Development and validation of interpretable machine learning models for dynamic prediction of prognosis in acute...

This study aims to develop and validate interpretable machine learning (ML) models to dynamically predict mortality risk among intensive care unit (ICU) patients diagnosed with acute pancreatitis complicated by acute kidney injury (AP-AKI).

The clinical data in the training set, including demographic characteristics, laboratory indicators, scoring systems, treatment modalities, and clinical management strategies, were obtained from three large-scale medical databases: the Medical Information Mart for Intensive Care, and the eICU Collaborative Research Database. The external validation set consisted of patients recruited from two independent hospitals. Predictive feature selection was conducted using univariate logistic regression, LASSO regularization, and multivariate logistic regression. Eleven machine learning (ML) algorithms-eXtreme Gradient Boosting (XGBoost), Logistic Regression (LR), Adaptive Boosting (AdaBoost), Decision Tree, Gaussian Naive Bayes (GNB), Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Bernoulli Naive Bayes (BernoulliNB), Linear Discriminant Analysis, LinearSVC, and Stochastic Gradient Descent (SGD)-were employed to develop predictive models. Finally, the SHapley Additive exPlanations (SHAP) method was applied to interpret the importance and directional effects of individual features.

Dynamic in-hospital mortality prediction was performed at 24 h, 48 h, and 7 days post-ICU admission, identifying nine to twelve variables respectively. The XGBoost model outperformed 10 other machine learning models, achieving training set AUROCs of 0.961 (95 % CI 0.95-0.97), 0.947 (95 % CI 0.94-0.96), and 0.968 (95 % CI 0.96-0.98) at these time points. The corresponding external validation results were 0.871 (95 % CI 0.79-0.95), 0.799 (95 % CI 0.66-0.94), and 0.667 (95 % CI 0.47-0.87). Regarding 90-day post-discharge mortality prediction, six variables were selected. The XGBoost model demonstrated superior performance, with a training set AUROC of 0.966 (95 % CI 0.96-0.97) and an external validation AUROC of 0.745 (95 % CI 0.61-0.88).

Web-based prognostic tools were developed to support clinical decision-making and optimize ICU bed resource management.