Medical Dialogue Summary

A local Flask application that helps clinicians turn a post-visit doctor-patient transcript into copy-ready EHR sections. The clinician pastes the full transcript, selects only the sections needed for that encounter, reviews each generated block, and copies it into the EHR.

This is a continuation of a Deep Learning course project for the University of Surrey. The original BART fine-tuning approach is preserved; the default backbone is now GanjinZero/biobart-large, a biomedical BART model released in 2022.

Current Application

Run the doctor-facing application:

python app.py

Open http://127.0.0.1:5002.

The workflow is deliberately section-based because the MTS-Dialog training data pairs each dialogue with a documented EHR section. It does not force irrelevant sections into every note.

• Paste the complete conversation, not an isolated snippet.
• Select the sections relevant to the encounter.
• Generate each section using the model's original dialogue-to-section task.
• Review the output before copying it into the EHR.
• Routine outputs such as None. and Noncontributory. are visibly flagged.

The older dataset-table explorer remains available:

python eval.py

Open http://127.0.0.1:5001.

Model

The committed adapter is biobart-large-lora/. It contains only the LoRA adapter and tokenizer; the 407M-parameter BioBART-large base downloads from Hugging Face on first use.

• Base model: GanjinZero/biobart-large
• Adapter: jonleed/clinical-note-biobart-large-lora
• Fine-tuning method: PEFT LoRA
• LoRA rank: 8
• LoRA alpha: 64
• LoRA dropout: 0.01
• Target modules: BART q_proj and v_proj
• Input context: 512 tokens
• Output budget: 512 tokens in the Flask app
• Decoding: deterministic beam search, 4 beams, repetition penalty 2.0, no-repeat 4-gram constraint

Results

Evaluation uses the same prompt and section-target construction as the original notebook. Scores use ROUGE with stemming on 100 validation dialogues, seed 42.

The evaluation output is checked in at results/eval-biobart-large-2026-08.json.

Controlled Model Study

Using the same validation split, seed, prompt construction, token limits, deterministic decoding, and ROUGE scorer:

The complete study record is results/model-study-2026-08.json. The results show the value of the original LoRA fine-tuning approach and support BioBART-large as the project's selected backbone.

Run the original notebook-style evaluation:

python eval_notebook_protocol.py --split val

Run the Flask integration evaluation through the actual test client:

python eval_app_integration.py --app note --n 100
python eval_app_integration.py --app section --n 100

The BioBART-large adapter produced ROUGE-L 0.4719 through the Flask integration path. This confirms that the served application and evaluation use the same adapter, prompt, input length, and deterministic decode.

Setup

git clone --recurse-submodules https://github.com/jonleed/Medical-Dialogue-Summary.git
cd Medical-Dialogue-Summary
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.py

Configuration

Reproducing Training

train_lora.py mirrors the original notebook while using BioBART-large by default:

python train_lora.py \
  --base_model GanjinZero/biobart-large \
  --output_dir ./biobart-large-lora \
  --per_device_batch 2 \
  --grad_accum 4 \
  --epochs 20

The original notebook recipe is retained in Medical_Dialogue_Generation.ipynb. Training uses the MTS-Dialog training split, the original prompt and target construction, 512-token inputs, 256-token targets, learning rate 2e-4, weight decay 0.01, and per-epoch validation ROUGE. The effective batch size is 8 in the recommended command.

Architecture

graph LR
    A["Full doctor-patient transcript"] --> B["Clinician selects EHR sections"]
    B --> C["Original section prompt"]
    C --> D["BioBART-large"]
    D --> E["LoRA adapter"]
    E --> F["Deterministic beam search"]
    F --> G["Reviewable EHR section"]

Project History

The original project used BART-base with LoRA. The restoration preserves that approach and improves the deployable artifact by:

• Loading the base model and adapter separately through PEFT
• Using the biomedical BioBART-large backbone
• Increasing serving input context from 128 to the 512-token training context
• Providing clinician-selected section generation in demo

This project is educational and is not a medical device.

Credits

• Dataset: Ben Abacha et al., MTS-Dialog, EACL 2023
• Biomedical base: GanjinZero/biobart-large
• Frameworks: PEFT, Transformers, Flask

License

MIT