A NEW explainable machine learning model has shown superior accuracy in predicting mortality risk for acute pancreatitis (AP) patients admitted to intensive care units (ICUs) compared to traditional clinical scoring systems. The study introduces an XGBoost-based model that not only provides more precise predictions but also explains its decision-making process, potentially transforming how AP patients are managed in critical care settings.
Current clinical risk scoring systems, such as the APACHE IV, SOFA, and BISAP scores, have limitations in accurately assessing mortality risk for AP patients in ICUs. In response, researchers developed a gradient-boosting machine learning model (XGBoost) and validated it using data from two public databases: the Medical Information Mart for Intensive Care (MIMIC) for training and the eICU Collaborative Research Database (eICU-CRD) for validation. The study aimed to improve upon existing methods by offering both higher accuracy and explainability in predictions.
The XGBoost model outperformed traditional scoring systems significantly, achieving an area under the receiver operating characteristic curve (AUC) of 0.89 (95% CI: 0.84–0.94). This indicates a high level of accuracy in distinguishing between patients who were at risk of death and those who were not. In comparison, the APACHE IV, SOFA, and BISAP scores demonstrated lower accuracy. Notably, when the model was set to 100% sensitivity for predicting death, it achieved a specificity of 38%, compared to just 1% for APACHE IV and BISAP scores, and 16% for the SOFA score.
A key innovation of this model is its use of SHapley Additive exPlanations (SHAP), a method that makes the decision-making process of machine learning algorithms more transparent. SHAP identifies and ranks the factors that most influence each individual prediction, helping physicians understand why a patient is classified as high-risk. This level of transparency can build greater trust in AI tools among healthcare professionals and facilitate more informed decision-making.
The authors suggest that this model could be particularly valuable in identifying patients with very low risk who could be safely managed in general wards rather than the ICU, potentially reducing healthcare costs and resource strain.
Aleksandra Zurowska, EMJ
Reference
Jiang M et al. Explainable machine learning model for predicting acute pancreatitis mortality in the intensive care unit. BMC Gastroenterol. 2025;DOI: 10.1186/s12876-025-03723-3.
