Workflow question: Can you restrict automatic NER pre-annotation to simplify the annotation process?

[LeJudith]

Thanks for the help so far!

Problem

I am trying to use Medcat to annotate main clinically relevant findings in pathology reports using a large pretrained SNOMED International CDB in Medcattrainer. My goal is not to annotate every SNOMED concept in the text. I am mainly interested in the main clinically relevant findings, for example histological type, grade, ER/PR/HER2/Ki-67, margins, lymphovascular invasion (as highlighted in blue in the screenshot). Currently, each document is automatically pre-annotated with many concepts such as ("material", "size", "cells", "protocol", etc.) that are not relevant for my use case. However, I still need to manually mark as "terminate" or "incorrect" before I can submit. This makes the annotation process quite slow. I guess that my use case is also quite different from the intended use of Medcat, but I am wondering if there is a better way.

I have a curated whitelist of ~200 clinically relevant CUIs per organ. The CUI File project filter restricts concept lookup but does not restrict automatic pre-annotation, so irrelevant concepts outside the whitelist are still automatically recognised in grey.

Questions

• Is there a supported way to disable automatic NER pre-annotation entirely, while keeping the full CDB available for manual concept lookup?
• Alternatively, can I restrict automatic pre-annotation to only the CUIs in the project CUI File, while keeping the full CDB available for manual lookup when concepts are missing from the whitelist?
• Or would it be better to build a small CDB containing only the ~200 whitelisted concepts? If so, can I still manually search and annotate concepts outside that CDB using the CDB search filter, or does every concept need to be fully added including CUI, name, and synonyms — which also seems inefficient?
• Or is the best approach to continue with the full CDB and terminate unwanted concepts as negative training examples?

[mart-r]

Are you sure you've set up the CUI filter correclty? Because what you're describing is exactly what this does. It applies a filter to the model and only annotations concepts you've added to the filter.

I.e if I add only the concept 117259009 (Microscopy) to the filter, I see this output (the text is copied from image to formatting isn't great and may have some other differences, but should still show the gist of it):

[mart-r]

Please note that the CUI file is expected in a JSON format, i.e from the admin page:

[LeJudith]

Hi,

Thanks for the quick reply!

I did have the CUI filter in a JSON file already, but I just tried your example and it indeed works. I think the issue was that I had already opened most of the documents once before applying the cui_filter.json. After applying the filter, the automatically recognized concepts in those previously opened documents didn't change, so I assumed the filter wasn't working. I now tested it on documents I hadn't opened yet, and there it behaves exactly as expected.

Sorry for the confusion, and thanks a lot for your help! This will make my life a lot easier 😊

[mart-r]

For documents you've already opened, you probably want to click the "Reset Document" button on the top right: