ESTRO 2025 - Abstract Book

S3844

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

3738

Digital Poster Artificial Intelligence in the prediction of clinical response in patients with COVID-19 pneumonia treated with low-dose pulmonary radiotherapy. Judith Osuna Ramírez 1,2 , Berta Piqué Smith 2 , Marta Canela Capdevila 2 , Raquel García Pablo 2 , Alberto Martínez Caballero 3 , Paula Maixé Brull 1,2 , Andrea Jiménez Franco 2 , Juan Manuel Jiménez Aguilar 2 , Rocío Benavides Villarreal 1,2 , Laura Torres Royo 1 , Johana Cristina Acosta Artiga 1 , Pablo Araguas Mora 1 , Víctor David Calderón Ávila 1 , Jordi Camps Andreu 2 , Meritxell Arenas Prat 2 , Victor Hernandez Masgrau 4 1 Department of Radiation Oncology, Hospital Universitari de Sant Joan, Reus, Spain. 2 Unitat de Recerca Biomèdica, Institut d'Investigació Sanitària Pere Virgili, Reus, Spain. 3 Medical Image Analysis and Biometry Laboratory, Universidad Rey Juan Carlos, Madrid, Spain. 4 Department of Medical Physics, Hospital Universitari de Sant Joan, Reus, Spain Purpose/Objective: This study aims to develop a predictive model capable of determining the clinical response to lung-targeted anti inflammatory Low-Dose RadioTherapy (LDRT) in patients suffering from pneumonia caused by COVID-19 1 using multi-omic data. Material/Methods: Clinical, analytical, and radiomic characteristics from 50 patients with COVID-19 pneumonia and treated with LDRT were analyzed. Clinical data, including demographic and treatment data, were collected from medical records. Blood tests performed prior to LDRT provided biochemical and hematological data. Additionally, 1688 radiomic features related to texture, shape, and intensity distributions were extracted from simulation computed tomography (CT) images using PyRadiomics 2 library. Predictive models were trained and tested with four data preprocessing aproaches. Two of them addressed class imbalance using Synthetic Minority Oversampling TEchnique (SMOTE) while others did not. In addition, two different feature selection methods were compared (manual or minimum Redundancy and Maximum Relevance algorithm). Finally, three Supervised Machine Learning algorithms were trained and evaluated for binary classification, including Random Forest, Logistic Regression, and Decision Tree, and their performance was assessed through Cross Validation.

Results: The models that provided the best results in each approach are presented below.

For imbalanced data with manually selected features, Random Forest achieved a sensitivity of 0.87 and a specificity of 0.57. With automatically selected features, Random Forest showed a sensitivity of 0.90 and a specificity of 0.43. For balanced data, with manually selected features, Random Forest reached a sensitivity of 0.84 and a specificity of 0.77. With automatically selected features, it had a sensitivity of 0.83 and a specificity of 0.81 (see figures 1 and 2).