MediCode

Inspiration

Payment errors in healthcare systems today make up more than $300 billion annually in unnecessary spending - about 10 cents to every dollar spent according to a 2019 report published in JAMA. ACA Marketplace Plans denied 17% of medical claims in 2021. Medical claim errors occur on a massive scale, causing lost revenue and big costs for healthcare provider offices and hospitals. Medical coding errors are a major reason for denied claims.

This inefficiency prevents hospitals from being paid for the services they provide and result in a poor patient and healthcare professional experience. The bulk of these errors come from incorrect medical coding, which is the process of extracting diagnostic and therapeutic codes by analyzing the clinical note, a process done by a combination of doctors and medical coders.

What it does

We built an AI copilot that enables a medical coder to input their clinical note and obtain:

• a summary with action items
• CPT/HCPCS (therapeutic) codes and ICD-10 (diagnostic) codes that can be cross-referenced against real codes (to prevent hallucination)
• citations from the clinical note for why those codes were chosen.

This assistant helps the healthcare professional determine the most accurate and comprehensive set of codes, so the claim made to the payor is not denied and to boot revenue.

How we built it

We built a React App with a chatbot that calls the GPT-4 API, and the infrastructure to cross-reference against codes from CMS to ensure high accuracy.

Challenges we ran into

Prompting can be finicky, and sometimes it is required to prompt multiple times to home in on the right answer. Eventually we were able to find a mix of prompts that provide decent results along with explanations that reference the original clinical note.

Accomplishments that we're proud of

Working copilot UI that leverages both LLMs directly and the infrastructure to verify these codes by cross-referencing the CMS CPT / HCPCS and ICD-10 codes.

What we learned

LLMs are great at identifying useful embeddings and summarization, currently we use it directly through prompting though training a direct classifier with these embeddings could improve performance with more clinical note data. Guided chatbot workflows that closely interact with the user and work with external systems,

What's next for MediCode

Integration with EHR systems that currently use this such as Epic Systems (which has a third-party developer program and has 70+% penetration in the hospital market) to better fit within current workflows. Automatically populate claim with medical codes determined by MediCode copilot.

Demo: https://www.loom.com/share/88edd5aa4dc845209c5af78a2df78eb6

Built With

• javascript
• openai
• python
• react