Clinical Sentiment Classification on Intensive Care Unit Clinical Notes

Abstract

Doctors' notes reflect their impressions, clinical sense, and opinions on patients' progress. This information is not documented elsewhere in patients' health records and is essential for doctors' daily clinical decisions. Decisions in critical care require an overall perception of the patient's clinical progress. Sentiment analysis tools are valuable for capturing judgments or perceptions from text. Despite this value, the literature indicates a lack of sentiment analysis methods that can function effectively on intensive care unit (ICU) clinical texts. Patients' privacy and confidentiality requirements have restricted the availability of open clinical text, contributing to this issue. Moreover, it remains unclear which of the several available approaches should be adopted to develop an effective sentiment model for ICU notes. Moreover, developing automated tools or AI models for analysing clinical notes requires involving clinicians in the process for validation by comparing machine output against clinicians' sentiment labels. Because of the difficulty of labeling, there are no public datasets of ICU notes labeled with sentiment classes, which is a significant issue hampering the development and evaluation of machine learning methods. The absence of sentiment classifiers which are able to distinguish polarity or more formally 'clinical sentiment' of ICU clinical notes has led to missing the value of doctors' impressions in these notes.

The aim of this thesis is to investigate which sentiment analysis method is most effective in recognizing clinical sentiment in ICU clinical notes. Clinical sentiment is clearly defined and specified for the task, avoiding the ambiguity and inconsistency found in related work. The thesis provides a comparison framework by which the effectiveness of different sentiment methods are evaluated, exploring which approach to sentiment and which model provides the best classification. In response to lack of labeled data, the thesis contributes an automatic labeling strategy, based on clinical events, for creating training samples, and a ground truth dataset labeled by clinicians for evaluation. In experimentation, different methods, models, domains, notes' content, and machine learning techniques are investigated.

The major contributions of this thesis are a sentiment ICU model, which demonstrates excellent capability in performing clinical sentiment classification; and the first ground truth clinical sentiment dataset to support future work on this problem. This dissertation addresses a research gap that has been highly under-researched. Its contributions support clinical linguistics, NLP applications and ultimately clinicians' decisions and healthcare practices.