JMIR preprint: recursive memory raises zero-shot clinical coding F1 from 0.318 to 0.605 after 20 iterations.
This methodological study evaluates a recursive learning architecture for automated clinical coding without task-specific examples. The key result is a stepwise gain from a 0.318 F1 baseline to 0.527 with recursive memory, then to 0.605 after 20 iterations.
The work is currently a preprint under journal review, so the evidence card presents it separately from peer-reviewed publications while still linking users to the source manuscript.
The headline numbers from the paper, in plain language.
How the model performed
Higher is better
Data: Up from 0.318 at baseline, after 20 iterations — JMIR Preprints
Scores reported on a 0–1 scale (AUC, kappa) are drawn ×100.
| Metric | Value | What it means |
|---|---|---|
| F1 · zero-shot clinical coding | 0.605 | Up from 0.318 at baseline, after 20 iterations |
| F1 with recursive memory | 0.527 |
Natalia Castaño-Villegas, Raúl Escandón, Katherine Monsalve, Jose Zea, Laura Velásquez. A Recursive Learning Architecture for Zero-Shot Automated Clinical Coding, a methodological study. JMIR Preprints. 2026. https://preprints.jmir.org/preprint/98279
@article{castanovillegas2026recursive,
title = {{A Recursive Learning Architecture for Zero-Shot Automated Clinical Coding, a methodological study}},
author = {Natalia Castaño-Villegas and Raúl Escandón and Katherine Monsalve and Jose Zea and Laura Velásquez},
journal = {JMIR Preprints},
year = {2026},
url = {https://preprints.jmir.org/preprint/98279},
note = {Preprint}
}
1 min read
1 min read
1 min read
For hospitals and insurers
We run the model on your own clinical records with your team, under your privacy and governance rules.
Start a research partnership