Large language models to analyze healthcare incidents

In a national project, Region Västmanland is now collaborating with, among others, the Region Västra Götaland and Region Skåne to take both technology and necessary data sharing to the next level.

The purpose of registering what are called incident reports (see fact box) is to improve patient safety. By identifying patterns among the incidents that occur, it becomes possible to take preventative measures.

The work already done in Region Västmanland focuses on errors related to medication. Approximately 400 people in the region have received training to be able to make this type of registration. However, the reporting is still done in a very heterogeneous way, and it has also been established that up to 30 percent of all medication incidents are missed.

Together with Mälardalen University, Anders Krifors and his colleagues chose to work with the pharmacists who have the deepest domain knowledge in the area.

"We succeeded in building an AI solution that can perform the classification with the same high quality as these experts. I think that is incredibly exciting, and it shows that the technology works. Hopefully, it can be expanded to more areas," says Anders Krifors.

This is the step that the participants in the project Regional collaboration on the development of AI tools in healthcare now want to take together. Magnus Kjellberg, Director of AI Competence Center at Sahlgrenska University Hospital and Head of AI office at the Region Västra Götaland, is one of the Region Västra Götaland’s representatives in the project.

"Healthcare incidents are a very important application case. But it also only scratches the surface of all the possible applications we have in healthcare for similar solutions. We have an endless amount of text throughout our organizations, and many cases where we could use this language ability not just for deviations, but also for other types of indicators, statistics, and more," says Magnus Kjellberg, continuing:

The AI models responsible for the classification are obviously an important part of the project. But making data accessible between the regions is also a central aspect. Incidents are not necessarily evenly distributed across the country's regions, and if data can be shared nationally, it increases the potential value.

Unlike many other healthcare-related AI initiatives, this project has an advantage: Completed incidents are public records, and thus differ from journal data. Furthermore, in the type of analysis the project is conducting, the identity of the patient involved is not relevant.

Lorna Bartram, Project Manager at AI Sweden, sees this as a great advantage:

More about Region Västmanland's project

Read the academic article about Region Västmanland's LLM project regarding medication incidents.

More about the project in the AI Sweden Podcast

Anders Krifors and Magnus Kjellberg joined the AI Sweden Podcast to share further insights into the project.