An Explainable AI Approach to Predicting Radiation Pneumonia in Head & Neck Cancer

Abstract

Radiation pneumonia is a serious complication for head and neck cancer patients receiving radiation therapy, so it is important to estimate the risk accurately. In the study, a combination of XGBoost and LightGBM models is used, with XGBoost as the meta learner, to estimate radiation pneumonia. To manage the large difference between the number of pneumonia and non-pneumonia cases (568 pneumonia vs. 1914 non-pneumonia cases), the SMOTETomek resampling approach was used on the training data. Moreover, with SHAP analysis, the changes in how each feature affected the predictions between the original imbalanced and the balanced data were clearly seen. Based on the experimental evaluation over the considered dataset, the precision, recall, F1-Score, F2-Score, and ROC-AUC have improved significantly. The study reveals that the approach proposed can predict well while giving simple and meaningful insights to doctors. The developed model may enable oncologists to understand and provide insights into AI-driven decision-making to make informed decisions.

Ähnliche Arbeiten