🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
ReadBack: Describe the Patients in Your Own Words
In a hospital, a verbal order is read back before anyone acts on it: the receiver repeats it, the sender confirms. ReadBack applies the same rule to the questions clinicians ask of patient data. Finding a cohort means waiting for an analyst or working through a filter panel with dozens of attributes and hundreds of codes. A language model writes the query in seconds, but it can be wrong in ways that look right, for example, "and" read as "or". For a study, an audit or a recall, a wrong list that looks right is worse than no list. ReadBack targets the error that matters: wrong filters that reach the clinician with nothing on screen saying so. The model only fills the filter panel. It never writes SQL and never sees a patient record. Nine deterministic checks compare the request, the filters and the catalog. Every filter shows its provenance. The explanation is not generated, avoiding hallucinations. Clinicians keep the tool they know: the model fills the panel, they correct it. We took 23 requests, each written in English and in Italian, and had six open models on AMD GPUs draft the filters. Every draft is either exact or wrong. If wrong, marked or unmarked. Accuracy follows the models' capacity, from 25 to 45 exact out of 46, with no unmarked wrong filter reaching the clinician. The benefits of ReadBack are several: Records stay inside: the model sees the request and the catalog, and the same open weights fit a single AMD Instinct MI300X on site. Every draft and every run is logged. Uncertainty is visible. Any value the clinician didn't specify is marked Assumed, every doubtful reading is marked Check on the filter it concerns. No autonomous decisions. It selects patients from criteria a person has accepted. It does not diagnose, rank or recommend. Sensitive data handled with care. No patient record reaches the model; weights can run on the hospital's own AMD hardware.
18 Oct 2026
