ESTRO meets Asia 2024 - Abstract Book

S224

Interdisciplinary – Mixed sites/palliation

ESTRO meets Asia 2024

Republic of. 3 Department of Otorhinolaryngology-Head and Neck Surgery, Yonsei University Wonju College of Medicine, Wonju, Korea, Republic of

Purpose/Objective:

International guidelines recommend single or hypo-fraction for palliative radiotherapy (PRT). However, due to the inherent difficulty in accurately predicting life expectancy, the customization of end-of-life care for individual patients poses a significant challenge. This study aims to develop a machine learning (ML) based mortality prediction model for PRT tailored to life expectancy.

Material/Methods:

A retrospective analysis encompassed 318 patients who expired after receiving PRT for advanced cancer from March 2013 to July 2023. Various model algorithms employing 22 variables were used to predict mortality within 30 days from initiation of PRT: extra trees, random forest, light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), support vector machine (SVM), and multilayer perceptron (MLP). We evaluated each model's performance using area under the receiver operating characteristic curve (ROC-AUC), accuracy, precision, recall, specificity, and F1 score.

Results:

The entire patient cohort was randomly allocated into training and test sets at a ratio of 7:3. The variable importance ranking altered with each execution of the code. However, the top seven variables (neutrophil lymphocyte ratio, albumin level, age, performance status, number of metastasis involved organs, number of received chemotherapy regimens, and admission during PRT) remained constant. After conducting hyper parameter tuning to enhance the performance of the models, the extra trees model exhibited the best performance. Among the models, SVM and MLP were not suitable for evaluating variable importance. LightGBM model was found to be over-fitted. We excluded those models. There was no significant difference in prediction performance among the random forest (ROC-AUC, 0.70), XGBoost (ROC-AUC, 0.64), and Extra trees (ROC-AUC, 0.71) models.

Conclusion:

We developed a ML-based mortality prediction model for tailoring PRT to life expectancy. High-importance variables are those recognized to exert a substantial influence on prognosis. Considering these high importance variables may be helpful when determining of PRT in clinical practice. This prediction model may be utilized as a clinical decision support system in guiding of PRT scheme.

Keywords: AI, decision support system, palliative care

References:

1. Tseng YD, Krishnan MS, Sullivan AJ, Jones JA, Chow E, Balboni TA. How radiation oncologists evaluate and incorporate life expectancy estimates into the treatment of palliative cancer patients: a survey-based study. Int J Radiat Oncol Biol Phys. 2013;87(3):471-8