During lung cancer radiotherapy, tumor motion magnitude may change, resulting in either inadequate margins or non-ideal dose conformality (margins are unnecessarily too big). To predict large motion change using only patient-specific clinical information and the simulation 4DCT images, we proposed an ensemble learning model (ELM) with robust feature selection (FS). For this specific study, we minimized the false negative prediction so that we will not miss patients who might have a subsequent large change in tumor motion. Thus, we employed positive predictive values (PPV) as the evaluation metric instead of commonly used accuracy scores.Seventeen features were collected including both clinical factors and imaging features. The motion change is defined by motion difference in SI direction between the simulation 4DCT and mid-course repeat 4DCT with the clinical threshold set as 3 mm. The dataset consisted of 140 patients, where 31 showed large tumor motion change. Four hundred seventy-two sets of 4DCT images were analyzed in this study. The ELM adopts 5 conventional predictive models (logistic regression (LR), decision tree (DT), support vector machine (SVM), random forest (RF), and XGBoost (XGB)) as base predictors. Then, the results are employed as inputs for a two-layer neural network to obtain a final prediction. To enhance the feature sensitivity and to reduce data noise, we developed a robust FS strategy with sequential backward selection. Each criterion score is an average from 5-fold cross-valuation, and the target feature of each round is based on 1000 repeated experiments. The peak of the criterion score after sequential removal of the least sensitive features was used to select the optimal feature set.By employing PPV as the criterion function, all the base predictors ranked initial SI tumor motion as the most important feature. ROC curve with selected sensitive features consistently outperformed the one with all features. For ELM, the AUCs (mean std. ± dev.) were 0.63 ± 0.04 when all features were included and increased to 0.72 ± 0.02 for PPV-based selected features only. The prediction comparisons were performed for all base predictors and ELM with 100 repetitions. The PPV results (Table 1) demonstrated the proposed ELM obtained superior performance than each base predictor.We proposed an ELM to identify patients with potential large tumor motion change using only information prior to the first treatment. The ELM showed superior performance with PPV-driven metrics. We also developed a feature selection strategy for each base predictor, and the consistent superiority was demonstrated by ROC results. In future work, additional information from on-treatment evaluation CTs may further improve the prediction.