The Fifth Elephant 2025 Annual Conference CfP

Overview

• Electronic Health Records (EHRs) give clinicians decades of longitudinal data—labs, imaging, progress notes, doctor’s notes, device streams. This means large amounts of data being available for each patient, especially for those that with chronic illnesses.

• There is scope to help reduce the work load and offer better patient care by leveraging modern, state-of-the-art LLMs. Some examples are

  • Creating summaries of patient history
  • Categorisation and summarization of documents
  • Chat with patient records
  • Finding interesting diagnostics links through knowledge graphs
  • Building clinical decision support systems
  • Customizing summaries and reports for each specility, highlighting information that’s most relevant to a given doctor.

• But there are significant challenges as there is almost no margin for error. The output has to be right every time for doctors to confidently use the product. Any hallucination, missing information, or irrevalant inforamtion reduces the trust a doctor has in the tool. And doctors have no time for unreliable tools

• In this talk, we discuss some of the practical lessons learnt while building one such assistant.

  • Techniques for data augmentation
  • Using knowledge graphs
  • Retrieval techniques to ensure high relevance and accuracy
  • Chain of thought and self verification
  • Evaluation strategies to build confidence
  • Using compound AI systems
  • Guardrails and validations

Audience

Engineers using AI models to build critical systems that need to be right.

Speaker

I’m a software engineer with ~10 years of expirience. For the past 1 year, I’ve been studing AI models with the focus on building production grade applications using them. I’m interested in building relaiable, high quality products with AI that are critical and have a real world imact.