EVA - Emergency Virtual Assistant

• 

  Methodology

Inspiration

Emergency departments (EDs) serve as vital components of healthcare systems worldwide, providing critical care to millions of individuals every minute. In the United States alone, the year 2021 witnessed 139.8 million ED visits, with 40.0 million attributed to injury-related cases. Despite the essential role of EDs, overcrowding remains a pervasive issue, contributing to increased mortality rates, compromised quality of care, and staff burnout. This phenomenon is driven by various factors categorized as input, throughput, and output challenges, highlighting the need for effective strategies to enhance patient throughput and optimize treatment delivery.

Triage, a fundamental process in ED management, aims to prioritize patient care based on clinical urgency. However, the efficacy and safety of existing triage scales remain uncertain, necessitating further research to validate their utility in real-world settings. Additionally, EDs in the United States have experienced a notable increase in admissions over the past decade, underscoring the importance of ensuring adequate resource allocation to meet growing demands.

A number of triage scales based on patients’ vital signs and clinical complaints have been developed to facilitate effective triage. Most scales have five levels of priority, ranging from immediate assessment to non-urgent and examples include the Manchester Triage Scale, Canadian Triage and Acuity Scale. Emergency rooms face prolonged waiting times, prompting the development of applications that provide real-time updates on hospital wait times. This issue is exacerbated by a high patient-to-nurse ratio, with nurses spending a significant amount of time on patient triage, particularly for straightforward cases. To alleviate this burden and allow nurses to focus on more critical patients, there is a growing interest in automated triage systems incorporating conversational agents (chatbots) and decision-making classifiers.

The inspiration for EVA (Emergency Virtual Assistant) stemmed from the urgent need to optimize emergency room (ER) operations, where every second counts. Recognizing the potential for artificial intelligence to enhance decision-making and efficiency, our team was motivated to create a solution that supports healthcare professionals in prioritizing patient care effectively. The Manchester Triage System, known for its structured approach to assessing urgency, provided a proven framework to integrate with AI technology, ensuring our solution is both innovative and grounded in established medical practice.

What it does

Introduction video Demo EVA leverages AI to assist ER staff in triaging patients more quickly and accurately. By adhering to the Manchester Triage System, our application analyzes patient data upon arrival—symptoms, vital signs, and preliminary patient information—to assign a priority level. This process helps streamline patient flow, reducing wait times and ensuring that critical cases receive immediate attention. The system also provides initial recommendations for tests and potential treatments, aiding medical staff in making informed decisions rapidly.

How we built it

Building EVA involved a cutting-edge approach centered around the integration of Gemini LLM, orchestrated with the use of advanced agents utilizing the LangGraph and LlamaIndex libraries. Our current system runs on Gemini Pro, which has already proven effective, but we anticipate that an upgrade to Gemini 1.5 Pro will exponentially improve the quality of our AI’s predictions and interactions.

To manage and interpret the vast amount of patient data effectively, we implemented a Knowledge Graph. This graph stores intricate details about user symptoms and their interconnections, allowing for rapid retrieval and analysis that supports accurate triage decisions. Additionally, our application employs Retrieval-Augmented Generation (RAG) to enhance the knowledge base of our AI agents. This technology helps in dynamically retrieving information to augment the AI's responses, significantly reducing the incidence of hallucinations and ensuring that the generated outputs are both relevant and accurate.

The integration of these technologies enables our system to provide a robust, intelligent solution that enhances the efficacy of ER operations by assisting in accurate patient triage and management. The orchestration of various AI components ensures that EVA can handle complex medical scenarios, improving both the speed and quality of emergency healthcare services.

Challenges we ran into

One of the major challenges was ensuring the accuracy and reliability of the AI under varied and unpredictable ER conditions. Balancing the sensitivity and specificity of the triage system to minimize both false positives and false negatives required meticulous tuning of our models. Additionally, integrating the AI system into the existing ER workflow without disrupting current practices posed a significant logistical challenge.

Accomplishments that we're proud of

We are particularly proud of developing an AI system that adheres to the rigorous standards of the Manchester Triage System. Our platform not only enhances triage accuracy but also reduces the cognitive load on ER staff, allowing them to focus more on patient care rather than administrative tasks. Successfully integrating advanced AI into a critical healthcare setting, and receiving positive feedback from preliminary tests in simulated environments, has been immensely rewarding.

What we learned

Throughout this project, we learned a great deal about the complexities of emergency medical care and the potential for AI

Built With

• geminiai
• langchain
• langgraph
• llamaindex
• neo4j
• pinecone
• postgresql
• prisma
• streamlit