Rutgers Robert Wood Johnson Medical School New Brunswick, NJ
D. Singh1, S. K. Jabbour2, and M. L. Reyhan3; 1Rutgers Robert Wood Johnson Medical School, Manalapan, NJ, 2Department of Radiation Oncology, Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, 3Department of Radiation Oncology, Rutgers Cancer Institute of New Jersey, Rutgers University, New Brunswick, NJ
Purpose/Objective(s): Pneumonitis is a dose-limiting side effect of radiotherapy (RT) and occurs at higher rates in patients treated for non-small cell lung cancer (NSCLC). Machine learning prediction of high-risk patients based on radiomics is a promising clinical tool for minimizing pulmonary toxicity. This project sought to retrospectively determine the predictive accuracy of a machine learning model based on radiomics for NSCLC patients receiving RT and durvalumab immunotherapy. The model was then applied to patients undergoing adaptive radiotherapy (ART). Materials/
Methods: A previously established and validated predictive machine learning algorithm of pneumonitis in patients undergoing immunotherapy was applied to stage III NSCLC patients undergoing RT followed by immunotherapy (n=15). Two radiomics features, skewness and angular variance of sum of squares, were extracted from simulation CT images to predict pneumonitis. Demographic data (age and smoking status) and dose volume histogram data (lung V5Gy and V20Gy) were incorporated to improve the model. Patients’ pneumonitis status within 6 months of completion of radiation therapy were identified to evaluate predictive accuracy using sensitivity and specificity. Imaging data was then acquired from stage III NSCLC patients undergoing ART and immunotherapy (n=14). The machine learning model (incorporating radiomics, demographics, and DVH values) was applied to the initial simulation CT from these patients and predictive accuracy was evaluated. Results: The baseline model for traditional RT as derived from the original publication had a sensitivity: 33.33% and specificity: 83.33%. When incorporating demographics and DVH values, the sensitivity and specificity increased to 66.67% and 91.67%, respectively. When applied to the ART patients, model sensitivity and specificity decreased to 33.33% and 50%, respectively. The model overpredicted patients who developed pneumonitis. This overprediction may demonstrate the importance of ART in mitigating pneumonitis. Conclusion: A machine learning model to predict radiation pneumonitis in NSCL patients undergoing radiation therapy and immunotherapy was developed and validated. When applied to patients undergoing ART, the model overpredicts patients who will develop pneumonitis. Further prospective studies are needed to better understand this overprediction. Abstract 2137 – Table 1
