Duke Health leaders have selected seven high-potential innovation projects, two from Duke Anesthesiology faculty, to be implemented as part of the Duke Institute for Health Innovation’s (DIHI) annual call for project ideas in “Generative AI & Large Language Models: AI solutions to improve staff and clinician efficiency, patient journey and outcomes.” Selected proposals were deemed innovative, scalable and those that could have tremendous impact on Duke’s clinical enterprise.
The first grant award (one-year, $25,000), focusing on harnessing large language models (LLMs) to summarize intricate Intensive Care Unit notes, promises to enhance shift turnover communication, while the second grant award (one-year, $25,000), centered around retrieval augmented generation, aims to develop a conversational interface for all perioperative policies. Both initiatives are set to significantly improve health system operations and patient care, marking a pivotal step forward in medical technology.
The project titled, “HandoverHero - Elevating ICU Communication,” pioneered by Duke Anesthesiology’s Drs. Sachin Mehta (principal innovator), Michael Kent, Mihai Podgoreanu, Madhav Swaminathan, and Padma Gulur, entails the construction of an AI-powered assistant capable of summarizing selected clinician notes within the Cardiothoracic Intensive Care Unit (CTICU) to facilitate concise but accurate/comprehensive turnover communication at shift turnover and assist with prioritization of work. In addition to AI-driven summarization/semantic analysis, this system will provide the opportunity to explore feature extraction from clinical text within an ICU environment for utilization in existing ICU models.
“Within the high volume/pace environment of the CTICU, which is characterized by numerous complex data streams, critical daily attending turnovers are variable in content/quality, rely on “by hand” summarization, and lack consistent semantic stratification,” says Mehta. “This tool will solve the need for consistent, concise, and standardized communication between attending CTICU physicians who care for a dynamic critically ill population.”
The second project titled, “DUPLEX.AI (Duke Health Policy Understanding and Learning through Extraction with AI),” pioneered by Duke Anesthesiology’s Drs. Kent (principal innovator) and Gulur, with nursing collaborators Katy Clark, Erin Hamtil and Samantha Fiske, entails the construction of an AI-powered assistant capable of interpreting and responding to user inquiries regarding Duke’s policies and processes. Utilizing natural language processing (NLP), it will comprehend queries, retrieve relevant policy documents, and present them to users in a digestible format. Moreover, the AI assistant will handle multi-level inquiries, thereby providing detailed insights into specific policy aspects, ensuring healthcare providers have nuanced, in-depth data for informed decision-making.
“We are using AI tools, known as retrieval augmented generation, to generate a tool where we will collect all perioperative policies and use an AI model to create a conversation interface for providers and nurses. Instead of having to look up some random policy, we will have a chat interface where someone can ‘talk to’ or ‘ask questions’ of policies to get an immediate answer,” says Kent. “Our goal is for this project to serve as the template Retrieval Augmented Generation System for the Duke University Health System that can be applied across many other departments.”
Alongside providing funding for these projects, the DIHI team will collaborate with various Duke Health and Duke University partners to provide access to data, analytics, statistical analysis, as well as machine learning and AI resources. They will also oversee project and implementation management.