2 months ago
NEW YORK, March 11, 2024 /PRNewswire/ -- An artificial intelligence instrumentality tin person doctor-written notes that summarize patients’ infirmary visits into accurate, laic language, a caller study found.
The investigation focuses connected discharge notes written by doctors to seizure patient’s wellness position successful nan aesculapian grounds arsenic they are discharged from nan hospital. Effective summaries are basal for diligent information during these transitions successful care, but most are filled pinch method connection and abbreviations that are difficult to understand and summation diligent anxiety, opportunity nan study authors.
To reside nan problem, NYU Langone Health has been testing nan capabilities of a shape of artificial intelligence (AI) called generative AI, which develops apt options for nan adjacent connection successful immoderate condemnation based connected really billions of group usage words successful discourse connected nan internet. A consequence of this next-word prediction is that nan specified generative AI “chatbots” person go bully astatine replying to questions successful realistic, elemental language, and astatine producing clear summaries of analyzable texts. However, AI programs, which activity based connected probabilities alternatively of “thinking,” whitethorn nutrient inaccurate summaries and truthful are meant to assist, not replace, quality providers.
To research generative AI, NYU Langone Health successful March 2023 received entree to GPT4, nan latest instrumentality from OpenAI, nan institution that created nan celebrated chatGPT chatbot. NYU Langone Health licensed 1 of nan first “private instances” of nan tool, which freed hundreds of its frontline clinicians to research pinch AI-based solutions to objective problems utilizing existent diligent data, while adhering to national standards that protect nan diligent privacy.
One of nan first studies by researchers utilizing GPT4, publishing online March 11 successful JAMA Network Open, looked astatine really good nan instrumentality could person 50 diligent discharge notes into patient-friendly language. Specifically, moving discharge notes done generative AI dropped nan reports from an eleventh-grade reference level connected mean to a sixth people level, nan golden modular for diligent acquisition materials.
The squad besides classed nan AI discharge study translations utilizing nan Patient Education Materials Assessment Tool (PEMAT), which generates a percent people based connected 19 factors connected nan expertise of patients to understand immoderate portion of reference material. GPT-4 translator raised PEMAT understandability scores to 81 percent, up from 13 percent seen pinch nan original doctor-written discharge reports from nan aesculapian record.
The investigation squad designed nan study to look astatine AI capacity by itself arsenic a technological question: How acold could it spell independently erstwhile translating discharge reports?
“GPT-4 worked good unsocial pinch immoderate gaps successful accuracy and completeness, but did much than good capable to beryllium highly effective erstwhile mixed pinch expert oversight, nan measurement it would beryllium utilized successful nan existent world,” says elder study writer Jonah Feldman, MD, aesculapian head of Clinical Transformation and Informatics wrong NYU Langone Health’s Medical Center Information Technology (MCIT ) Department of Health Informatics. “One attraction of nan study was connected really overmuch activity physicians must do to oversee nan tool, and nan reply is very little. Such devices could trim diligent worry moreover arsenic they prevention each providers hours each week successful aesculapian paperwork, a awesome root of burnout.”
To measurement nan accuracy of nan AI instrumentality translations, nan authors besides asked 2 physicians to reappraisal nan AI discharge summary for accuracy based connected a 6-point scale. The reviewing physicians awarded conscionable 54% of nan AI-generated discharge notes nan champion imaginable accuracy rating. They besides recovered that conscionable 56% of notes created by AI were wholly complete. These results, however, must beryllium considered successful context, opportunity nan authors. For instance, they say, nan results signify that, moreover astatine nan existent capacity level, providers would not person to make a azygous alteration successful much than half of nan AI summaries reviewed.
Feldman notes that generative AI devices are sensitive, and asking a mobility of nan instrumentality successful 2 subtly different ways whitethorn output divergent answers. The accomplishment required to framework nan questions asked of chatbots successful a measurement that elicits nan desired response, called punctual engineering, combines intuition and experimentation. Physicians and nurses, pinch their heavy knowing of individual cases and nuanced aesculapian contexts, are champion positioned to technologist prompts, opportunity nan authors, and without learning to constitute machine code.
Within weeks, nan investigation squad will beryllium launching a programme interviewing patients waiting to beryllium discharged whether AI-generated reports are clear and adjuvant aft expert review. By nan summer, nan squad expects to motorboat a aviator programme to merge GPT4-generated, physician-reviewed laic connection discharge summaries to patients connected a larger scale.
“Having much than half of nan AI reports generated being meticulous and complete is an astonishing start,” says first study writer Jonah Zaretsky, MD, Associate Chief of Medicine astatine NYU Langone Hospital—Brooklyn. “Even astatine nan existent level of performance, which we expect to amended shortly, nan scores achieved by nan AI instrumentality propose that it tin beryllium taught to admit subtleties.”
Along pinch Feldman and Zaretsky, NYU Langone study authors were Jonathan Austrian and Yindalon Aphinyanaphongs from nan MCIT Department of Health Informatics, Jeong Min Kim and Saul Blecker, and Department of Medicine, Division of Hospital Medicine; Yunan Zhao from nan Department of Population Health; Samuel Baskharoun from nan Department of Medicine astatine NYU Grossman Long Island School of Medicine, and Ravi Gupta from NYU Langone Health’s Long Island Community Hospital.
Contact: Gregory Williams, [email protected]
View original contented to download multimedia: https://www.prnewswire.com/news-releases/ai-model-makes-hospital-notes-patient-friendly-302085555.html
SOURCE NYU Langone Health System
