Deep-Learning Model Assists Researchers in Obtaining Useable EHR Data

With the goal of identifying a more efficient method of deciphering clinical notes, researchers from MIT described how a deep-learning model helped clean up her data, leading to information extraction and more accurate analyses of patient data.

EHRs play a distinct role in healthcare, assisting providers in several ways. Ten years ago, the US government decided to encourage the adoption of EHRs, as it believed that these systems would lead to improvements in care, the press release notes.

Even though encouraging EHR use aimed to ease provider burdens, it became clear that data contained in EHRs are often disorganized and rife with jargon and abbreviations that many cannot easily comprehend.

This issue, along with the goal of creating a single model to extract information, which can work well within various hospitals and learn from limited amounts of labeled data, led MIT researchers to act. Affiliated with the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), the researchers believed that large language models could proficiently extract and analyze EHR data. Thus, they used a GPT-3 styled model to perform those tasks.

“It's challenging to develop a single general-purpose clinical natural language processing system that will solve everyone's needs and be robust to the huge variation seen across health datasets. As a result, until today, most clinical notes are not used in downstream analyses or for live decision support in electronic health records. These large language model approaches could potentially transform clinical natural language processing,” said David Sontag, PhD, MIT professor of electrical engineering and computer science and principal investigator in CSAIL and the Institute for Medical Engineering and Science, in a press release.

Researchers are increasingly applying large language models (LLMs), like GPT-3, to clean data by expanding jargon and acronyms and identifying medication regimens. The MIT team created a small dataset to evaluate the extraction performance of LLMs. They found that these models could achieve 86 percent accuracy at expanding acronyms within data without any labels.

The GPT-3, which is the model used by the team, works by inputting a clinical note, which results in a "prompt," that is a question about the note that the system answers, according to the press release. This extraction and cleaning of data can enable personalized clinical recommendations.

“Prior work has shown that these models are sensitive to the prompt's precise phrasing. Part of our technical contribution is a way to format the prompt so that the model gives you outputs in the correct format,” says Hunter Lang, a PhD student at CSAIL and author on the paper, in the press release.

Further, the team intends to expand the model to cover languages other than English, add methods for quantifying uncertainty in the model, and work to achieve similar results with open-sourced models.

“AI has accelerated in the last five years to the point at which these large models can predict contextualized recommendations with benefits rippling out across a variety of domains such as suggesting novel drug formulations, understanding unstructured text, code recommendations or create works of art inspired by any number of human artists or styles,” said Parminder Bhatia, head of machine learning for low-code applications leveraging large language models at AWS AI Labs, in the press release. Bhatia was not involved in the research.