AURA-ED
An LLM that reads raw emergency-department data and writes the risk summary a clinician usually assembles in their head while running between rooms.
What eats an ED physician's day
Emergency physicians spend roughly a fifth of their week on indirect work: documentation, chart review, the reading and re-reading that never touches a patient directly. Most of the AI aimed at this problem transcribes. It writes down what was said. It does not tell you what any of it means. For a team project in UVA's Healthcare for Data Science course, we wanted to see how far past transcription a language model could actually go.
22% of an ED physician's week goes to documentation and chart review.Most tools transcribe. We wanted one that could summarize.
The synthesis, done by a model
We built an ingestion pipeline that pulls triage vitals, lab results, comorbidity scores, and demographics from 118,385 Stanford ED encounters in the MC-MED dataset. On top of it sits a language model that generates structured risk narratives for the conditions where minutes matter: sepsis, acute kidney injury, stroke, pulmonary embolism, acute coronary syndrome, heart failure.
Between the raw data and the model, a clinical flagging layer maps continuous vitals to categorical risk levels and turns cryptic ICD codes into readable comorbidity categories. A Tableau dashboard sits alongside for exploratory risk analysis. We scored the output on three axes, not one: predictive accuracy, hallucination rate, and equity across demographic groups.
Why preprocessing was the real project
The model was the visible part. The work was everything upstream of it. Before an LLM could say anything clinically accurate, the data needed outlier filtering, threshold checks against physiological norms, and a translation layer that turned column headers no human would recognize into actual medical terminology. Get that wrong and the model produces fluent, confident nonsense. Getting it right is most of what made the summaries trustworthy.