Specialized Session 4: NLP-based Methods for Medical Documents

Textual analysis plays a key role in medical research, aiding automated diagnosis from clinical narratives. This study applies causal forests to Electronic Health Records (EHR) to uncover causal links between language and diagnosis probabilities for diabetes and adrenal gland disorders using the MIMIC-III dataset. Findings reveal significant linguistic impacts on diagnosis probability and heterogeneity in treatment effects, with specific words identifying high-risk subgroups. These results highlight the utility of causal inference for advancing natural language processing in healthcare.

The lack of benchmark datasets for medical documents limits the reproducibility and generalizability of clinical Natural Language Processing (NLP) research. The European Clinical Case Corpus (E3C) offers a valuable multilingual resource, yet its limited annotations restrict its usability. In this study, we propose a novel clustering approach that extracts key medical concepts using LLaMAntino-3, an Italian Large Language Model, before applying transformer-based embeddings. Our results demonstrate that clustering these extracted concepts yields more coherent groupings than direct full-text clustering. By comparing different clustering strategies and evaluation methods, we provide insights into high-level annotations that can facilitate the creation of a structured benchmark dataset for medical NLP.

In this work, we propose an approach to tackle the challenge of data scarcity in clinical research. We devise an approach to convert the European Clinical Cases Corpus (E3C), an existing publicly available dataset annotated on several dimensions, into annotated Case Report Forms (CRFs). Our proposed pipeline is composed of two steps, the first leveraging the original data annotations to build a set of CRF protocols and the second filling the CRFs with the information contained in the original data. The result is a new and diverse dataset of CRFs filled by with human-based annotations. We depict the methodology, highlight its challenges, propose a baseline for the task, and release the dataset, hoping to empower researchers and clinicians to develop innovative approaches to tackle this critical task.