Ishika Nahar
All work
IndependentApplied AIHealthcareResearch

Style-Adaptive Clinical Note Extraction

Classify a physician’s writing style first, then route clinical notes to a style-adapted extraction prompt — so terse, narrative, and hedged notes don’t fail silently under one generic LLM prompt.

Organization
Independent exploration
Timeline
2026
My role
Designed and built an interactive HTML prototype for style-aware clinical note extraction.
Interactive demo

The problem

Oncologists write radically differently — some terse and abbreviated, some narrative, some uncertainty-heavy. A single generic LLM extraction prompt treats every note the same, which creates confident but wrong extractions that can slip past review.

Why it matters

At clinical-data scale, style variation is a silent failure mode. Adapting the prompt to the writer’s style (instead of forcing one template) can surface the fields a human abstractor would actually catch.

Data & inputs

  • Synthetic oncology notes illustrating terse, narrative, and uncertainty-heavy styles.
  • Side-by-side generic vs style-adapted field extraction for comparison.

Approach

Classify writing style → run a generic baseline prompt → run a style-adapted prompt → compare extracted fields and accuracy so the gap is inspectable.

  1. 01NoteOncology clinical text
  2. 02Style classifyTerse · narrative · hedged
  3. 03Generic extractOne-size prompt baseline
  4. 04Adapted extractPrompt tuned to that style
  5. 05CompareFields + accuracy gap

My contribution

  • Built the interactive comparison surface for style-adaptive extraction.
  • Encoded three physician styles with explicit generic vs adapted outcomes.
  • Grounded the idea in published work on physician-specific NLP and note preprocessing.

Technical details

  • Client-side HTML/JS proof of concept — not a trained production model.
  • Extraction results are scripted comparisons for walkthrough clarity.

Challenges & decisions

  • Keeping exploratory UI honest about what is simulated vs measured.
  • Showing why style adaptation helps without overclaiming clinical readiness.

Results

  • Shipped a clickable demo: pick a note style, run extraction, see where the generic prompt fails.

Limitations

Exploratory idea demo, not a completed research study or clinical tool. Scores are illustrative for the walkthrough.

What I learned

For clinical NLP, adapting to how a physician writes can matter as much as the model you call.