Since the last few years, the clinical NLP community has worked incredibly hard to extract crucial variables hidden in digital records. This idea of digitizing medical records came up because of a law that the U.S. government approved almost ten years ago to modernize the healthcare industry. These computerized records contain a vast amount of valuable information outside the purview of clinical studies, such as the ideal dosage of a given medication for patients within a specific body mass index (BMI) and genetic history, among other things.
Training several machine learning models is necessary for information extraction. However, the jargon and acronyms used in these doctor’s notes make it extremely difficult for computers to comprehend using the currently available approaches. Furthermore, distinct models are needed for various hospitals and doctors, and training each model involves diligent annotation work, which is a time-consuming and expensive procedure.
Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) turned to large language models to build one ideal system that could cater to multiple hospitals without requiring large amounts of training data. To execute tasks like extracting medical regimens and expanding clogged-up jargon and acronyms, the researchers employed a large GPT-3-style model. Future clinical suggestions could be made more specific by using the extracted clean data.
Due to many limitations on using public resources for testing LLMs, procuring medical data for training purposes is an extremely challenging task in itself. A system that learns from a small amount of labeled data, works well at many hospitals, and uses a single model to extract multiple sorts of information would be an ideal model. However, developing a single, all-purpose clinical natural language processing system is incredibly difficult. The team curated their own dataset by incorporating smaller, freely accessible medical snippets.
MIT’s large language model approach will completely transform clinical natural language processing. Their research shows that, despite not being trained especially for the clinical domain, LLMs like InstructGPT perform well at zero and few-shot information extraction from clinical literature. This discovery also provides room for scaling since it allows for the rapid construction of new models, even for many use cases, without needing to manually annotate data. After thorough evaluations, the researchers concluded that their work could extend acronyms with 86 percent accuracy. Additionally, the team’s newly devised procedures increased this score’s accuracy to 90% without needing re-annotation.
An enormous amount of publicly accessible training data is used to teach LLMs to complete sentences and anticipate the next most likely word. Even if older, simpler models like earlier GPT versions or BERT performed well at collecting medical data, manual data annotation is still a labor-intensive process. Through their research, the team has eliminated the need to train separate machine learning models for various tasks, such as extracting medication and side effects from medical records and disambiguating common abbreviations. The researchers not only expanded abbreviations but also showed how to use LLMs to handle various NLP tasks that call for more structured outputs, such as relation extraction, token-level sequence categorization, and span recognition.
The team’s efforts to get the model to produce outputs with the right structure also greatly reduced the time needed for post-processing. However, one such limitation when it comes to this approach of analyzing healthcare data is in situations where a patient’s confidential information is communicated over the internet to an LLM provider like OpenAI. Making the model smaller so it may be utilized on-site is an alternative to this method.
In terms of future work, the group intends to broaden their study to include languages other than English and create new techniques for quantifying model uncertainty. To achieve comparable results, they also intend to leverage open-source models. Unstructured clinical notes in the healthcare sector present special issues as opposed to general-purpose text. This is due to the extensive usage of acronyms and the irregular linguistic patterns employed by various healthcare facilities. In this regard, the research conducted by MIT researchers presents a fascinating standard for utilizing the strength of large language models for several clinical NLP tasks.
Check out the Paper and MIT Source. All Credit For This Research Goes To Researchers on This Project. Also, don’t forget to join our Reddit page and discord channel, where we share the latest AI research news, cool AI projects, and more.
Khushboo Gupta is a consulting intern at MarktechPost. She is currently pursuing her B.Tech from the Indian Institute of Technology(IIT), Goa. She is passionate about the fields of Machine Learning, Natural Language Processing and Web Development. She enjoys learning more about the technical field by participating in several challenges.