Local Recurrence Prediction After Carbon-ion Radiotherapy for Early-stage Non-small Cell Lung Cancer Using Machine Learning.
Predicting local recurrence remains challenging in carbon-ion radiotherapy (CIRT) for non-small cell lung cancer (NSCLC). In this study, we aimed to develop and validate a machine learning model to predict local recurrence after CIRT for early-stage peripheral NSCLC.
We retrospectively analyzed patients treated with CIRT at our institution between 2010 and 2020. An Extreme Gradient Boosting classifier using clinical parameters was developed to predict local recurrence within 24 months after CIRT using a nested threefold cross-validation framework. The optimization objective was the area under the receiver operating characteristic curve (ROC-AUC). The prediction model was evaluated using the area under the precision-recall curve (PR-AUC) and survival analysis. The patients were stratified into risk groups for 2-year local control. Model interpretability was explored using SHapley Additive exPlanations (SHAP).
A total of 124 patients were evaluated. Ten (8.1%) patients experienced local recurrence within two years. The 2-year local control rate was 91.0%, with a median follow-up period of 44.9 months. The prediction model showed an ROC-AUC of 0.622. Furthermore, the PR-AUC was 0.145, and the 2-year local control rates were 94.0% in the low-risk group and 65.0% in the high-risk group (log-rank test, p<0.01). SHAP highlighted the importance of the Brinkman Index, C-reactive protein level, and solid component tumor diameter.
A machine learning model based on clinical parameters may predict 2-year local recurrence after CIRT for early-stage peripheral NSCLC.
We retrospectively analyzed patients treated with CIRT at our institution between 2010 and 2020. An Extreme Gradient Boosting classifier using clinical parameters was developed to predict local recurrence within 24 months after CIRT using a nested threefold cross-validation framework. The optimization objective was the area under the receiver operating characteristic curve (ROC-AUC). The prediction model was evaluated using the area under the precision-recall curve (PR-AUC) and survival analysis. The patients were stratified into risk groups for 2-year local control. Model interpretability was explored using SHapley Additive exPlanations (SHAP).
A total of 124 patients were evaluated. Ten (8.1%) patients experienced local recurrence within two years. The 2-year local control rate was 91.0%, with a median follow-up period of 44.9 months. The prediction model showed an ROC-AUC of 0.622. Furthermore, the PR-AUC was 0.145, and the 2-year local control rates were 94.0% in the low-risk group and 65.0% in the high-risk group (log-rank test, p<0.01). SHAP highlighted the importance of the Brinkman Index, C-reactive protein level, and solid component tumor diameter.
A machine learning model based on clinical parameters may predict 2-year local recurrence after CIRT for early-stage peripheral NSCLC.
Authors
Mochida Mochida, Miyasaka Miyasaka, Kubo Kubo, Yoshida Yoshida, Okano Okano, Kawamura Kawamura, Ohno Ohno
View on Pubmed