Mortality Prediction Model for General Surgeries Using a Small Data Set with Explainable Artificial Intelligence

Abstract

This study introduces a mortality prediction model for general surgeries, enhanced by Explainable Artificial Intelligence (XAI) to help surgeons anticipate postoperative outcomes and identify high-risk patients. Using a deep learning approach on a limited dataset, synthetic data generation via a variational autoencoder (VAE) was employed to simulate real-world accuracy. The deep learning model trained with VAE-augmented data emerged as the best performer in comparison to other machine learning models, including RF, KNN, extreme gradient boosting, support vector machines, and logistic regression., achieving the highest F1 score and balanced precision and recall.
Patient data—including morbidities, laboratory results, and postoperative complications—was processed through various models and evaluated on accuracy, F1 score, and AUROC metrics. The VAE data augmentation improved the performance of most models, especially complex ones such as decision trees, random forests, gradient boosting, and XGBoost. However, simpler models like logistic regression and support vector machines (SVC) struggled with VAE-augmented data. The ensembling approach incorporating Ensembles of VAE, Flipout in last layer, Flipout in all layers and Bayesian model was used to improve prediction, The ensemble model, VAE-Flipout Last Layer and Flipout All Layers Ensemble, demonstrated enhanced predictive accuracy and reliability surpassing other ensemble models . Calibration techniques, including Temperature Scaling, Platt Scaling, and Isotonic Regression, were applied to ensure robust probabilistic outputs. The VAE-Flipout Last Layer and Flipout All Layers Ensemble achieved an F1 score of 0.77, a Brier score of 0.0254, and a ROC-AUC of 0.94.
In terms of model explainability, LIME and SHAP identified the features influencing mortality, including Sepsis, Postoperative Urea, Small Bowel Resection, Omentoplasty ASA Classification, Chronic Liver Disease, and postoperative biomarkers like SGPT and bilirubin. LIME provided local insights tailored to individual predictions, while SHAP revealed a global perspective across all instances, consistently highlighting these key features. This consistency reinforces the relevance of identified factors in patient outcomes.
Future research should focus on expanding the dataset through advanced augmentation techniques, like GANs, and on refining calibration metrics, such as Expected Calibration Error (ECE) and Maximum Calibration Error (MCE). This study offers a valuable AI-driven tool to improve patient prognosis and postoperative outcomes.