Capítulo de livro Acesso aberto Revisado por pares

Argument Mining on Clinical Trials

2018; Linguagem: Inglês

10.3233/978-1-61499-906-5-137

ISSN

1879-8314

Autores

Tobias Mayer, Elena Cabrio, Marco Lippi, Paolo Torroni, Serena Villata,

Tópico(s)

Natural Language Processing Techniques

Resumo

In the latest years, the healthcare domain has seen an increasing interest in the definition of intelligent systems to support clinicians in their everyday tasks and activities. Among others, this includes novel systems for the field of Evidence-based Medicine. The latter relies on the principle of critically appraising medical evidence and combining high quality evidence with the individual clinical experience of the practitioner with respect to the circumstances of a patient to achieve the best possible outcome. Hence, most of the proposed intelligent systems aim either at extracting information concerning the quality of evidence from clinical trials, clinical guidelines, or electronic health records, or assist in the decision making processes, based on reasoning frameworks. The work in this thesis goes beyond the state-of-the-art of currently proposed information extraction systems. It employs Argument Mining methods to extract and classify argumentative components (i.e., evidence and claims of a clinical trial) and their relations (i.e., support, attack). An Argument Mining pipeline is proposed and further enhanced to integrate additional information inspired by prevalent biomedical frameworks for the analysis of clinical trials. These extensions comprise the detection of PICO elements and an outcome analysis module to identify and classify the effects (i.e., improved, increased, decreased, no difference, no occurrence) of an intervention on the outcome of the trial. In this context, a dataset, composed of 660 Randomized Controlled Trial abstracts from the MEDLINE database, was annotated, leading to a labeled dataset with 4198 argument components, 2601 argument relations, and 3351 outcomes on five different diseases (i.e., neoplasm, glaucoma, hepatitis, diabetes, hypertension). Various Machine Learning approaches ranging from feature-based SVMs to recent neural architectures have been experimented with, where deep bidirectional transformers obtain a macro F1-score of .87 for argument component detection and .68 for argument relation prediction, outperforming current state-of-the-art Argument Mining systems. Additionally, a Proof-of-Concept system, called ACTA, was developed to demonstrate the practical use of the developed argument-based approach to analyse clinical trials. This demo system was further integrated in the context of the Covid-on-the-Web project to create rich and actionable Linked Data about the Covid-19.

Referência(s)