In this article I will try to point out some of the various ways that companies are building products to do medical records indexing and summaries and some of the pitfalls to watch out for.

First off I believe that indexing and summarizing medical records is a job for AI and not for humans moving forward. The speed, cost savings and accuracy that AI can give are tailor made for this type of work.

Here is the challenge. How were the models built and what data were they trained on?

To start, let me point out the difference between extractive and abstractive AI models. The best way I have heard to explain this, so I copied it, is an example of different ways to do a book report.

One way to do a book report is to give you a book and for you to tell me, in your own words, who the characters were in the book and what happened in the story along with how you felt or interpreted that book for yourself. This would be an abstractive way of doing the book report.

Another way would be to give you a highlighter and to tell you to highlight all of the most important parts of the book and then build me a chronology with only the extracts of what was highlighted. This would be an extractive report.

When it cones to medical records indexing and summaries what we don't want is for AI to make up things or give opinions when it comes to medical records so it so super important for these models to be trained on how to do the extractive method.

Here is the issue in that, where do you get the data. What you would need is literally thousands of sets of personal health information that is already indexed and summarized to train these models. Since many of the folks that are building these models don't have all of that data they are using a combination of some extractive and abstractive, and in many cases some human interaction, to build their indexes and summaries.

While these read well and often can have some level of insights pulled that were not actually spelled out in the record itself, they are inherently flawed in the area of clinical accuracy.

Because Gemini is a medical records retrieval company, and has been for 20 years, we already had hundreds of thousands of indexed and summarized medical records data to train AI with. Using purely extractive models JudyAI will accurately index and summarize medical records in minutes instead of days and at a much lower cost than many of the other providers out there.

Please reach out if you would like a free trial or to find out more about this service and Gemini is happy to help.