2024 World Conference on Lung Cancer (WCLC) - ePosters-EP.05B.05 Integrative Prediction Model for Radiation Pneumonitis: Genetic and Clinical-Pathological Factors Utilizing Machine Learning

EP.05B.05 Integrative Prediction Model for Radiation Pneumonitis: Genetic and Clinical-Pathological Factors Utilizing Machine Learning

Back to course

Pdf Summary

The study aimed to improve prediction models for radiation pneumonitis (RP), a common side effect of thoracic radiotherapy (RT) in lung cancer patients. Traditional models relying on clinical and dosimetric data have limited accuracy, prompting the integration of genetic factors through machine learning (ML) techniques.<br /><br />A cohort of 59 lung cancer patients treated with RT was examined, focusing on genetic variations (SNPs) in TGF-β1 and BMP genes. These genetic markers, combined with clinical variables such as age, sex, smoking history, and dosimetric data like total dose and target volume, were used to develop predictive models.<br /><br />The study employed both statistical and ML methods, including logistic regression and various feature selection techniques. ML models used were Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Machine (SVM), evaluated with nested cross-validation for performance optimization. <br /><br />Results showed that integrating genetic data improved model accuracy significantly. The combined model, which included clinicopathological factors and genomic variables (with key predictors being age, smoking history, PTV volume, and BMP2 rs1979855), achieved an AUC of 0.822, indicating superior predictive accuracy over the clinical-only model (AUC 0.741). Among feature selection methods, the wrapper-based set performed best, with XGBoost showing the highest accuracy for predicting severe RP.<br /><br />The study concludes that incorporating genetic data markedly enhances the predictive capability for severe RP, which can facilitate personalized treatment planning. Such models can potentially forecast radiation-related toxicities more accurately, allowing for customized RT approaches based on individual risk profiles.

Asset Subtitle

Hong In Yoon

Meta Tag

Speaker Hong In Yoon

Topic Pulmonology and Staging

Keywords

radiation pneumonitis

thoracic radiotherapy

lung cancer

genetic factors

machine learning

predictive models

TGF-β1

BMP genes

feature selection

personalized treatment