Liver transplantation is critical for end-stage liver disease, but limited donor availability necessitates prioritizing patients on waiting lists. Predictive models like the Model for End-stage Liver Disease (MELD) are used for organ allocation and survival probabilities, but MELD's effectiveness is debated. This study aimed to develop machine learning models to predict postoperative survival at 1-month, 3-month, and 1-year intervals using preoperative data.

The dataset, after excluding missing or invalid data, comprised 454 patients with 52 features each. Leave-One-Out cross-validation was used to address data imbalance. K-Nearest Neighbor imputation handled missing values, ensuring robustness. Feature selection was performed using Decision Trees (DT) and Random Forests (RF), incorporating both clinically used and new features.

Various algorithms were evaluated, including DT, RF, Logistic Regression, Gaussian Naive Bayes (GuassianNB), and Linear Discriminant (LD) Analysis, to predict survival outcomes.

Results indicated that DT outperformed other feature selection methods, while GuassianNB excelled in predicting 1-year survival with an area under the curve of 0.61, a sensitivity of 0.98, and an F1-score of 0.89, demonstrating superior discrimination power. The LD model combined with RF feature selection was superior for 1-month and 3-month predictions. Additionally, a performance comparison of models for 1-year survival using MELD features and various selection methods was analyzed.

The study demonstrates that advanced machine learning models, particularly GuassianNB and LD Analysis with robust feature selection methods, can improve the prediction of postoperative survival in liver transplant patients. These findings could lead to better patient prioritization and outcomes in liver transplantation.