Combining multiple models to generate consensus: application to radiation-induced pneumonitis prediction

Area of Science:

• Oncology
• Medical Physics
• Radiotherapy

Background:

• Predicting radiation-induced pneumonitis is crucial for lung cancer radiotherapy.
• Model fusion can enhance prediction robustness by leveraging complementary model strengths.

Purpose of the Study:

• To fuse predictions from multiple nonlinear multivariate models for improved radiation-induced pneumonitis risk estimation.
• To identify consensus features predictive of pneumonitis from fused model outputs.

Main Methods:

• Fusion of four distinct nonlinear multivariate models: Decision Trees, Neural Networks, Support Vector Machines, and Self-Organizing Maps.
• Cross-validated predictions averaged from each model to mitigate training set dependency.
• Area Under the Receiver Operating Characteristics Curve (AUC) used for performance evaluation.

Main Results:

• The fused model achieved an AUC of 0.79, outperforming individual models with lower variance.
• Identified five key consensus features for pneumonitis prediction: chemotherapy, equivalent uniform dose (EUD) across different exponents, lung volume receiving >20-30 Gy, female sex, and squamous cell histology.
• Fused patient outcome results were fitted to a logistic probability function for interpretability.

Conclusions:

• Model fusion offers a more robust and accurate method for predicting radiation-induced pneumonitis risk.
• Consensus features identified provide valuable insights for personalized radiotherapy planning and risk stratification.
• The logistic probability function enhances the clinical utility of the fused prediction model.