ESTRO 2020 Abstract Book

S828 ESTRO 2020

were collected (clinical, histological, and therapeutic variables [including dosimetric data], among others). The end points of analysis (development of acute esophagitis, cough, dyspnea or pneumonitis) were estimated and scored using Common Terminology Criteria for Adverse Events version 4.0. For the analysis, patients were grouped according grades (grade <2 vs grade ≥2). The cases from both groups were balanced by random elimination of samples from the majority group. The following forward selection techniques were applied for exploratory purposes: 1) Minimum Redundance-Maximum Relevance (mRMR); 2) Relief; 3) Random Forest (RF); and 4) Information Gain (IG). In addition, three subsetting methods were used for that purpose: 1) Correlated-based Feature Selection; 2) Boruta; and 3) Chi – squared filtering (ChiSq). A voting method (VM) which considers the features selected by two or more corresponding selection methods was applied. The characteristics selected in each case have been used to train the following classifiers: 1) Support vector machine (SVM); 2) Artificial Neural Network (ANN); 3) Linear Regression (LR); and 4) Naïve Bayes (NB). All selection of characteristics-classifiers combinations have been trained by 5-fold cross validation (80% of samples for training and 20% for testing) and their output were evaluated by means of the area under the ROC curve (AUC) for each RT-induced toxicity prediction. Furthermore, the same predictive models for toxicity have been trained using the common features used and recommended by 2 experienced radiation oncologists. Results For the training dataset, the rates of acute grade ≥2 esophagitis, cough, dyspnea or pneumonitis were 34% (N=204), 31% (N=123), 19% (N=115), and 28% (N=166), respectively. The models that showed a higher AUC for predicting acute esophagitis, cough, dyspnea or pneumonitis were VM+LR (AUC 0.88), RF+LR (AUC 0.86), Boruta +LR (AUC 0.79), and ChiSq +LR (AUC 0.90), respectively. The number of features used for each model was 58, 38, 16, and 31 respectively. When using the features (N=27) used and recommended by 2 experienced radiation oncologists the AUC for predicting acute esophagitis, cough, dyspnea or pneumonitis were 0.79, 0.64, 0.69, and 0.56, respectively. Conclusion Feature selection methods improve accuracy in toxicity prediction. This tool could help to enrich lung cancer guidelines and may be useful for guiding RT intensity in an individualized therapy. PO-1534 Tumour response modelling to radiochemotherapy with gemcitabine for pancreatic cancer B. Costa Ferreira 1,2,3 , J. Dias 3,4 , P. Mavroidis 5,6 , H. Rocha 3,4 , A. Gomes 1 1 Porto Polytecnic, School of Health, Porto, Portugal ; 2 Aveiro University, I3N Physics Department, Aveiro, Portugal ; 3 Coimbra University, Institute for Systems Engineering and Computers at Coimbra, Coimbra, Portugal ; 4 Coimbra University, CeBER and FEUC, Coimbra, Portugal ; 5 University of North Carolina, Department of Radiation Oncology, Chapel Hill, USA ; 6 Karolinska Institutet and Stockholm University, Division of Medical Radiation Physics, Stockholm, Sweden Purpose or Objective To assess tumour response as a function of the dose of gemcitabine (GEM) in concomitant radiochemotherapy (RCT) of locally advanced pancreatic cancer (LAPC). Three models describing this tumour response to chemotherapy (CH) were tested: the constant, the linear, and the Hill model. Material and Methods Clinical data was retrieved from the literature where details on patients, therapy and response to treatment were provided. The response to concomitant RCT with GEM

Results The incidence of lymph node metastasis (LNMs) was 58% in both cohorts. The AUCs of the clinical, radiomics and combined models were 0.79, 0.69 and 0.82 in the developmental cohort, and 0.65, 0.63 and 0.69 in the external validation cohort, with good calibration demonstrated. The ROC plots are shown in Figure 2. In comparison, the AUC of current cN-stage in development and validation cohorts was 0.60 and 0.66, respectively. For overall survival, the combined clinical and radiomics model achieved the best performance to discriminate the external validation cohort (X 2 6.08, CI = 0.60, df 1, p = 0.01).

Conclusion Accurate diagnosis of LNMs is crucial for predicting prognosis and guiding treatment decisions. Despite obtaining signal for improved prediction in the development cohort, the models using PET radiomics derived from the primary tumour were not fully replicated in an external validation cohort. PO-1533 Feature selection methods improve accuracy in radiation toxicity prediction for lung cancer F. Núñez-Benjumea 1 , S. González-García 1 , J. Moreno- Conde 1 , A. Moreno-Conde 2 , J. Cacicedo 3 , B.D. Delgado 4 , M. Borrego 4 , S. Perez Luque 4 , C.L. Parra-Calderón 1 , J.L. Lopez Guerra 5 1 Institute of Biomedicine of Seville- IBIS / Virgen del Rocío University Hospital / CSIC / University of Seville, Biomedical Informatics- Biomedical Engineering and Health Economics, Seville, Spain ; 2 Virgen Macarena University Hospital, Biomedical Informatics- Biomedical Engineering and Health Economics, Seville, Spain ; 3 Cruces University Hospital, Radiation Oncology, Barakaldo, Spain ; 4 Virgen del Rocío University Hospital, Radiation Oncology, Seville, Spain ; 5 Virgen del Rocío University Hospital, Radiation oncology, Sevilla, Spain Purpose or Objective The project named S32 aims to develop a Learning Health System to predict radiation-induced toxicity. This tool will allow physicians know in advanced the effects of radiation therapy in lung cancer patients. Material and Methods The lung cancer dataset includes clinical information of 596 patients (training dataset) treated with thoracic radiation therapy (RT) from 2013 to 05/2019 and 46 patients (validation dataset) treated from 06/2019 to 09/2019 at 2 Institutions. For each patient, 59 variables