Automating disease diagnosis and cause-of-death classification from medical narratives using event extraction and temporal ordering
Year: 2021
Type: dissertation
Abstract: The digitization of health records has provided a wealth of data that can be used for machine learning tasks such as automated diagnosis of illness or cause of death. However, crucial contextual information such as time and duration have often been omitted from such models. Although the words and phrases describing time can be modeled with semantic representations such as word embeddings, these representations often fail to capture the temporal value of such time phrases. The goal of this work is to present models for identifying symptoms and time information in the text of different types of health documents (while explicitly modeling temporal value), determining the chronological order of the extracted symptoms, and classifying the documents according to diagnosis or cause-of-death. Since the style and content of health text documents can be widely variable, these models are evaluated on three different types of health records: clinical notes (formal written health records), verbal autopsies (informal written records of deaths that occurred outside of health facilities), and transcripts of clinical conversations (verbal dialogues between medical staff, patients, and caregivers). We conduct end-to-end disease classification using neural network models trained on free-text narratives. The system includes supervised learning models for event and time extraction, listwise temporal ordering of events, and classification. We find that classification models that include event timeline representations perform better than methods that use only the raw text of the narrative. Furthermore, we present several models and metrics for listwise temporal ordering, including the ability to model simultaneous events. These models can be applied to different types of medical narratives, and can improve the performance of automated classifiers which can be used in real-world contexts to improve cause-of-death coding efficiency and clinical decision support.
Source: TSpace (University of Toronto)
Author: Serena Jeblee
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3178630086