An interface for clinicians: finding crucial information with language models in electronic health records

Abstract

Machine learning is often employed to solve tasks autonomously, but it is frequently necessary and more useful to aid domain experts rather than replacing them. Using machine learning to provide aid in specialized domains has three primary challenges: 1) identifying useful information, 2) providing trustworthy information, and 3) obtaining supervision for training. Pretrained Language Models (PLMs) can address these challenges by providing mechanisms of interpretability through their use of language and making efficient use of distant supervision. This thesis explores using PLMs in the context of the unstructured text and images in Electronic Health Records (EHRs) and takes the approach of creating models with varieties of interpretability in service of the primary goal of surfacing relevant information to aid clinicians. Importantly, we only assume access to data naturally present in the EHR, relying on PLMs to make use of this data without any manual annotations. This work starts with looking at the use of internal mechanisms to point to information of clinical relevance in the raw inputs, first in medical record text and subsequently in images in a multimodal setting. The second part focuses on using the zero-shot capabilities of recent Large Language Models to extract structured information from unstructured notes for the purposes of interpretability, efficiency, and supervision. We conclude by emphasizing the viability of using Pretrained Language Models as an interface to Electronic Health Records that can identify information to help clinicians avoid mistakes and become more efficient.--Author's abstract

Ähnliche Arbeiten