MedClear translates doctor-speak into human-speak. Fine-tuned FLAN-T5-base (248M params) that simplifies clinical notes, medical terms, and discharge summaries into plain language patients can understand.

Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("DTanzillo/medclear-v2-base")
model = AutoModelForSeq2SeqLM.from_pretrained("DTanzillo/medclear-v2-base")

text = "simplify: Patient underwent laparoscopic cholecystectomy for acute cholecystitis. EBL minimal. POD1: afebrile, tolerating PO diet."
inputs = tokenizer(text, return_tensors="pt", max_length=512, truncation=True)
outputs = model.generate(**inputs, max_new_tokens=256, num_beams=4)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Data

23,157 examples across multiple granularity levels:

Level	Examples	%
Terms	4,989	21.5%
Phrases	6,660	28.8%
Sentences	8,000	34.5%
Flashcards	2,689	11.6%
Paragraphs	574	2.5%
RAG-augmented	245	1.1%

Key insight: 50% of training is term/phrase level. The model learns vocabulary mappings first, then composes them into simplified text.

Results

Metric	Raw FLAN-T5	MedClear
ROUGE-1 F1	0.13	0.36
ROUGE-2 F1	0.05	0.13
ROUGE-L F1	0.10	0.22
Eval Loss	--	1.712

Limitations

Can hallucinate on complex multi-fact clinical notes
Best used with RAG pipeline (MedlinePlus) for verification
Not a substitute for professional medical advice

Demo

Try the live demo: MedClear on HuggingFace Spaces

Duke University Hackathon 2026