TN HIMSS Hackathon Generative AI

Healthcare professionals have long struggled with cumbersome prior authorization (PA) procedures, leading to significant delays in patient care.

The American Medical Association (AMA) recently published compelling statistics revealing that 94% of physicians face delays, 88% describe the PA burden as unreasonably high, and many physicians and their staff spend an average of 14 hours each week completing PAs.

The healthcare industry needs an innovative solution to streamline and optimize this cumbersome PA process. The current inefficiencies result in delayed patient care and contribute to increased burdens on healthcare professionals.

To solve this problem, we theorized that we could build a solution incorporating an open-source large language model (LLM) trained on clinical data to write individualized, high-quality PA letters to dramatically accelerate the process of PA letter creation. To put this theory to the test, this is how we approached the problem:

We decided to build a software solution that generates high-quality PA letters in seconds with just a single click.  Our approach prioritizes the use of open source components. The goal of using open source was to showcase how healthcare organizations, regardless of their size, can effectively harness the power of AI.

To build our solution, we identified the primary components of the project:

1. An open-source AI language model​
2. A simulated EHR database​
3. An intuitive user interface
4. Secure and HIPAA compliant

We began the project by identifying the model we wanted to use for our generative AI. We started at Hugging Face, a machine learning and data platform that helps users build, deploy, and train machine learning models. Within Hugging Face, we identified LLAMA 2 as the best model for our generative AI project.

LLAMA 2 is a sizeable open-source language model renowned for its versatility and accuracy in generating text-based content. We created an on-premises instance of LLAMA 2. We used our in-house, high-performance AI server to fine-tune it. We employed medical terminology and anonymized transcribed physician and patient conversations to improve inference quality. This fine-tuning allowed us to provide segmented tuning across medical fields. Then we incorporated prompt engineering to aid the model in creating the layout, content, and tone of the letters generated.

We wanted our patient encounter data to closely mimic the real world. To achieve this, we utilized data from SyntheticMass, a source that provides simulated healthcare data that mirrors the demographics and healthcare visits of the state of Massachusetts. This Electronic Medical Records (EMR) data was loaded into an Azure SQL database.

In real-world healthcare settings, most PA letters are still manually written. The process includes sourcing multiple data pages that live within different locations. To streamline this time-consuming process, we wanted to create a simple interface with a login screen, a work queue of prescriptions requiring prior authorization, and an area to edit and approve the generated letters.

To securely keep patient healthcare data private, healthcare companies must follow strict HIPAA compliance standards. As long-time consultants to many large healthcare organizations, our team knows how to develop software that adheres to HIPAA requirements. For this project, we implemented multi-factor authentication, encrypted all data in transit and at rest, and established secure HTTPS service calls between the Azure and AWS environments.