SURA: Generative AI to find new ways of managing health

Challenges  

“The information in the genomic reports comes in very heterogeneous formats. So, searching for something in them would require a lot of time and dedication. That’s where we found a solution together with our partner Datalytics. We decided to use generative models to extract information from the PDFs easily and transfer it to a standardized database,” explains Catalina Bustamante, head of technology at SURA’s Omics Science Center.

From this situation, the main challenges were:

• Processing large volumes of data.

• • Interoperability: it was necessary to combine genetic reports with medical history data.
  • The existence of a large amount of unstructured but highly relevant information.

• Difficulty in interpretation.

Strategy: What did we do?

The goal was to automatically extract information from more than 10,000 PDFs based on the variables we needed and compile it into a standardized report.

“Searching these databases requires technical knowledge of genetics, which is uncommon among clinical staff. However, they do have a clear understanding of what kind of questions to ask and the terms to use. In this sense, we designed a bot using generative AI that allows natural language queries to be made to the database,” adds Bustamante.

Together with the SURA team, Datalytics developed a solution based on a private instance of OpenAI, fed with information from patient genetic reports. Moving forward, we aim to integrate medical history, radiology, and other data.

It is a chat —private and restricted within Microsoft Teams— that uses generative AI to query data. Physicians interact using natural language, which facilitates access to and understanding of genomic data.

To achieve this:

• We standardized the data: We used a strategy powered by generative AI that normalizes the types of data received and helps machines interpret them. From there, genetic data could be queried just like any medical history.
• We used GPT to access the extracted, curated information available in a database. This allows natural language questions, such as: “What are the patient’s family medical histories?” GPT not only provides easy access to all available information but also summarizes it without adding potentially erroneous interpretations.

Physicians who are not specialists in genetics can ask questions naturally, such as: “How many patients have X variant and also suffer from X pathology?” or “How many patients have a parent younger than 40 with X disease?”

GPT is crucial because it spares geneticists from having to run queries manually, as it translates these questions into a query for the database. Once done, the AI agent constructs text that delivers the response to physicians in natural language.