ESTRO 2025 - Abstract Book

S1382

Clinical - Lung

ESTRO 2025

3082

Digital Poster Pathologic Complete Response prediction with Machine Learning using Electronic Health Records (EHRs) Carlo Greco 1 , Edy Ippolito 2 , Claudia Tacconi 2 , Michele Fiore 2 , Aurelia Iurato 2 , Domenico Paolo 3 , Valerio Guarrasi 3 , Rosa Sicilia 4 , Paolo Soda 3,5 , Sara Ramella 2,6 1 radiation oncology unit, Fondazione Policlinico Universitario Campus Bio-Medico, Rome, Italy. 2 Radiation Oncology Unit, Fondazione Policlinico Universitario Campus Bio-Medico, Rome, Italy. 3 Unit of Computer Systems & Bioinformatics, Department of Engineering, Università Campus Biomedico di Roma, Rome, Italy. 4 Unit of Computer Systems & Bioinformatics, Department of Engineering, Università campus biomedico di Roma, Rome, Italy. 5 Department of Radiation Sciences, Radiation Physics, Biomedical Engineering, Umea University, Umea, Sweden. 6 Radiation Oncology Unit, Università Campus Biomedico di Roma, Rome, Italy Purpose/Objective: Pathologic response (pCR) after neoadjuvant therapy is associated with significant improvements in event free survival for patients with resectable NSCLC. Prediction pCR tafter neoadjuvant therapy is essential for determining appropriate strategy and guiding tailoring treatment. A non-invasive toosl to predict pCR are lacking. Our study aims to develop ensemble learning models using Electronic Health Records (EHRs) to predict pCR in NSCLC after radiochemotherapy. Material/Methods: In this work, an inhouse dataset was used. EHRs and clinical data processed through different unimodal deep learning approaches and a multimodal deep learning approach to predict prognosis as a binary classification over pCR.To predict prognosis as a binary classification over pCR, we utilized EHRs and clinical data processed through different unimodal deep learning approaches and a multimodal deep learning approach. Unimodal EHRs: the proposed method includes a three-stage pipeline: (i) Encoding clinical entities in EHRs using a Named Entity Recognition (NER) system. (ii) Combining them with a hierarchical attention mechanism where Token embeddings from the NER system were weighted via a soft attention layer to produce sentence embeddings, processed with another attention layer to generate a patient representation. (iii) Predicting the binary pCR outcome using machine learning. Clinical data were used directly as input for the machine learning to predict pathological complete response (pCR) and then patient representations generated by the hierarchical attention mechanism from EHRs were evaluated with clinical data (multimodal). These combined representations served as input to the machine learning model for predicting pCR. For Training we used 10-fold cross-validation to train the models and for the binary pCR prediction we used several machine learning models. Results: We collected 134 EHRs related to first visit before treatment and include a comprehensive array of patient information, such as personal data, medical history, histology, imaging reports, physical examinations, . We included the most relevant clinical features: sex, age at diagnosis, cancer histology type, cTNM status We also included the most relevant some clinical features: sex, age, histology, and cTNM stage. The AUC obtained with use of the EHRs, clinical features and the combination of both were 68.9,61.4 and 68.9 respectively (Fig. 1).