ESTRO 38 Abstract book

S1034 ESTRO 38

standardizing, analyzing and reusing retrospective lung cancer patients’ information to generate radiation- induced toxicity predictive models that could be actioned in real-time during the treatment planning process. A prospective trial will be developed in a real environment in order to elucidate the accuracy of the predictions yielded by the system and its impact in health-related quality of life, among other endpoints. In this work, the preliminary results of the feature selection methods and prediction models investigated for this purpose are presented. Material and Methods As a preliminary step for the radiation-induced toxicity prediction process, several Feature Selection (FS) methods based on either feature sub-setting or forward selection techniques have been tested along with two different prediction models based on Support Vector Machines (SVM) and Artificial Neural Network (ANN) for predicting radiation-induced acute esophagitis and pneumonitis. The following sub-setting algorithms were applied to both SVM and ANN prediction models: 1) Correlation-based Feature Selection (CFS); 2) Random Forest-based (Boruta); and 3) Chi-Squared filter. In addition, the prediction models were also trained with those features which appeared in two or more of the FS methods listed above (which we called the “polling” method). In addition, the following forward selection techniques were applied only to the SVM-based prediction model for exploratory purposes: 1) Minimum Redundance- Maximum Relevance (mRMR); 2) Relief; 3) Random Forest (RF); and 4) Information Gain (IG). The SVM-based prediction model was also trained with those features which appeared in two or more of the FS methods listed above. Furthermore, and for comparison purposes, the performance of the SVM-based prediction method was also tested without applying any FS method, i.e., with all the available features. Results FS methods and predictive models have been tested with a retrospective dataset with information gathered during routine care for the last 5 years. Radiation-induced prediction models were developed for the two most common side effects found in the dataset: acute esophagitis (N = 406) and acute pneumonitis (N = 408). The models were trained with the 80% of the samples, leaving the remaining 20% for testing purposes. Accuracy of the prediction was measured in terms of AUC (Table).

right; L=patient’s left. Dark green=Genioglossus muscle/Tongue; Dark yellow=Submandibular gland; Light green=Hyoglossus muscle; Light yellow=Geniohyoideus muscle; Pink=Upper pharyngeal constrictor muscle; Purple=Base of tongue; Red=Anterior digastric muscle; White=Mylohyoideus muscle.

Results Dose to several swallowing structures was associated with dysphagia post-RT in univariate analysis. When applying a multivariate model, the mean dose to the larynx and the epiglottis as well as the maximum dose to the contralateral submandibular gland were associated with PAS≥4, PAS≥6 and PAS≥4 as well as PAS≥6 respectively. The mean dose to the contralateral submandibular gland and the maximum dose to the contralateral anterior digastric muscle were associated with DESdC≥3. Figure 2. Dose-volume histogram for epiglottis, the structure with the best discrimination power for dysphagia according to MVA and associated statistically significant volume differences between dysphagia and non-dysphagia patients.

Conclusion Dose-response relationships were found for specific dysphagia endpoints. EP-1902 S32: A decision Support System to predict radiation toxicity in lung cancer patients F. Núñez Benjumea 1 , J. Moreno Conde 1 , A. Moreno Conde 1 , S. González García 1 , M.J. Ortiz Gordillo 2 , J. Riquelme 3 , M.D.C. Fernández Fernández 2 , C.L. Parra Calderón 1 , J.L. Lopez Guerra 2 1 Hospital Universitario Virgen del Rocio, Group of Technological Innovation, Sevilla, Spain ; 2 Hospital Universitario Virgen del Rocio, Radiation oncology, Sevilla, Spain ; 3 University of Sevilla, Department of Computer Language and Systems, Sevilla, Spain Purpose or Objective Decision support systems are a growing class of tools with the potential to impact healthcare. The S32 project aims to develop an informatics infrastructure oriented towards

Conclusion The best prediction model found for predicting acute esophagitis was the ANN trained with the features yielded by the CFS method. The highest AUC found for predicting acute pneumonitis was related to the model built upon the SVM and Polling of the sub-setting algorithms investigated. This tool could help to define new lung patients care protocols based on predictive models for patient toxicity. EP-1903 Learning from scanners: radiomics correction modeling I. Zhovannik 1,2 , J. Bussink 1 , R. Fijten 2 , A. Dekker 2 , R. Monshouwer 1 1 Radboud University Medical Center, Radiation Oncology,