A machine learning-based model for assessing the risk of new-onset liver injury following cardiac surgery under cardiopulmonary bypass

Abstract

Objective: This study aimed to develop and validate a predictive model for assessing the risk of new-onset liver injury following cardiac surgery under cardiopulmonary bypass (CPB), using non-redundant and informative features extracted from electronic health records. Materials and Methods: We employed machine learning algorithms including Generalized Additive Model (GAM), Random Forest, XGBoost, LightGBM, and Fully Convolutional Network (FCN) to construct the models using data from 5,364 patients at a large medical center in China, and validated these models with an independent dataset of 1,207 patients from another center. A three-stage feature selection process was used to refine the input variables. Results: The GAM model displayed the best performance with good predictive accuracy in both internal and external validations, despite a noticeable performance decline in the external dataset potentially due to differences in feature distributions. The most impactful factors included CPB time, cryo time, and preoperative bilirubin levels. Conclusion: The predictive model developed provides a valuable tool for early identification of patients at risk of postoperative liver injury, potentially aiding in preventive treatment planning. However, the model requires further prospective validation and optimization for broader application across different medical centers. The model's robustness against clinical practice variations highlights its potential utility in improving patient safety and reducing healthcare costs.