Author: Ibtisam Almajnooni, Siyong Kim, Nathaniel Miller, Elisabeth Weiss, Lulin Yuan 👨🔬
Affiliation: Virginia Commonwealth University 🌍
Purpose: Radiation-induced esophagitis (RE) is a common concern in lung cancer IMRT. Recent studies have indicated that the risk of radiation side effects varies greatly with patients’ baseline clinical conditions. In this study, we utilize machine learning (ML) methods to better identify the effect of various clinical conditions on the risk of RE.
Methods: This study included 102 consecutive patients with locally advanced lung cancer treated in our department. The clinical endpoint was radiation-induced esophagitis (RE) graded according to CTCAE v5 criteria. The clinical features extracted from the patient charts included demographic information, as well as defined clinical conditions and dosimetric factors.
Four ML algorithms (Logistic Regression [LR], least absolute shrinkage and selection operator [LASSO], Random Forest [RF], and light gradient boosting machine [LightGBM] were employed to predict RE grades (0(44 patients), 1(32), 2(24) and 3(2) combined) based on clinical features and dosimetric parameters. The dataset was split into a training set (n=81) and a test set (n=21). The models were trained and validated using stratified 5-fold cross-validation. The performance of each model was evaluated using the area under receiver operating characteristics curve (AUC).
Results: The average AUC for the prediction of three RE grades on the test dataset from LASSO, LR, RF and LightGBM are: 0.76, 0.77, 0.79 and 0.79, respectively. The maximum and mean esophagus dose, use of H2 blockers prior to radiotherapy, esophagus length irradiated to 30, 40 and 50 Gy, and age were the most valuable influencing features identified by RF and LightGBM.
Conclusion: Our analysis demonstrates the capability of ML algorithms to predict esophagitis using dosimetric and clinical factors. Future considerations include esophageal sub-volume effects and validating these models on larger datasets to address class imbalance and improve the models’ performance.