By Erik Calcina (Junior Researcher at Jozef Stefan Institute)
.
On April 20, 20206, we took part in the OHDSI Europe Symposium in Rotterdam. The event brought together people working with real-world health data, including researchers, clinicians, and developers, all focused on making better use of healthcare data through the OMOP Common Data Model and open tools.
PREPARE had a strong presence this year. We contributed four posters and a software demo, all of which were featured during the main event. One of our works was also selected for a presentation, delivered by Erik Calcina.
Turning clinical notes into structured data
A key part of PREPARE is standardising data across multiple partners so it can be used to develop AI models for personalised rehabilitation. While structured information can be directly mapped into the OMOP CDM, a significant amount of relevant data is still stored in unstructured clinical notes.
These notes contain inconsistent terminology, abbreviations, and a lot of variation depending on the clinician. Before this data can be used, it first needs to be extracted and structured.
To address this, we presented an iterative annotation pipeline that combines AI with human input. Instead of annotating everything from scratch, a model first generates suggestions, which are then reviewed and corrected. These corrections are used to improve the model, and the process repeats.
With each iteration, the model becomes more useful. Annotation time dropped from almost three minutes per note to around 45 seconds. At the same time, smaller models performed comparably to larger ones, while being much easier to run in real clinical settings.
Why this step is important for PREPARE
This approach makes it more realistic to build high-quality clinical datasets without a large manual burden. That is an important step if we want to use clinical data for prediction models and personalised rehabilitation.
Related to this work, we also developed the PREPARE Extraction Tool (https://github.com/E3-JSI/PREPARE-Extraction-Tool) which was presented as a software demonstration by Tinkara Meterc. This is a semi-automated, end-to-end pipeline designed to improve efficiency and interoperability while reducing manual effort for clinical staff.
The pipeline consists of three main components:
- Semi-automated extraction of clinically relevant information from medical records
- Semantic clustering of extracted entities to consolidate similar concepts
- Mapping of these concepts to standardised OHDSI vocabularies available through Athena
Final thoughts
The symposium was a good opportunity to connect with others working on similar challenges and to exchange ideas around data standardisation and analysis. For PREPARE, it was also a chance to share our progress and show how different parts of the project contribute to making clinical data more usable in practice.
.
Posted by: Nayara Güércio (Trilateral Research)