Refeeding syndrome (RFS) is a severe metabolic complication characterized by profound electrolyte disturbances, fluid imbalance, and organ dysfunction. It occurs when nutritional support is restarted in patients with prolonged fasting or severe malnutrition, driven by core pathophysiological changes that result from rapid insulin secretion recovery and subsequent intracellular electrolyte shifts (1). In critically ill patients, RFS is closely associated with increased mortality and prolonged hospital stays (2). Although consensus definitions and high-risk population criteria have been proposed by institutions such as the American Society for Parenteral and Enteral Nutrition (ASPEN) (3, 4), early identification remains challenging in clinically complex populations.

Patients with severe acute pancreatitis (SAP) are at extremely high risk of RFS. On one hand, SAP-induced systemic inflammatory response, hypermetabolic consumption, and gastrointestinal dysfunction rapidly lead to malnutrition (5, 6). On the other hand, early enteral nutrition—the cornerstone of SAP treatment—serves as a classic trigger for RFS (7). However, existing RFS prediction models primarily focus on general intensive care unit (ICU) or hospital-wide populations (8–10). Their predictive variables (e.g., specific urinary biomarkers) or targets (e.g., isolated hypophosphatemia) lack applicability and specificity in the unique pathophysiological context of SAP. Thus, developing an interpretable RFS risk prediction tool tailored to SAP patients, based on routine clinical data, is crucial for precise early warning and personalized nutrition management.

Recently, machine learning (ML) has demonstrated significant advantages in medical prediction models. It captures complex non-linear relationships between variables, potentially outperforming traditional logistic regression (LR) (11, 12). Building on this, we systematically collected clinical data from SAP patients, constructed RFS risk prediction models using multiple ML algorithms, and enhanced model interpretability via SHapley Additive exPlanations (SHAP) analysis. The present study aims to develop a dedicated RFS risk early warning tool for SAP patients, with a core focus on identifying high-risk individuals through routine clinical indicators before the initiation of nutritional support. It targets the state where RFS risk exists without overt electrolyte abnormalities, rather than diagnosing existing electrolyte disturbances or confirmed RFS, ultimately providing data support for the clinical development of individualized monitoring and preventive intervention plans.

This study constructed and validated RFS risk prediction models for SAP patients using six ML algorithms based on single-center retrospective data. The GBM model showed good predictive performance and generalization ability, with seven easily accessible clinical features identified via LASSO regression and SHAP analysis. This provides a reliable tool for early RFS identification in SAP patients.

The seven identified predictors have clear clinical and pathophysiological significance. Electrolyte disturbances are central to RFS (19), and our model highlights high pre-refeeding serum potassium/sodium and low serum calcium as key warning signs. In SAP patients, elevated pre-refeeding serum potassium/sodium may reflect electrolyte distribution abnormalities or impaired excretion due to stress, dehydration, or underlying renal dysfunction (20). This “apparently normal or elevated” state often masks true electrolyte depletion, leading clinicians to overlook potential risks without targeted supplementation during enteral nutrition initiation. Once enteral nutrition starts, restored insulin secretion drives rapid intracellular shifts of potassium, phosphorus, and magnesium, causing acute serum depletion and typical RFS manifestations (21). Hypocalcemia is common in SAP, associated with peripancreatic fat necrosis (calcium soap formation), hypomagnesemia, and vitamin D metabolism disorders (22); refeeding further exacerbates calcium-phosphorus imbalance (23), increasing neuromuscular instability. Beyond electrolytes, DM history and diuretic use directly elevate metabolic imbalance risk by impairing insulin sensitivity and promoting renal electrolyte loss. Elevated BUN indicates dehydration or hypercatabolism, a marker of systemic metabolic disturbance (24). Finally, gastrointestinal decompression—an important procedure-related factor—signals severe gastrointestinal dysfunction and exacerbates electrolyte loss via digestive fluid depletion (25), compounding metabolic challenges during refeeding.

It is important to note that the core function of this model lies in risk early warning and stratification, and its prediction logic and clinical application require an in-depth understanding through the following three aspects. First, SHAP analysis (Figure 6B) reveals that key predictors such as serum potassium, serum sodium, serum calcium, and blood urea nitrogen affect RFS risk mainly through their relative fluctuations within the normal range. For SAP patients with laboratory indicators within the normal range, serum potassium or sodium close to the upper limit, or serum calcium tending to the lower limit, may indicate a pseudo-normal state of intracellular electrolyte depletion or potential metabolic disorders. The model can sensitively capture such subtle changes, assisting clinicians to strengthen monitoring or implement preventive interventions before initiating enteral nutrition. Second, diuretic use not only reflects the volume management needs or comorbidities (such as heart failure) of SAP patients but may also cause electrolyte loss through promoting renal excretion, thereby increasing the risk of RFS. Through feature integration in the multivariate machine learning model, this study has controlled its confounding effect at the statistical level. However, due to the retrospective nature of the study, details such as the specific types, dosages, and clear indications of diuretics could not be collected. When applying this model clinically, the contribution of this predictor needs to be comprehensively interpreted in combination with the patient's volume status, comorbidities, and medication background. Additionally, NRS scores assessed at admission were generally low in this study. The core reason is that the score is designed around chronic nutritional depletion indicators, while the nutritional risk of SAP patients stems from acute stress response, gastrointestinal dysfunction, and hypermetabolic state. Such acute risks are difficult to capture by traditional scoring tools relying on chronic indicators. This phenomenon further confirms the limited applicability of traditional nutritional risk scores in SAP populations, highlighting the need for developing disease-specific RFS risk prediction models.

Compared with previous studies, this is the first to focus on SAP patients—an extremely high-risk RFS population—and systematically compare multiple ML algorithms for model construction. For example, Zhan et al. (8) developed a well-discriminating model for general ICU patients, but included non-routinely monitored indicators (e.g., urinary epithelial cell count) with questionable applicability in SAP. Choi et al. (9) used XGBoost to predict refeeding hypophosphatemia, but the model targeted only one RFS component and included heterogeneous study populations, failing to reflect SAP-specific metabolic and inflammatory features. Zhang et al. (10) constructed an RFS prediction model for neurocritical patients, identifying risk factors such as drinking history and fasting duration, but the pathophysiological core (consciousness disorder-related nutritional intake impairment) differs fundamentally from SAP's pancreatitis-mediated digestive failure and hypercatabolism. Additionally, their model used traditional logistic regression, failing to capture complex non-linear relationships or quantify feature contributions via interpretability algorithms, limiting precision and acceptance in clinical translation. Our study fills the gap of SAP-specific RFS prediction models, with LASSO-selected predictors (e.g., electrolytes, gastrointestinal decompression) closely aligned with SAP clinical course, ensuring biological rationality and clinical relevance. Furthermore, SHAP analysis enhances the interpretability of the complex ML model, translating predictions into intuitively understandable clinical decision support. All predictors are derived from routine clinical data without requiring additional tests, ensuring excellent clinical practicality and wide applicability.

Nevertheless, this study has several limitations. First, the single-center, retrospective design and limited sample size may introduce selection bias. Center-specific practices such as enteral nutrition protocols and electrolyte monitoring strategies may affect model performance and generalizability. Future plans include multicenter external validation and single-center prospective validation, enrolling SAP patient cohorts from centers in different regions and with varying medical levels. The Bootstrap method will be used to calculate optimism-adjusted performance metrics for more rigorous evaluation of the model's generalization ability and clinical applicability. Second, this study focuses on SAP patients, so the model's extrapolation to mild and moderately severe acute pancreatitis patients requires caution; the SAP cohort was not further stratified by severity indicators such as Marshall score and duration of organ failure, and residual disease heterogeneity may impact model performance. Future studies will incorporate these severity indicators to optimize the model and expand its application scope through multicenter research. Third, due to the retrospective design, clinical data were not fully retrievable. Specific types and indications of diuretics were not collected, nor were details of nutritional support such as infusion rate and caloric targets—key modifiable factors in RFS management (26). Fourth, to improve the specificity of RFS diagnosis, this study adopted a stricter definition than the 2020 ASPEN criteria. While this helps reduce the inclusion of non-specific cases, it may exclude some early or mild RFS cases, potentially affecting the model's ability to identify the full spectrum of RFS risks. Future research can validate the model's performance under a broader definition.

In summary, this study developed and validated an explainable ML prediction model for RFS in SAP patients. The GBM model demonstrates satisfactory predictive performance and generalization ability, requiring only seven easily accessible clinical features. This tool enables early identification of high-risk SAP patients before nutritional support initiation, facilitating timely interventions and optimized nutrition management to improve patient outcomes.