Using Generative AI to Effectively Triage

Problem Background

Effectively triaging patients is a key factor in improving access to world-class cancer care and reducing the administrative burden for staff and clinicians. Patient service representatives (PSR) staff, must review detailed, complex, and lengthy triage instructions. Reconciling these important, but often times difficult to understand triage instructions, can result in mismatched routings of patients, leaving patients and their families frustrated with delays and physicians questioning why patients were routed to them. When PSR staff have questions on a triage decision, this is often escalated to a nurse navigator and/or a clinician, leading to additional delays.

Proposed Solution

GuideMyTriage (GMT) addresses this issue by providing PSR staff with a tool that they are able to communicate with directly allowing them to ask questions about the proper way to triage a patient. Interactions are performed in a human centric approach and allows quick guidance from carefully and continuously curated decision trees for triaging.  The results – patients are quickly scheduled for the right clinic for the right time without the need to escalate to a nurse manager or physician.

Healthcare Innovation CIC Takes on Initial Prototyping

The Healthcare Innovation CIC worked closely with a leading medical center in the US to develop an impactful prototype highlighting how this triage process could be reimagined.

We worked closely to understand the goals and requirements of the proposed effort and framed three potential use cases to tackle:

  1. Intake & Triage – When a patient is referred, a Patient Service Representative (PSR) is responsible for triaging the patient to the appropriate disease service line (DSL) to schedule an appointment. They rely on complex guidelines that can be confusing for the PSR to follow. The PSR might schedule the patient with the incorrect DSL, taking up a precious appointment slots and creating delays for the patient and provider. The use case is for the PSR to chat with a chatbot after receiving and reviewing the referral, to ensure the patient is triaged to the correct DSL to schedule an appointment within 24 hours, and for the patient to be seen within 7 business days.
  2. Rapid Access Clinic Care Guidance – A chatbot that can provide Rapid Access Clinic (RAC) Advanced Practice Providers (APPs) guidance on the right set of studies/tests (labs, imaging, etc.) to order and follow-up appointments and referrals to request. The APP in the RAC is that first touch point for many patients seeking timely care. This chatbot will provide the APP guidance on the first steps in care for these patients entering the RAC.
  3. After-Hours Operator – Outside of regular business hours (M-F 9AM-5PM), a chatbot that helps the operator direct the patient to the correct staff or provide the operator with correct information to relay to the patient. The current problem is that operators, after hours, on weekends and holidays, have directly connected the patient to clinical staff such as medical oncology fellows who are over-burdened by administrative work calls that could have been completed by office staff instead of a clinician (e.g., notifying patients about upcoming appointment details, lab results, etc.). Administrative work where the patient could have simply been instructed by the operator to call the clinical office the next business day.

Our innovation center decided to take on prototyping for use case #1 and utilized generative AI technologies in correspondence to a knowledge domain of procedure documents to prove out the feasibility of the concept.

We evaluated several Large Language Models (LLM’s) with various prompts including prompts that displayed reasoning of choices. The following demonstrates a correct routing choice the bot was able to achieve based on procedure documentation within the bot’s knowledge base.

Additionally, we validated that the bot would prompt the user for more information if it was unclear of where to route the patient.

Call to Action

Interested in integrating the concepts of this challenge into your production workloads? Contact us and we can discuss how AWS Cloud and the AWS Partner Network can help accelerate your goals.

Artifacts

DeliverableDescriptionLink
PRFAQFictional press release for the GuideMyTriagehttps://github.com/UC-CIC/ccc-docs/blob/main/ccc.prfaq.pdf
Use case summary & architectureArchitecture diagram & use casehttps://github.com/UC-CIC/ccc-docs/blob/main/ccc.usecase.pdf
Prototype (Technical)Technical prototype leveraging AWS Gen AI LLM Chatbot Acceleratorhttps://github.com/UC-CIC/aws-genai-llm-chatbot-ccc/tree/ccc-mod