D1.1 - Requirements Task Definition (Analysis of requirements for applying NLP techniques to the unstructured portion of ED EHRs and definition of the NLP tasks)

This document aims to outline the necessary elements to set up the Natural Language Processing (NLP) component within the Work Package 1 of the project. More precisely, this deliverable reports on Task 1.1, “Context analysis and task definition”, aiming to analyze the requirements for applying NLP techniques to the unstructured portion of EHRs in EDs and the definition of the NLP tasks useful for emergency physicians and nurses.

According to the project plan, the task includes the following activities:
● Analysis of the context and the requirements for each emergency department of the project (Section 2). There are three main aspects we have been working on: an initial workflow involving the NLP component, in collaboration with Astir (Section 2.1); an anonymization component ensuring that the NLP component manages data without sensitive information during the training of the model (Section 2.2); an initial study of the hardware requirements of the NLP component at inference time, in order to test the compatibility with current hardware available in EDs (Section 2.3).
● Assessment and selection of the reference resources to be used for NLP processing (Section 3). We are considering two kinds of resources: non-annotated resources, necessary for pre-training large language models in the medical domain (Section 3.1) and annotated resources, necessary to train a language model on the specific tasks of eCREAM (Section 3.2).
● Definition of the NLP task, in terms of the clinical entities to be automatically extracted (Section 4). In the first year of the project here we have been working on the following aspects: in collaboration with IRFMN and Orobix, we have provided a definition of the eCREAM tasks related to NLP, presented as the eCREAM benchmark (Section 4.1); an assessment of the available language models in the medical domain, including initial work on a T5-large medical model (Section 4.2).