2437 - Prediction of Esophagitis following Intensity-Modulated Radiotherapy for Lung Cancer Using Statistical and Machine Learning Methods

Purpose/Objective(s): Radiation esophagitis (RE) remains a clinically relevant problem for lung cancer patients even with highly conformal radiotherapy techniques. Here, we predicted the risk of RE based on a comprehensive list of clinical and dosimetrical factors using different statistical and machine learning (ML) methods.

Materials/Methods: Lung cancer patients (N=102, stage I and II 27%, stage III and IV 73%) underwent IMRT to a median delivered dosef 60 Gy (range 30-70) in 1.5 (BID) – 3 Gy fractions with concurrent chemotherapyn 67/102 (66%) patients. RE grades were recorded prospectively according to CTCAE v5 criteria. Risk factors for RE were analyzed with univariate analysis using chi-square test and independently using 4 machine learning algorithms (Logistic Regression [LR], least absolute shrinkage and selection operator [LASSO], Random Forest [RF], light gradient boosting machine [LightGBM]). The dataset was split into a training set (n=81) and a test set (n=21). 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: Median age was 67 years (range 44-89), 56 (55%) were female and 43 (42%) were African American. 42 had G0, 31 G1, 26 G2, 2 G3 and 1 G4 RE. On univariate analysis, G = 2 RE was significantly related with esophagus length irradiated to 30, 40, and 50 Gy, percentage of esophagus = 30 Gy, esophagus mean dose, irradiation of full circumference with prescribed dose (p<0.0001); length treated to 40 Gy and percent volume receiving 30 Gy were most significant (p=0.000002). Concurrent chemotherapy, treatment with famotidine prior to RT, Caucasian race, younger age and female patients also had a higher risk (p=0.0002 - 0.027). Applying ML methods, mean AUC for the prediction of each RE grade (0, 1, and 2,3 combined) on the test dataset were: 0.84, 0.82, and 0.65, respectively. Esophagus length irradiated to 30 and 40 Gy, pretreatment use of famotidine and chemotherapy were the most valuable influencing features identified by both RF and LightGBM. Of note, percentage of esophagus = 60 Gy, a standard dose constraint, was not significantly related with grade = 2 esophagitis (p=0.14 from univariate analysis).

Conclusion: Grade = 2 esophagitis occurs in 28% of patients despite using highly conformal radiotherapy techniques. Both univariate and ML-based analysis indicate strong predictive strength of intermediate dose levels in addition to concurrent chemotherapy. Adding constraints in the intermediate dose range may lead to better plan quality and lower toxicity risk. Interestingly, contralateral esophagus sparing was not a significant factor for RE risk.