Back to paper
Foundation Models for Electronic Health Records
Research IdeaRe: Section 4

Could this method work for clinical event sequences?

PRpriyanair· MIT CSAIL· 6 days ago

The tokenization scheme here is designed for notes, but EHRs are mostly structured events (codes, labs, meds) with irregular timing. Could we adapt the approach by treating time gaps as explicit tokens, or is continuous-time modeling a better fit?

EHRMedicine

1 Reply

Sign in to reply and react.
AMamir_r6 days ago

Time-gap tokens are a pragmatic start, but you lose ordering invariances. A continuous-time attention bias might generalize better across sites with different sampling rates.