Large Language Models in Emergency Medicine

Dr. Hopkins's presentation focused on the rise of ambient medical AI scribes and other uses of LLMs in Emergency Medicine.

In his presentation, Dr. Hopkins discusses the scoping review article by Preiksaitis et al., which systematically reviews the broader applications of LLMs in emergency medicine (EM). The article highlights the potential of LLMs to transform EM by enhancing clinical decision-making, optimizing workflows, and improving patient outcomes.

The review identifies four major themes for LLM applications in EM:

  • Clinical Decision-Making and Support: This includes using LLMs for tasks like triage, diagnosis, treatment recommendations, and risk stratification.
  • Efficiency, Workflow, and Information Management: LLMs can aid in charting, summarizing medical records, extracting information from unstructured data, and identifying patterns.
  • Risks, Ethics, and Transparency: The authors emphasize the importance of addressing potential biases, ensuring data privacy, and developing transparent and accountable AI systems.
  • Education and Communication: LLMs can be used for physician training, patient education, and even generating scientific content.

He also discusses medical AI scribes and the potential legal and ethical challenges they present, particularly in the context of a review article from the Canadian Medical Protective Association (CMPA). He highlights the CMPA's concerns about AI scribes, emphasizing the potential for "hallucination" of information, introduction of biases, and the necessity for clinician review of AI-generated notes for accuracy and completeness before inclusion in medical records. He underscores that physicians are ultimately responsible for the information documented in patient charts, regardless of the technology used. Errors in AI-generated notes, such as missing or incorrect information, could lead to serious consequences, including hospital and college complaints, human rights complaints, and legal action.

Further concerns addressed by Dr. Hopkins and the CMPA include:

  • Data Privacy: The transmission of patient data to external servers, potentially outside of Canada, raises concerns about data security and encryption standards.
  • Consent: The use of AI scribes requires proper patient consent, especially if audio recordings of encounters are being stored. Separate consent forms may be required, similar to those used for audio or video recordings.
  • Data Retention: Clear guidelines are needed regarding the duration of recording retention and the process for their destruction.

Dr. Hopkins also shared his own experience developing an app prototype to better understand the technology. He describes a process involving audio recording of the patient-physician interaction, followed by transcription and analysis by a large language model (LLM) to extract key information and format it as a clinical note.

He demonstrates the app's impressive capabilities, including:

  • Generating Comprehensive Clinical Notes: The app can create detailed notes based on patient encounters, including the patient's ID, chief complaint, history, physical exam, assessment, and plan.
  • Customizable Note Formats: Physicians can adjust prompts to specify desired sections and information order within the clinical note, tailoring it to their individual preferences.
  • Differential Diagnosis Support: Scribbler can suggest potential diagnoses based on the patient's presentation.
  • Investigation and Treatment Suggestions: The app can propose relevant investigations and treatment options based on the patient's context.
  • Patient-Friendly Discharge Documents: Scribbler can generate discharge summaries in plain language, including reasons for ED return, tailored to different reading levels and languages.

However, Dr. Hopkins acknowledges the importance of "Retrieval Augmented Generation" (RAG) to improve the accuracy and reliability of AI scribes. RAG involves providing the LLM with access to relevant external information, such as the patient's medical history or up-to-date medical literature, to augment its knowledge base. This allows for more informed and accurate outputs, especially for complex or rare cases.

    Both Dr. Hopkins's presentation and the scoping review article emphasize the transformative potential of AI scribes and LLMs in healthcare, while also acknowledging the need for careful consideration of the associated ethical and practical challenges. Further research and development are crucial to harnessing the power of these technologies to improve patient care and optimize physician workflows.

    Back to top