Postiglione, Marco (2023) Knowledge Graphs for Next-Generation Health Science Applications. [Tesi di dottorato]

Anteprima

Testo POSTIGLIONE_MARCO_36.pdf Download (10MB) | Anteprima

Abstract

The ongoing digitization of medical records, clinical charts, and health archives has markedly enhanced the accessibility of these critical documents, thereby ushering in a new era in the field of medicine. This data serves as a foundational cornerstone for the field of "precision medicine", whose primary objective is to elevate the standards of personalized diagnoses and therapies by harnessing the individualized attributes of patients, including but not limited to lifestyle factors, medical histories and genomic information. However, a notable challenge is that about 80% of healthcare data is unstructured, comprising textual elements like clinical notes and discharge summaries, and remains largely unexplored. Traditional Natural Language Processing (NLP) algorithms, when applied to clinical scenarios, have largely depended on shallow matching techniques, template-based approaches, and non-contextualized word embeddings. These approaches exhibit limitations in capturing nuanced contextual semantics. Although there have been significant advancements in the broader NLP domain through language models able to effectively leverage contextual information, many of these general-purpose NLP algorithms face challenges when applied to specific clinical NLP tasks that necessitate specialized biomedical knowledge, especially in low-resource languages where there is a lack of annotated datasets. This thesis delves into the multifaceted domain of few-shot learning techniques aimed at extracting information from clinical textual data. A pivotal focus is placed on data augmentation strategies and the amalgamation of multiple datasets into a unified model to enhance the learning efficacy. In this work, a novel representation of patients' medical histories is proposed through the introduction of Temporal Knowledge Graphs, which provide a structured framework for encapsulating chronological clinical information. Furthermore, a specialized model is developed, which utilizes recurrent units of Graph Convolutional Neural Networks (GCNs) to effectively harness the temporal dependencies within the data. This approach aims to anticipate potential health disorders a patient may encounter in the future, presenting a significant stride towards proactive healthcare management. The validation of the proposed framework is carried out on two distinct datasets: the publicly accessible MIMIC-III database and a private dataset furnished by the Department of Advanced Biomedical Sciences at the University of Naples Federico II. The latter offers a unique lens into clinical narratives compiled in Italian, thereby broadening the evaluation spectrum and demonstrating the capability of the framework in handling multilingual clinical text. Through rigorous evaluation, this thesis underscores the potential of harmonizing clinical notes with structured temporal data representation in advancing predictive healthcare analytics.

Downloads

Actions (login required)

Modifica documento