1102 - Development of a Machine Learning-Based 30-Day Survival Prediction Model for Patients Undergoing Palliative Radiation Therapy to Establish Optimal End-of-Life Care

Purpose/Objective(s):

Palliative radiation therapy (PRT) is delivered to alleviate symptoms in patients with terminal cancer. However, administering PRT to those with limited life expectancy may inadvertently compromise their quality of life, necessitating accurate survival prediction to optimize clinical decision-making. This study hypothesized that machine learning (ML) models utilizing primarily standard laboratory parameters as predictors could accurately predict 30-day mortality following PRT.

Materials/Methods:

This study analyzed retrospective electronic medical record (EMR) data from patients with terminal cancer who received PRT at a single institution between 2013 and 2022. A total of 18,854 patients were included in the final analysis after excluding those with missing survival or laboratory data, or age < 20 years. The input variables comprised 34 features, including 29 laboratory parameters and 5 demographic and treatment-related factors. The target variable was defined as 30-day mortality, representing death occurring within 30 days following the final PRT session. Missing laboratory values were imputed using Multiple Imputation by Chained Equations. Continuous variables were standardized, and class imbalance was addressed by applying the Synthetic Minority Oversampling Technique solely to the training set. The overall cohort was stratified into a training set (n=15,083) and an internal validation set (n=3,771). Five ML algorithms—Random Forest (RF), Support Vector Machine (SVM), XGBoost, CatBoost, and Logistic Regression (LR)—were trained and optimized using the training set. Model performance on the internal validation set was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity, and SHAP values were utilized to interpret feature importance.

Results:

Among 18,854 patients, 1,590 (8.4%) died within 30 days. This mortality rate was consistent across both training and internal validation sets. As shown in the table, all five ML models demonstrated robust predictive performance with AUCs exceeding 0.85. Logistic Regression and SVM achieved the highest AUC of 0.860. SHAP analysis identified albumin, lymphocyte count, and bilirubin as the most significant predictors; specifically, lower albumin/lymphocyte levels and higher bilirubin levels were associated with increased mortality risk.

Conclusion:

The developed ML models, utilizing primarily standard laboratory parameters, demonstrated high accuracy and robustness in predicting 30-day mortality following PRT. These findings suggest that ML-based survival prediction can serve as a valuable decision-support tool to optimize clinical decision-making and resource allocation for terminal cancer patients. External validation is planned to confirm generalizability.