BioSum Reliable

An AI-powered biomedical text summarization system that produces structured, evidence-backed insights from research literature.

BioSum Reliable is an intelligent biomedical NLP system that converts complex research documents into structured summaries. The system combines extractive NLP, transformer-based abstractive summarization, and biomedical entity recognition to produce interpretable and evidence-supported summaries of biomedical literature.

---

Demo

🚀 Live Demo 👉 https://huggingface.co/spaces/saloni-1919/biosum-reliable

OR

Watch the Video!

Web Interface

---

Key Features

• Evidence-based extractive summarization
• Transformer-based abstractive summarization
• Biomedical entity recognition
• Structured research summaries (Objective, Methods, Results, Conclusion)
• Evidence scoring for transparency
• REST API with FastAPI
• Interactive Swagger API documentation
• Docker-ready deployment

---

Use Cases

BioSum Reliable can be used in several biomedical and research workflows:

• Summarizing biomedical research papers
• Extracting key findings from clinical studies
• Supporting literature review for researchers
• Assisting medical professionals in reviewing research
• Integrating biomedical summarization into AI pipelines

---

Example Output

Structured Research Summary

Objective

This study evaluated metformin response in adults with diabetes.

Methods

We reviewed 100 patients and compared baseline glucose and HbA1c values.

Results

HbA1c improved after treatment and fewer patients required insulin rescue therapy.

Conclusion

Metformin improved glycemic control in this cohort.

---

Extractive Summary

Important sentences selected directly from the research document using statistical scoring and biomedical heuristics.

---

Final Summary

A refined summary generated using a transformer model trained on biomedical literature.

---

Key Biomedical Entities

Example extracted entities:

• Disease: Diabetes
• Drug: Metformin
• Drug: Insulin
• Measure: HbA1c

Example System Output

---

System Architecture

BioSum Reliable processes biomedical text through several NLP stages.

1. Input biomedical research text
2. Sentence segmentation and preprocessing
3. Biomedical entity extraction
4. Evidence-based sentence scoring
5. Extractive summarization
6. Transformer-based abstractive summarization
7. Structured summary generation
8. Final summary with supporting evidence

---

Technology Stack

• Python
• FastAPI
• PyTorch
• HuggingFace Transformers
• spaCy NLP
• Uvicorn ASGI Server

This project demonstrates a production-style machine learning pipeline combining NLP, transformer models, and API deployment.

---

Running the Project

Clone the repository

git clone https://github.com/saloni-1919/biosum-reliable.git
cd biosum-reliable

Install Dependencies

Install all required Python packages using the requirements file.

pip install -r requirements.txt

Run the API Server

Start the FastAPI application.

uvicorn app.main:app --reload

Open the API Documentation

Once the server starts, open your browser and navigate to:

http://127.0.0.1:8000/docs

This will open the interactive Swagger API interface where you can test the summarization API.

Example API Request

Endpoint

POST /api/summarize

Example Request Body

{
  "text": "Biomedical research text...",
  "target_sentences": 5,
  "abstractive": true
}

Example Response Includes

• Structured summary
• Extractive summary
• Final summarized output
• Key biomedical entities
• Evidence-ranked sentences

---

Project Structure

biosum-reliable
│
├── app
│   ├── api
│   ├── core
│   ├── ml
│   ├── services
│   └── main.py
│
├── docs
│   ├── demo.gif
│   ├── interface.png
│   ├── output1.png
│   ├── output2.png
│   └── output3.png
│
├── tests
├── LICENSE
├── Dockerfile
├── Procfile
├── pyproject.toml
├── requirements.txt
└── README.md

---

Model Training

Note: Model weights are excluded from the repository due to GitHub file size limits. The model can be reproduced using the training script above.

The abstractive summarization component was trained using a transformer-based model on biomedical research literature.

Dataset

PubMed Scientific Papers Dataset

Base Model

BART Large CNN

Training Configuration

Parameter	Value
Training samples	200
Validation samples	50
Epochs	1
Learning rate	2e-5

Training Results

• Final training loss: 2.57
• Validation loss: 1.97

Reproduce Model Training

python app/ml/train_biomedical_summarizer.py

---

Model Architecture

BioSum Reliable follows a hybrid summarization pipeline combining extractive NLP techniques with transformer-based abstractive summarization.

Architecture Overview

The system processes biomedical research articles through the following stages:

1. Text Input

   • Biomedical research abstract or document is provided by the user.

2. Preprocessing

   • Sentence segmentation
   • Text normalization
   • Section detection (Objective, Methods, Results, Conclusion)

3. Biomedical Entity Recognition

   • Identification of domain-specific entities such as:
     • Diseases
     • Drugs
     • Clinical measurements
     • Treatments

4. Evidence-Based Sentence Scoring

   • Sentences are ranked based on relevance using statistical scoring.

5. Extractive Summarization

   • Top-ranked sentences are selected to produce an evidence-supported summary.

6. Transformer-Based Abstractive Summarization

   • A fine-tuned BART model generates a concise natural-language summary.

7. Structured Research Output

   • Results are organized into structured sections:
     • Objective
     • Methods
     • Results
     • Conclusion

8. Evidence Transparency

   • The system highlights supporting sentences used to generate the summary.

Summarization Pipeline

Biomedical Text
        │
        ▼
Sentence Segmentation
        │
        ▼
Biomedical Entity Recognition
        │
        ▼
Evidence Sentence Scoring
        │
        ▼
Extractive Summary
        │
        ▼
Transformer Abstractive Model (BART)
        │
        ▼
Structured Research Summary

Output Components

The system generates multiple outputs for interpretability:

• Structured summary (Objective, Methods, Results, Conclusion)
• Extractive summary
• Abstractive final summary
• Key biomedical entities
• Evidence-ranked supporting sentences

This multi-stage architecture ensures both interpretability and high-quality summarization for biomedical literature.

---

Future Improvements

• Train on larger biomedical datasets
• Improve entity recognition accuracy
• Add citation extraction
• Support full research paper summarization
• Deploy model using HuggingFace inference API

---

License

This project is licensed under the MIT License.

See the LICENSE file for details.

Author

Saloni Nathani
GitHub: https://github.com/saloni-1919