Project 2: Patient Note Summarizer¶
Mentors: To be announced
Problem: Dense clinical notes bury the handful of things a patient actually needs to act on (medications, follow-up, restrictions) under clinical shorthand most patients can't read.
Context: Runs on mtsamples for test notes, comparing an open-source summarization model (e.g. facebook/bart-large-cnn via Hugging Face transformers) against LLM prompting with a "summarize for a patient, 5th-grade level" instruction. No training.
Goals: Turn a dense clinical note into a short, patient-facing bullet summary that keeps the actionable parts and drops the clinical shorthand.
Deliverables: A file/text upload flow with a 3-5 bullet summary panel, showing both backends' output side by side (dedicated summarization model vs. general LLM with a summarization prompt) so students can see where a task-specific model differs from prompting a general one.
Showcase: TBD
References:
- Comparing Two Model Designs for Clinical Note Generation; Is an LLM a Useful Evaluator of Consistency?: compares a dedicated summarization model against LLM prompting on clinical notes, the exact side-by-side this project's deliverable produces.
- ClinicalT5: A Generative Language Model for Clinical Text (EMNLP Findings 2022): an open-source, task-specific summarization model fine-tuned on clinical text, the same "dedicated model vs. general LLM" comparison point as
bart-large-cnnin this project.