S91 Machine Learning and Deep Learning Models for Predicting Acute Kidney Injury in Acute Pancreatitis: A Systematic Review and Meta-Analysis

Introduction: Acute kidney injury (AKI) is a severe complication of acute pancreatitis (AP), with mortality reaching up to 75%. Traditional scoring systems often underperform due to AP’s heterogeneity and AKI’s multifactorial nature. Artificial intelligence methods, including machine learning (ML) and deep learning (DL), show promise for early AKI prediction by capturing non-linear clinical patterns; however, their application in AP remains limited. This systematic review and meta-analysis assesses current artificial intelligence-based models for AKI prediction in AP, identifies gaps in validation and implementation, and guides future research and clinical integration. Methods: We searched PubMed and Google Scholar through May 2025, supplemented by snowballing, to identify studies that developed or validated ML/DL models for predicting AKI in AP using area under the curve (AUC) as the performance metric. Seven retrospective studies using clinical or public datasets (e.g., MIMIC) were included. Models were categorised as: (1) linear (e.g., LR, LASSO), (2) tree-based/ensemble (e.g., RF, XGBoost), (3) neural/deep learning (e.g., ANN, DL), and (4) kernel/probabilistic (e.g., SVM). Extracted data included predictors, validation methods, and AUCs. Pooled AUCs were estimated using a random-effects model; heterogeneity was assessed via I², and robustness was evaluated with leave-one-out analysis. Results: Across 7 studies (n = 4,549), ML/DL models outperformed logistic regression in predicting AKI in AP, with pooled AUCs of 0.83 (95% CI: 0.82–0.85; I² = 65.9%) vs 0.80 (95% CI: 0.78–0.83; I² = 54.5%). Tree-based models achieved the highest subgroup performance (AUC = 0.86), followed by kernel/probabilistic (0.84) and neural/deep learning models (0.82). Forest plots demonstrated consistency with low heterogeneity, and leave-one-out analysis confirmed robustness (AUC range: 0.82–0.85). Common predictors included creatinine, CRP, IL-6, SOFA, APACHE II scores, age, and fluid balance. Few studies incorporated external validation or interpretability tools. Conclusion: ML/DL models, particularly tree-based models and SVM, surpass LR in predicting AKI in AP, with a pooled AUC of 0.83, by effectively handling complex data for early risk stratification. However, limited interpretability (e.g., lacking SHAP), inconsistent predictors, and scarce external validation hinder clinical adoption. Future research must enhance model transparency, validation, and clinical integration to improve early risk prediction and patient outcomes.