Predicting lung radiotherapy-induced pneumonitis using a model combining parametric Lyman probit with nonparametric...

Area of Science:

• Medical Physics
• Radiation Oncology
• Clinical Informatics

Background:

• Lung radiation-induced pneumonitis is a significant toxicity affecting lung cancer patients undergoing radiotherapy.
• Accurate prediction of pneumonitis risk is crucial for optimizing treatment plans and minimizing adverse events.

Purpose of the Study:

• To develop and validate a predictive model for lung radiation-induced Grade 2+ pneumonitis.
• To enhance the predictive accuracy beyond traditional dose-based metrics by incorporating clinical variables.

Main Methods:

• A predictive model was developed using data from 234 lung cancer patients, incorporating the Lyman normal tissue complication probability (LNTCP) metric.
• Weighted nonparametric decision trees were integrated with LNTCP, utilizing both dose and non-dose patient-specific factors.
• A boosting process enhanced model accuracy, with performance evaluated via 10-fold cross-validation and a simplified model extracted for practical application.

Main Results:

• The developed model achieved an area under the receiver operating characteristics curve of 0.72, significantly outperforming the LNTCP metric (0.63, p=0.005).
• Key predictive variables identified include LNTCP, gender, histologic type, and chemotherapy and treatment schedules.
• Injury prediction varied widely based on non-dose factors, highlighting the importance of personalized risk assessment.

Conclusions:

• Integrating non-dose factors via decision trees significantly improves radiation pneumonitis prediction compared to dose-based models alone.
• The simplified model provides a practical tool for assessing pneumonitis risk, aiding in personalized radiotherapy planning for lung cancer patients.