterm infants.

BACKGROUND

To evaluate podoplanin (PDPN) as a circulating biomarker for early necrotizing enterocolitis (NEC) risk stratification and to determine whether an interpretable machine-learning framework integrating PDPN with prediagnostic clinical and inflammatory variables could improve individualized risk estimation.

METHODS

PDPN expression was examined in two public NEC intestinal transcriptomic datasets. Serum PDPN was then measured by enzyme-linked immunosorbent assay in a nested case-control cohort of 132 preterm infants (66 with NEC and 66 gestational-age- and birth-weight-matched controls). All candidate predictors were obtained before NEC onset in cases and before the matched reference time in controls. Restricted cubic spline, receiver operating characteristic, precision-recall, LASSO, multivariable logistic regression, 11 machine-learning algorithms, SHAP, and nomogram analyses were performed.

RESULTS

PDPN was upregulated in NEC intestinal tissue. Serum concentrations were higher in infants who developed NEC than in controls (48.36 ± 22.89 vs 19.75 ± 6.26 ng/mL, P < 0.001) and in Bell stage III than in Bell stage II NEC [69.43 (58.91-77.95) vs 30.59 (24.06-42.79) ng/mL, P < 0.001]. PDPN correlated weakly with the C-reactive protein-to-albumin ratio and high-sensitivity C-reactive protein, achieved an area under the receiver operating characteristic curve of 0.905 and an area under the precision-recall curve of 0.925, and remained independently associated with NEC after multivariable adjustment. CatBoost showed the most favorable internal performance among 11 models, with PDPN as the leading SHAP contributor. The nomogram achieved an internally assessed area under the curve of 0.974.

CONCLUSIONS

Prediagnostic serum PDPN, particularly when combined with routinely available variables, may support early individualized NEC risk stratification. Prospective multicenter validation and mechanistic studies are warranted.