MedCoderAI

Problem & Motivation

There are over 135,000 medical billing coders in the US¹ who transcribe medical notes into insurance billing codes.  It's an extremely complex task, that requires reading vast amounts of documentation and then mapping it to one of 7000 PDCs... which results in an estimated 2.5 million medical coding errors each year.
 

Data Source & Data Science Approach

Data sourced from https://physionet.org/content/mimiciv/2.2/

Previous approaches:
Truncate discharge notes

Our approach:
Summarized the discharge notes with a LLM, to avoid data loss from truncation.
 

Evaluation

Precision and Recall / True Positive Rate and Sensitivity

Key Learnings & Impact

New approaches and tools can have a huge impact!

This is about using LLM to summarize and using the encodings instead of text.

Acknowledgement

Lots of thanks for our amazing professors Danielle Cummings & Fred Nugen whose insight, advice and guidance helped turn our project into a reality.

Notes

1.  https://www.zippia.com/medical-biller-coder-jobs/demographics/