AI and Multi-Omics Could Predict Necrotizing Enterocolitis Before Symptoms Strike

Necrotizing enterocolitis, or NEC, is one of the most feared diagnoses in any neonatal intensive care unit. The disease strikes predominantly in preterm infants, often beginning with subtle, easily missed warning signs before accelerating into intestinal necrosis, sepsis, multiple organ dysfunction and, in many cases, death. Even when infants survive, particularly those who require surgery, they may face persistent intestinal dysfunction, impaired growth and lasting neurodevelopmental problems. Despite decades of progress in biomarker discovery, intensive care and perioperative management, mortality and long-term morbidity remain stubbornly high. Now, an editorial published in the World Journal of Pediatrics argues that the field has been aiming at the wrong target: instead of trying to diagnose NEC earlier, clinicians and researchers should be trying to predict it before any injury has begun, using artificial intelligence to weave together layers of biological data that no single biomarker could ever capture.

The authors, a team led by Yan-Jun Wu and Hao Chen of Guangdong Provincial People’s Hospital together with Yong-Yan Shi of Shengjing Hospital of China Medical University, draw a sharp conceptual line between three distinct stages in the evolution of the disease. Preclinical prediction means estimating a future risk of NEC before any clinically apparent manifestation exists. Early detection means identifying a subclinical or already evolving disease process before a conventional diagnosis is established. Diagnosis itself means confirming the presence of NEC on the basis of current clinical, laboratory and imaging evidence. The distinction matters because much of what the literature celebrates as ‘prediction’ actually falls into the second category. Models that incorporate abdominal signs, inflammatory markers or imaging abnormalities can indeed help clinicians recognize NEC at an earlier stage, but those variables often reflect disease that is already underway. By the time such warning signs emerge, inflammatory and tissue-damaging processes may have been progressing for some time.

High discrimination in models built on near-diagnosis features, the authors caution, should therefore be interpreted carefully, since it may represent enhanced early detection of evolving injury rather than genuine prediction. A more appropriate framework comes from studies based on information available much earlier in life. One example is GutCheckNEC, a clinical risk index developed by Gephart and colleagues that estimates baseline susceptibility by integrating early factors such as gestational age, transfusion exposure, sepsis, feeding practices and the NEC burden of the individual unit. Another line of evidence comes from newborn-screen acylcarnitine studies, which suggest that metabolic vulnerability may be detectable shortly after birth and associated with later NEC development. These approaches are conceptually distinct from models driven by features emerging close to diagnosis. Truly preclinical prediction, the editorial argues, should rely on data available before clinical suspicion arises, define a clinically meaningful prediction window, and separate upstream risk signals from early manifestations of the disease itself.

The case for a fundamentally new approach rests on the biology of NEC, which is a dynamic, multilayered process in which intestinal immaturity, microbial dysbiosis, nutrient and energy metabolic disturbance, epithelial barrier disruption and inflammatory amplification interact over time. The long search for a single ideal biomarker has been central to NEC research, but the authors contend that this strategy is intrinsically limited. A single marker may reflect one dimension of vulnerability, such as microbial imbalance, metabolic stress, epithelial injury or host inflammation, but it is unlikely to capture the coordinated biological changes that precede clinically apparent intestinal injury. Recent studies illustrate both the promise and the fragmentation of the field: fecal proteomic signatures have indicated protein-level alterations before diagnosis, microbiota and volatile metabolite profiles have pointed to ecological disruption in the days preceding clinical onset, and other work has linked NEC risk to gut microbial shifts, tricarboxylic acid cycle metabolites, bile acid imbalance, blood proteomic alterations and fecal amino acid patterns.

Although each of these signals is biologically informative, they represent distinct layers of a complex disease process, and their predictive performance varies with sampling time, study population, outcome definition and analytic design. The central challenge, the authors argue, is not the absence of candidate biomarkers but the limited capacity of isolated signals to provide a stable, generalizable and clinically actionable representation of NEC risk. Their proposed solution is multi-omics integration: assessing susceptibility across biological layers simultaneously and linking those signals into a time-aware, biologically coherent risk framework. By integrating longitudinal multi-omics information, NEC prediction could move beyond isolated biomarker associations toward mechanism-informed risk stratification. Artificial intelligence, in this vision, is not a replacement for clinical or biological reasoning but a tool for integrating information that is difficult to handle with conventional statistical methods, particularly as bedside monitoring and omics technologies generate data that are highly dimensional, heterogeneous and time-varying.

Technically, the editorial outlines several ways heterogeneous datasets can be combined during model development. Feature-level integration merges appropriately normalized clinical and omics variables within a single model. Representation-level approaches transform each modality into lower-dimensional representations before joint analysis, while decision-level integration combines predictions from separate modality-specific models. For NEC, temporal differences across modalities must also be considered, because microbiome, metabolomic, proteomic and host-response signals may be collected at different points along the disease trajectory. Conventional statistical methods remain useful for testing predefined associations but are often limited in capturing nonlinear relationships, interactions across biological layers and longitudinal changes. AI-guided algorithms have already been applied to related tasks in NEC research, including combining abdominal radiographs with clinical, laboratory or physiologic data such as vital signs and near-infrared spectroscopy-derived oxygenation measurements to support diagnosis, and using early perinatal variables for surgical NEC risk assessment. Explainable AI work has highlighted features such as bowel peristalsis, C-reactive protein, albumin, bowel thickness and procalcitonin.

The authors are careful to temper expectations. The value of AI should not be reduced to incremental improvements in predictive performance, they write, and AI should be viewed as a tool for organizing and interpreting complex biological data rather than overcoming underlying biological uncertainty. Its usefulness depends on the quality, timing, interpretability and biological relevance of the input data. This leads to their call for mechanism-driven prediction. Current models often rely on clinical risk factors such as gestational age, birth weight, feeding exposure, antibiotic use and conventional inflammatory markers. These variables help identify infants with baseline vulnerability, but they do not fully explain why some high-risk infants progress to NEC while others with similar clinical profiles do not. A mechanism-driven model would integrate clinical information with microbiome, metabolomic, proteomic and host-response data to construct a dynamic and biologically interpretable risk assessment system, clarifying not only which infants are at risk but which disease pathways may be driving that risk.

Emerging omics studies provide a foundation for this shift. Plasma proteomics has revealed NEC-related changes in biosynthetic pathways, extracellular matrix remodeling, immune responses and glycosaminoglycan levels, while aptamer-based serum proteomics has identified candidate proteins with discriminatory potential. Metabolomic studies have reported alterations in bile acid profiles, tricarboxylic acid cycle metabolites and acylcarnitine patterns, suggesting that NEC risk reflects converging disturbances in microbial ecology, host metabolism, immune activation and intestinal integrity. The temporal dimension is equally central. Fecal proteomic signatures have been detected before NEC diagnosis, urinary intestinal fatty acid-binding protein has been reported to rise within days before clinical onset, and superior mesenteric artery Doppler findings obtained before feeding exposure have shown potential for identifying infants at increased risk. Microbial shifts, including changes in the ratio of Bacteroidetes to Firmicutes before NEC onset, further suggest that ecological disturbance may precede overt disease.

What would this look like at the bedside? A clinically useful model, the authors argue, should do more than assign a static probability of disease. It should identify biological susceptibility, explain why risk is increasing, and determine when an infant is entering an actionable high-risk state. A rising risk trajectory could prompt closer bedside monitoring and reassessment of feeding tolerance, lead clinicians to review modifiable exposures such as feeding strategies and unnecessary antibiotic use, and trigger prompt laboratory and imaging evaluations if risk continues to climb. It could also support the timely use of established preventive strategies, including human milk and other evidence-based interventions. In this way, prediction would shift from single-time-point risk estimation toward continuous risk monitoring, and from empirical classification toward biologically interpretable prediction of the preclinical transition.

Substantial hurdles remain before any of this reaches routine practice. Current studies are limited by retrospective designs, missing data, small sample sizes, heterogeneous NEC definitions and potential selection bias, and even studies with stronger preclinical sampling strategies require validation in larger multicenter cohorts because NEC events are relatively rare. Multi-omics data are vulnerable to sample handling effects, platform and batch effects, temporal biological variation and high dimensionality, all of which can undermine reproducibility and complicate standardization. The authors call for prospective, multicenter cohorts with standardized sampling, harmonized NEC definitions, prespecified prediction windows and clearly defined endpoints, with models specifying when each variable becomes available and whether external validation and calibration are performed. The goal of next-generation NEC prediction, they conclude, should not be the discovery of more risk factors but the identification of the mechanisms that drive disease onset and progression, with AI serving to transform multi-omics data into interpretable, actionable and clinically trustworthy decision support for the most fragile infants.

Subject of Research: AI-enabled multi-omics prediction of necrotizing enterocolitis in preterm infants

Article Title: Beyond early diagnosis: toward AI-enabled multi-omics prediction of necrotizing enterocolitis

Article References: Wu, Y.-J., Lin, T., Meng, M.-J., Zhang, X.-T., Zhang, K.-X., Zhuo, Z.-W., Sha, W.-H., Shi, Y.-Y., & Chen, H. (2026). Beyond early diagnosis: toward AI-enabled multi-omics prediction of necrotizing enterocolitis. World Journal of Pediatrics. https://doi.org/10.1007/s12519-026-01112-8

Image Credits: AI Generated

DOI: 10.1007/s12519-026-01112-8

Keywords: necrotizing enterocolitis, preterm infants, artificial intelligence, multi-omics, biomarkers, neonatology, microbiome, metabolomics, proteomics, predictive medicine, machine learning, neonatal intensive care