K. Halder1, T. K. Podder2, F. Maria-Joseph1, Y. Zheng3, M. D. Mix2, and T. Biswas4; 1Indian Institute of Technology, Roorkee, India, 2SUNY Upstate Medical University, Syracuse, NY, 3University Hospitals, Case Western Reserve University, Cleveland, OH, 4Metro Health, Case Western Reserve University, Cleveland, OH
Purpose/Objective(s): Early prediction of pneumonitis for lung cancer patients can potentially enhance patient care management including treatment optimization, reduction in complications, and better post-treatment follow-up care. The aim of this study is to develop an AI-based pneumonitis prediction model using common clinical inputs for non-small cell lung cancer (NSCLC) patients treated with SBRT. Materials/
Methods: Our proposed AI-based classifier model integrates four conventional machine learning classifier models, they are: Artificial Neural Network (ANN), Support Vector Machine, Naive Bayes, and Decision Tree, along with an output predictor module. The output layer generates the pneumonitis prediction based on highest probability estimation. We have incorporated clinical and dosimetric data collected from 230 early-stage NSCLC patients who underwent SBRT at our institution between 2017 and 2020; thirty-two patients had pneumonitis (19 grade 1, 13 grade 2, none grade 3 or higher). Twenty-two common clinical and dosimetric features: Age, Sex, T-stage, aCCI, SUV_max, ECOG Performance status, Smoking status, History of COPD, Total lung volume, Target location, Dose per fraction, GTV_max dose, GTV_min dose, GTV_mean dose, PTV_max dose, PTV_min dose, PTV_mean dose, Conformity index, Homogeneity index, Lung_V5, Lung_V20, and Lung_mean dose were included in the designed classifier model. The entire dataset was divided into a 70:30 ratio, with 70% allocated to the training phase, and the designed model was subsequently tested on the remaining 30% patients. Results: Patients’ median age was 74 years (range: 49-91 yrs), 51% were female; 53% had stage T1b, and majority had adenocarcinoma. The SBRT prescriptions were 48Gy-60Gy in 3-5 fractions. As compared to other four models, the proposed combined prediction model yielded superior performance with an accuracy of 0.88, a ROC-AUC of 0.67, and sensitivities for both classes: 0.89 (class-1) and 0.80 (class-0), respectively. Conclusion: Results indicate that the combined AI-based model has much better capability of predicting pneumonitis than the other four classical models. The proposed model may potentially assist in clinical decision-making process and may improve clinical outcomes.
