Named Entity Recognition for Medical Texts

Overview

Swedish patient records hold large amounts of information locked in free text — hard to search, even harder to aggregate. Turning this unstructured information into useful data is a real challenge.

But that's where a project comes in, carried out by two of our current employees, with the goal of efficiently extracting valuable data from Swedish patient records using Named Entity Recognition (NER).

In the project, several BERT models and data augmentation techniques were evaluated, with the potential to significantly improve NER results on Swedish patient records.

Results

• Data augmentation significantly improved system performance, especially on smaller datasets.
• Augmenting 50% of the training data gave results comparable to using the full original dataset without augmentation — halving the need for annotated data.

This project shows how the right technology and methods can help us extract valuable information from Swedish patient records, which can contribute to a more enlightened and data-driven healthcare.

The work was carried out by two people who now work at Fiive, before they joined us. It is research work rather than a client delivery.

Tech

Next steps

Want to use ML to structure text data or build better decision support? Contact us and we’ll set up an intro call.