MedJev: Fast Clinical Variable Extraction — With an Important Asterisk on the Numbers

Curating clinical variables from free-text notes is tedious. General LLMs can help, but processing thousands of notes can be slow and costly. Inspired by Jev and the open-source community, we’re releasing MedJev to turn clinical notes into structured fields on consumer GPUs.
A 0.8B model + 43 MB LoRA adapter, trained to extract 11 predefined clinical variables.
On our benchmark of 2,895 held-out notes:
⚡ 62 ms per note, 4.6× faster than the Jev API
🎯 87.8% accuracy vs. 60.2% for Jev
🏠 Runs locally, patient text stays on your network
Data & Code: github.com/JunMa11/MedJev
Model: https://lnkd.in/dyebBewd

MedJev, built by Jun Ma on the open-source "kev" decision-model architecture, is a 0.8B model plus a LoRA adapter that extracts 11 predefined clinical variables from free-text notes in a single forward pass — about 62ms per note on one GPU, running locally so patient text never leaves your network.

On the project's own held-out benchmark, MedJev reports 87.8% agreement versus 60.2% for querying the general-purpose Jev API directly, at roughly 4.6x the speed. Important context: these accuracy figures measure agreement with "silver" labels generated by GPT-4 against a medical template, not clinician-verified ground truth — so the numbers reflect how well MedJev reproduces a GPT-4 extraction, not validated clinical accuracy. The project's own documentation states it is a research artifact, not for clinical use, and has not been validated against clinician judgment.