I am using chunkresolve_snomed_findings_clinical to do the entity linking on Medmentions dataset. Are there any possible ways on Spark NLP for Healthcare to improve the performance? #300
-
|
Hi, I am using chunkresolve_snomed_findings_clinical to do the entity linking on the Medmentions dataset. The implementation has followed the demo https://colab.research.google.com/github/JohnSnowLabs/spark-nlp-workshop/blob/master/tuto[…]n_Trainings/Healthcare/3.Clinical_Entity_Resolvers.ipynb. The UMLS results (using the UMLS concept table to link snomed resolution to UMLS CUIs) are very different from the Medmentions UMLS gold labels. The precision is about 16% and recall is less than 5%. Are there any possible ways to improve the performance? Thanks very much! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
We also run the same experiments a while ago but used Sentence Resolvers, not chunk resolvers.. and found out that we have 52% top-1, 67% top-5 accuracy.. mind that the original MedMentions paper has 45% F1 score.. to test this, you can run the same model we have sbiobertresolve_umls_major_concepts: This model returns CUI (concept unique identifier) codes for Clinical Findings, Medical Devices, Anatomical Structures and Injuries & Poisoning terms. |
Beta Was this translation helpful? Give feedback.
We also run the same experiments a while ago but used Sentence Resolvers, not chunk resolvers.. and found out that we have 52% top-1, 67% top-5 accuracy.. mind that the original MedMentions paper has 45% F1 score.. to test this, you can run the same model we have
sbiobertresolve_umls_major_concepts: This model returns CUI (concept unique identifier) codes for Clinical Findings, Medical Devices, Anatomical Structures and Injuries & Poisoning terms.