Patient-Safe Magnetic Resonance Imaging Examination by AI-based Medical Screening

Important results from the project

** Denna text är maskinöversatt ** The purpose of this project was to build an AI model for automatic identification of "implant terms" in medical records. The need for an automatic identification of implant terms spurs from safety reasons because patients who have an implant may be at risk if they undergo Magnetic Resonance Imaging (MRI). To fulfill this purpose, we created and evaluated a BERT model for focused terminology extraction. The model was devised to discover terms indicative of implants in Swedish EMRs.

Expected long term effects

** Denna text är maskinöversatt ** The identification of implant terms in EMRs is very challenging because implant terms and contexts are sparse and because EMRs are an extremely noisy type of document. Regardless these objective difficulties, manual evaluation by domain experts shows that our approach is highly rewarding, since a solid number of indicative terms were discovered by our BERT model. We will use the implant terms discovered by BERT as the gold standard (ground truth) to evaluate future versions of our model.

Approach and implementation

** Denna text är maskinöversatt ** The project was performed using a fruitful collaboration model for the work, including regular frequent zoom-meetings; this was an important feature. Collaboration occurred mainly between Region Östergötland and RISE, but Linköping University was also working intimately with the project. Computing resources and data were localized to Region Linköping due to the sensitive nature of the data, and they were accessed remotely, as agreed on by our legal advisors.