# AI Frontiers: Annie Hill (OpenAI DevDay)

## Метаданные

- **Канал:** OpenAI
- **YouTube:** https://www.youtube.com/watch?v=wCJejxhpz-w
- **Дата:** 15.11.2023
- **Длительность:** 7:53
- **Просмотры:** 14,083
- **Источник:** https://ekstraktznaniy.ru/video/11560

## Описание

Annie Hill is the Sr. Manager, Innovation & Digital Health Accelerator at Boston Children's Hospital

Boston Children’s Hospital is using GPT-4, function calling and retrieval across a number of projects to improve hospital operations, reduce administrative burden, help healthcare professionals access information more efficiently, and catch errors that could lead to issues in patient care.

## Транскрипт

### Segment 1 (00:00 - 05:00) []

-Hi, everyone. My name is Annie. I'm from the Innovation & Digital Health Accelerator group at Boston Children's Hospital. For those of you who don't know, Boston Children's is a world leader in providing compassionate, equitable, and family-centered care. We are ranked among the top pediatric hospitals in the country, and we are the primary pediatric teaching hospital for Harvard Medical School. I'm from the Innovation & Digital Health Accelerator group. We were founded about seven years ago as part of the hospital's commitment to digital health innovation. We work hard to make sure that pediatric healthcare has a seat at the table when it comes to the intersection of healthcare and technology. We have three different pillars that form the foundation of our work. Each of those pillars is responsible for a different function, but we really feel that generative AI sits at the intersection of what our team does. We've taken a collaborative approach to really understand where there are opportunities to leverage this technology, both at the enterprise and in the broader healthcare ecosystem. We see a lot of opportunities for this technology to deliver value to healthcare organizations. I'll just touch on a few of those today. Burnout is a huge factor for many healthcare organizations, and we are particularly interested in applications of generative AI that can reduce burden and really allow employees to focus on job roles and functions that they find fulfilling. As a healthcare organization, we are also interested in supporting our clinical teams and being as effective as possible in delivering high-value care, and supporting our patients through the patient experience. I'll also note that as an institution, we are committed to equity, diversity, and inclusion, so we are actually working on building an LLM implementation equity guideline that will really help us think about equity principles and EDI in everything we do related to AI. That includes things like research and some of our development efforts. When we started thinking about how generative AI could make an impact at the hospital, we started where we usually start these efforts, and that is with the experts on hospital pain points and challenges, our staff. We talked to our staff to try to understand where they were experiencing challenges, and then we matched that to where we understood this technology can make a difference, and where it aligned with our strategic priorities and some of those value propositions I mentioned on the last slide. We collected a number of pain points from across the hospital. We organized them into three buckets, operational challenges, clinical challenges, and research challenges. On the operational side, it's a bit of a catchall bucket because there's a lot that goes into operating a hospital system. One thing I do want to mention that I thought was interesting is this pain point around patient education. We develop a lot of patient education, and it is so important for us to be able to offer that education in not only different languages, but at different reading levels as well. We're a children's hospital, we have a wide audience, and so anything that can help us in being more efficient in translating that content to different scenarios would make a big impact. On the clinical side, we're not really looking at pain points and use cases that are true clinical decision-making at this point. We're more exploring opportunities to support our clinical teams, reduce burden, and really put the right information at their fingertips when they need it. Lastly, on the research side, a lot of the pain points we heard were around working with large data sets, particularly qualitative data. It's very time-consuming. Think interview data as part of a clinical study. The use cases we're looking at there are about using LLMs as a first pass analysis for those large qualitative data sets. Something that our group does every time we explore a digital solution to a particular pain point is consider, what can we leverage out of the box, and in what cases do we actually need to develop something ourselves? We've taken that same approach here and we're really excited to be exploring actual development of some LLM-powered applications, one of which is MedTutor. MedTutor is a medical tutor that will aim to guide healthcare learners, so medical students, residents, fellows, through a medical case. Medical cases form the basis of our current healthcare education system, but they're not personalized to the needs of the learner. They're not interactive in their current form, and they're not readily accessible. MedTutor is aiming to address some of those challenges. The model will ultimately be fine-tuned using hospital data that we're currently using for our medical education purposes, and we're planning to start pretty narrow, so generating a handful of cases to a particular disease area. Then the interactive platform will actually walk students through the case as if a mentor or a small group were walking them through it, adjusting responses based on what the student is saying in response. We expect this will not only increase learning opportunities

### Segment 2 (05:00 - 07:00) [5:00]

but we actually hope it will increase the quality of learning because the experience is tailored to the needs of the student. We're also looking at ways that LLMs can enhance the existing tools that we've developed at the hospital. Swirl was developed by a group of innovators who are affiliated with Children's, and it is a provider facing system that ingests data from different clinical sources, so think notes, labs, medications, orders, way-forms, and it aggregates that into a comprehensive real-time patient view. The team is looking at LLM integration to see if there's a way to allow providers to ask patient specific questions and receive answers that aggregate data from across the patient record. Providers were spending a lot of time looking in different platforms, looking across systems to try to find the most relevant information, and that can take a lot of time. So think, if a provider was seeing a new patient, they might need to know who was seeing this patient in the past? What medications have been prescribed? How have those medications performed? All that information requires digging. With this integration, a GPT-4 agent can actually build context across the patient record and provide responses that include information from different sources and then even link out to those sources for further validation. We expect that this will reduce the amount of time providers are spending searching for information. It'll put the information they need at their fingertips when they need it, which will allow them to provide higher quality care. The other integration opportunity we're looking at with Swirl is context-driven error detection and alerting. Error detection, obviously very important in healthcare, but traditional systems can bombard providers with irrelevant alerts and that can lead to alarm or alert fatigue. We are looking at-- and Swirl is aggregating patient data, so with this integration, the patient record could actually be monitored and potential errors could be identified with improved ability for evaluating alert accuracy. For example, this integration could allow the platform to recognize that a medication that was previously prescribed is no longer safe for a patient because of a new or recent medical event. Context-driven error detection can not only streamline the workflow for providers, but more importantly, it can improve patient safety and really reduce clinical errors, which is so important. These slides highlight just a few of the ways that Boston Children's is thinking about genitive AI. We are very excited about some of the announcements that came out today and we're eager to see how this technology continues to advance and we'll continue to evaluate and really try to understand the impact that this technology can have on healthcare not just today, but also in the future. Thank you.