Machine Learning-assisted Prediction of Fluid Overload and Its Nursing Implications in Patients With Acute Pancreatitis

Fluid overload (FO) is a critical complication in acute pancreatitis (AP), contributing to prolonged hospital stays and increased mortality. Traditional scoring systems like BISAP lack predictive accuracy for FO, placing a burden on nurses managing fluid therapy. Machine learning (ML) offered potential for early FO prediction, yet its integration into nursing practice remains underexplored. Using the MIMIC-IV database, we analyzed 3458 adult AP patients, with 22.8% experiencing FO (net fluid balance >2 L/24h plus clinical signs). Three ML models-Random Forest (RF), XGBoost, and Long Short-Term Memory (LSTM)-were developed to predict FO using clinical and nursing-derived features (eg, BUN, net fluid balance, and documentation frequency). Models were evaluated for accuracy, precision, recall, F1 score, and ROC AUC, with SHAP analysis for feature importance. A Gradio-based clinical decision support system (CDSS) was prototyped for nursing integration. LSTM outperformed RF and XGBoost, achieving an AUC of 0.92 (95% CI [0.90, 0.94]), accuracy of 0.85, and F1 score of 0.82, significantly surpassing BISAP (AUC 0.74, P <.001). Key predictors included BUN, hematocrit, and net fluid balance. Subgroup analysis showed improved performance in patients >65 years (AUC 0.95) but reduced accuracy in CKD patients (AUC 0.88). The CDSS provided real-time FO risk scores, enhancing nurse decision-making. ML models, particularly LSTM, enabled accurate FO prediction in AP, with potential to transform nursing practice through proactive fluid management. Integration into CDSS supports early interventions, reducing nurse workload and improving patient outcomes.