Medical Classification Project

This Python project classifies patient data into four distinct classes using two different methodologies: Semantic Routing and Modern_Bert. Additionally, it provides tools for visualizing sentence embedding vectors using t-SNE and U-MAP. For interactive plots, users can set the argument -i to enable this feature.

Prerequisites

Install dependencies using:

pip install -r requirements.txt

Files in the Project

• main.py: The main script to classify data or visualize sentence vectors
• Examples.xlsx: Contains a few example patient data for the classification models
• args.py: Handles command-line arguments
• modules/SemanticRoute.py: Implements the Semantic Routing method
• modules/ModernBert.py: Implements the ModernBert method
• utils/t_SNE.py: Script for generating t-SNE visualizations
• utils/UMAP.py: Script for generating UMAP visualizations
• utils/bert_finetuning.py: Script for finetuning ModernBert model

Usage Instructions

1. Classify Data

You can classify patient data into one of four classes using Semantic Routing or Modern_Bert. Run the main.py file with the following arguments:

Semantic Routing

python main.py --semantic

• By default, the dataset is set to Classification_v1.json and the embedding model is set to sentence-transformers/all-MiniLM-L6-v2

• To use a custom dataset or embedding model, add the arguments --dataset <path_to_dataset> or --embedding <embedding_model>

ModernBert

python main.py --bert

• Before using ModernBert, download the model weights from this link and place them in the Modeernbert_model folder.
• Both methods launch an interactive Gradio application where you can input text and receive classification results.

2. Visualize Sentence Vectors

t-SNE Visualization

python main.py --tsne -i

• To visualize a different dataset, add the argument --dataset <path_to_dataset>
• -i or --interactive: Optional flag to enable interactive plotting.

Example t-SNE plot:

UMAP Visualization

python main.py --umap -i

• To visualize a different dataset, add the argument --dataset <path_to_dataset>
• -i or --interactive: Optional flag to enable interactive plotting.

Example UMAP plot:

3. Example Usage

Classify Data Using Semantic Routing

python main.py --semantic

Generate t-SNE Plot (Interactive)

python main.py --tsne --dataset Classification_v1.json -i

Project Structure

.
├── main.py                  # Main entry point
├── args.py                  # Argument parser
├── Examples.xlsx            # Example dataset
├── modules/                 # Contains Semantic Routing and ModernBert implementations
│   ├── SemanticRoute.py
│   └── ModernBert.py
├── utils/                   # Utility scripts for visualizations and finetuning
│   ├── t_SNE.py
│   ├── UMAP.py
│   └── bert_finetuning.py
├── Datasets/                # Contains classification datasets
│   ├── Classification_v1.json
│   └── Classification_bert.json
├── Modeernbert_model/       # Contains ModernBert model and weights (download from the link before use)
└── Visual_plots/            # Contains t-SNE and UMAP plots (tsne_plot.png, umap_plot.png)

Notes

• Make sure to put your custom datasets inside the Datasets folder to work on it.
• Currently it supports only huggingface embedding models. It can further be extended to OpenAI embedding models or any other third-party provider.
• For ModernBert, download the model weights from this link and put them in the Modeernbert_model folder before running the classification.