Neonatal necrotizing enterocolitis (NEC) remains a devastating gastrointestinal emergency in neonates, particularly affecting preterm infants with birth weights below 1,500 g (1, 2). Current diagnostic reliance on Bell staging criteria and abdominal radiography faces critical limitations, including delayed detection of early pathophysiological changes (e.g., mucosal ischemia and bacterial translocation) (3, 4). x-ray is an important diagnostic modality for NEC, but its widespread use is limited by radiation concerns and the inconvenience of requiring transport to a radiology suite, rather than being readily available at the bedside. This diagnostic latency contributes to the persistent 20%–30% mortality rate despite advances in neonatal intensive care (5, 6).

Recent advancements in ultrasonography have demonstrated superior sensitivity in detecting preclinical NEC manifestations. High-resolution ultrasound can quantify bowel wall thickness (BWT) variations (>2.0 mm predictive of necrosis) and monitor mesenteric blood flow dynamics through Doppler indices (7, 8). Abdominal radiography remains a common initial imaging modality for evaluating neonatal abdominal pathology. However, as demonstrated by Silva et al. (2013), a normal radiographic gas pattern does not exclude the presence of significant intestinal abnormalities detectable by ultrasound, highlighting the potential for missed diagnoses when relying solely on radiography (9). Parallel developments in serum biomarkers, including PT, INR, APTT (10), C-reactive protein (11) and other serological have shown potential for risk stratification, Sharif demonstrated that low serum albumin (SA) concentration (≤20 g/L) on day 2 of NEC diagnosis is a significant predictor of surgical intervention in neonates with Bell's stage 2 NEC. This finding suggests that SA, in conjunction with other clinical and serological markers, may be a useful tool for identifying patients at higher risk of requiring surgery (12). Their findings suggest that monitoring coagulation parameters can aid in early identification of high-risk NEC neonates, potentially optimizing treatment strategies and improving outcomes (13, 14). While ultrasonography provides real-time visualization of intestinal dynamics, its diagnostic accuracy may be compromised by acoustic shadowing from bony structures and operator-dependent expertise, potentially leading to misinterpretation of early NEC signs. Conversely, serological biomarkers, though objective in quantification, exhibit significant interindividual variability due to fluctuations in host immune status and inflammatory cascades. These inherent limitations of standalone modalities underscore the suboptimal predictive performance when employing either approach in isolation. Emerging evidence suggests that integrating both modalities through machine learning algorithms may harness their synergistic diagnostic potential, thereby improving sensitivity for preclinical NEC detection and risk stratification.

Machine learning (ML) presents transformative opportunities for NEC prediction through multimodal data fusion. Leiva et al. (2023) provide a comprehensive overview of the use of machine learning in NEC biomarker discovery, while also acknowledging the challenges inherent in the field. They highlight the potential of machine learning to integrate multi-omics data with clinical features, phenotypes of progression, and predicted therapeutic targets, resulting in clinically meaningful information. This approach could lead to earlier diagnosis, more targeted therapies, and improved outcomes for infants with NEC (15). Contemporary studies further highlight ML's capacity to decode nonlinear interactions between temporal ultrasound features and biochemical trajectories (16). Our study innovatively expands this paradigm by systematically evaluating 12 ML algorithms on hybrid ultrasound-serological datasets, addressing critical gaps in neonatal predictive modeling.

This study demonstrates that an XGBoost-based fusion model incorporating ultrasound and serological markers (USPN model) significantly improves the prediction of NEC in neonates compared to models relying solely on ultrasound or serological data. This finding highlights the potential of integrating multimodal data and machine learning to enhance diagnostic accuracy and inform clinical decision-making in this vulnerable population. The superior performance of the fusion model suggests that NEC is a complex disease process that is best characterized by a combination of imaging and biomarker data.

Our results build upon previous research that has explored the use of ultrasound or serological markers for NEC prediction. For example, Wang et al. retrospectively analyzed 144 neonates with suspected or confirmed NEC, comparing abdominal ultrasound and plain x-rays for diagnostic accuracy and prognostication. Their study found that ultrasound was superior to x-ray in detecting portal venous gas and intestinal dilatation in confirmed NEC cases. Furthermore, ultrasound findings of intestinal dilatation, bowel wall thickening, and ascites were significantly associated with the need for surgery or death, suggesting their potential utility in predicting disease severity. While this study highlights the value of abdominal ultrasound, it did not incorporate serological markers, limiting its ability to leverage the combined predictive power of both modalities, as explored in our current research (17). In contrast to studies focusing solely on imaging, Garg et al. (18) investigated the clinical impact of NEC-associated sepsis and its relationship to inflammatory markers. They found that infants with NEC-associated sepsis had significantly higher CRP levels and lower platelet counts at NEC onset and 24 h after onset compared to those without sepsis. While Garg et al. primarily focused on the consequences of NEC-associated sepsis and its impact on outcomes like length of stay and mortality, our study aims to leverage serological markers, alongside ultrasound findings, for early prediction of NEC development, whereas our study demonstrates the synergistic effect of combining these modalities within a machine learning framework.

Several studies have also explored the use of machine learning for NEC prediction. Leiva et al. (15) systematically reviewed the potential of multi-omics (genomics, proteomics) combined with machine learning to identify NEC biomarkers, highlighting its ability to decode disease subtypes and therapeutic targets through heterogeneous data integration. However, their analysis revealed critical limitations in existing approaches: (1) reliance on single-omics data that may not fully capture the complexity of NEC; (2) small sample sizes in many studies, potentially leading to overfitting risks; (3) exclusion of imaging modalities like ultrasound, which limits early diagnostic utility. In contrast, our study advances the field by addressing these gaps. We rigorously compared 12 machine learning algorithms to optimize model generalizability, ultimately selecting XGBoost model with SHAP-based interpretability. Crucially, we integrated ultrasonographic markers with serological markers. This multimodal strategy not only improves sensitivity for early NEC but also addresses the “black-box”. By bridging imaging biomarkers with host response dynamics, we provide a clinically actionable tool that aligns with Leiva's call for “phenotype-aware AI models” while overcoming the translational barriers of pure omics approaches.

Although the predictive variables were collected within 24 h before the formal diagnosis of NEC, they reflect routine monitoring performed during the early phase of clinical suspicion, prior to overt disease recognition or intervention. The model was specifically designed to operate at this early stage, utilizing parameters triggered by subtle signs rather than clear NEC manifestations.

Elevated CRP and procalcitonin levels, indicative of systemic inflammation, were strong predictors of NEC in our study, consistent with the established role of inflammation in the pathogenesis of the disease. Zeng's (19). study demonstrate the similar result. Gaudin et al. (20) further emphasized the prognostic value of CRP, demonstrating a correlation between elevated CRP levels and the risk of post-NEC intestinal stricture. our findings, combined with Gaudin et al.'s work, highlight the continued clinical relevance of CRP, particularly in assessing disease severity and predicting long-term complications. The readily availability and widespread use of CRP and procalcitonin assays make it a valuable tool in the initial assessment of NEC risk, Lee's (21) findings were similar with ours.

Our study identified hypoalbuminemia as a significant risk factor for the development of NEC, consistent with the findings of Mohd Amin et al., (22) who demonstrated that a low albumin level, particularly when combined with elevated CRP (CRP/ALB ratio ≥ 3), is strongly associated with poor outcomes, including the need for surgical intervention and mortality in neonates with NEC. Hypoalbuminemia may reflect systemic inflammation, malnutrition, or increased vascular permeability, all of which are implicated in the pathogenesis of NEC. The liver's reduced capacity to synthesize albumin in preterm infants, compounded by the inflammatory response, may further exacerbate this condition. The CRP/ALB ratio, as highlighted by Mohd Amin et al., serves as a valuable prognostic tool, integrating both inflammatory and nutritional status, which aligns with our findings that hypoalbuminemia independently predicts NEC risk.

Reduced bowel peristalsis and increased bowel wall thickness, as assessed by ultrasound, reflect intestinal ischemia and inflammation, key features of NEC, Chen demonstrated that in patients with reduced or absent intestinal peristalsis, the incidence of NEC is ten times higher than in those with normal peristalsis (23). This finding is consistent with our study, which also identified reduced bowel peristalsis as an independent risk factor for NEC. Esposito et al., (24) in their comprehensive review of NEC imaging, further support the significance of these ultrasound findings. They highlight that in the early stages of NEC, when x-ray findings may be non-specific, ultrasound can reveal direct signs such as bowel wall thickening (generally considered pathological when exceeding 2.6 mm) and abnormal bowel wall echoic patterns, reflecting the loss of normal wall layering due to inflammation and edema. Our study's identification of increased bowel wall thickness as an independent risk factor aligns with this observation, reinforcing the value of ultrasound in detecting early intestinal changes indicative of NEC. Priyadarshi's findings are also similar to ours (25).

While these ultrasound-based features are strongly associated with NEC risk, analysis of misclassified cases revealed that their predictive performance may vary depending on the broader clinical context. In certain cases, reduced bowel peristalsis alone contributed disproportionately to high-risk predictions, even when inflammatory markers such as CRP remained low and bowel wall thickness was within normal range. These false positive predictions often occurred in neonates with transient ileus or non-NEC-related sepsis, suggesting that peristalsis, though sensitive, may lack specificity when interpreted in isolation.

Conversely, false negative predictions were more frequently observed in extremely preterm infants with early-stage or atypical NEC. In these cases, CRP and procalcitonin levels were often within normal limits, and bowel ultrasound findings were subtle or absent. Such presentations, while clinically recognized, may escape detection in models that rely solely on static, single-timepoint data.

These findings underscore the need to incorporate additional contextual and temporal information into future model iterations. Potential strategies include the use of time-series trends in inflammatory biomarkers, integration of gestational age and birth weight, and inclusion of comorbid conditions such as sepsis or hemodynamic instability. Furthermore, modeling feature interactions—such as interpreting reduced peristalsis as high risk only in the presence of elevated CRP—may enhance model specificity and reduce misclassification in borderline clinical scenarios.

In our cohort, only 13 neonates (6.8%) had a gestational age <32 weeks, which limited the feasibility of performing gestational-age–stratified analysis of ultrasound findings. Future multicenter studies with larger and more balanced cohorts are needed to investigate how NEC presentation varies with gestational maturity, as highlighted by Battersby et al. (26).

This study has several notable limitations. First, the retrospective design introduces potential selection and information biases, as data were extracted from existing medical records rather than prospectively collected. Although we implemented strict inclusion and exclusion criteria, residual confounding may still exist. Second, this was a single-center study conducted in a tertiary NICU, which may limit the diversity of patient populations, clinical practices, and imaging protocols—factors that can affect model generalizability. Third, the relatively small sample size (n = 191), while sufficient for initial model development and internal validation, increases the risk of overfitting and may not fully capture the heterogeneity of NEC presentations.

In fact, the discrepancy between the training AUC (0.97) and validation AUC (0.88) may reflect some degree of overfitting, despite the implementation of mitigation strategies such as 10-fold cross-validation, hyperparameter tuning via grid search, and SHAP-based feature selection. These techniques helped reduce dimensionality and overfitting risk, but further refinement is still necessary.

To improve the robustness and external applicability of the USPN model, future studies should focus on prospective validation in larger, multicenter cohorts. This would allow model calibration across varying institutions and patient subgroups, enhancing its clinical utility. In addition, the incorporation of further regularization techniques (e.g., L1/L2 penalty), model simplification, or ensemble averaging may help improve performance stability and reduce the risk of overfitting in future implementations.

In conclusion, our study demonstrates that an XGBoost-based fusion model incorporating ultrasound and serological markers significantly improves the prediction of NEC in neonates. This finding has important clinical implications and highlights the potential of integrating multimodal data and machine learning to enhance diagnostic accuracy and inform clinical decision-making in this vulnerable population. Future research should focus on validating our findings in larger multi-center studies and exploring the use of the model to guide clinical decision-making and to personalize treatment strategies.