icd crosswalk
r, clinicalbert, sapbert, similarity search
hospital records are full of diagnosis codes that were retired years ago. this pipeline embeds every code and matches it to its modern equivalent, replacing crosswalks that clinical coders used to build by hand.
- f1, icd-9 to icd-10-ca
- 0.42 → 0.52
- f1, icd-9 to icda-8
- 0.72 → 0.76
- trained scoring model, unseen codes
- 0.67and0.84
crossed-out numbers are the published baseline.
see the code
