Natural language processing to estimate RECIST response in cancer patients

Treatment response in oncology trials, evaluated according to the Response Evaluation Criteria in Solid Tumors (RECIST) guidelines, is commonly reported as free text in radiology reports, which require manual review for the response extraction. Moreover, in clinical practice, reports do not explicitly mention the RECIST response. We fine-tuned two deep- learning models for Natural Language Processing (NLP), CamemBERT and DrBERT for response classification from radiology reports and we tested them on less structured reports collected in clinical routine to detect instances of disease progression. Both model performed well on reports collected in trials settings (accuracy > 97%), and CamemBERT classified progression in less structured reports with 90% accuracy. Our study provides evidence on the feasibility of using NLP models to determine the treatment response from clinical notes.