The problem
One extraction prompt can't serve every physician's writing style
Tempus processes billions of clinical notes from hundreds of institutions. But oncologists write radically differently — some terse and abbreviated, some narrative, some uncertainty-heavy. A single generic LLM extraction prompt treats all notes the same, creating silent failures that pass human review undetected.
Tempus dataset scale
7B+ clinical notes across institutions
Current approach
One generic extraction prompt applied uniformly
The blind spot
Style variation causes silent extraction failures — confident but wrong
Proposed fix
Classify style first → route to adapted prompt
Interactive demo — select a note
Dr. A · Oncology
Terse / abbreviation-heavy
1. Classify style
→
2. Generic prompt
→
3. Style-adapted prompt
→
4. Compare
Generic prompt output
Style-adapted output
—
—
Why the gap exists
Research foundation
Brake & Schaaf (2024) — Personalized Clinical Note Generation from Doctor-Patient Conversations
Showed that modeling physician-specific writing style improves NLP output quality by up to 88.6% on key sections. Focused on note generation — this prototype applies the same insight to the extraction side.
Informatik / MDPI (2025) — Preprocessing of Physician Notes by LLMs Improves Clinical Concept Extraction
Demonstrated that writing style variability, abbreviations, and non-standard formatting degrade downstream NLP pipeline accuracy. Proposed normalizing notes before extraction — this work proposes adapting the prompt instead, preserving clinical nuance.