VHD Audio Classification

This project implements an audio classification system for Valvular Heart Disease (VHD) detection using Deep Learning (RNN-LSTM).

Dataset Preprocessing (preprocessing.ipynb)

The preprocessing script handles the reorganization of the dataset into training and testing sets.

• Source Directory: dataset/train
• Target Directory: dataset/test
• Split Ratio: 10% of the training data is moved to the test set for validation.
• Classes Processed:
  • AS (Aortic Stenosis)
  • MR (Mitral Regurgitation)
  • MS (Mitral Stenosis)
  • MVP (Mitral Valve Prolapse)
  • N (Normal)

Class Mapping

The model uses the following mapping for prediction outputs:

Class	Label	Description
0	AS	Aortic Stenosis
1	MR	Mitral Regurgitation
2	MS	Mitral Stenosis
3	MVP	Mitral Valve Prolapse
4	N	Normal

Model Architecture (vhd_model.ipynb)

The model utilizes Mel-Frequency Cepstral Coefficients (MFCC) features extracted from audio files to train a Recurrent Neural Network (RNN).

Feature Extraction

• Library: Librosa
• Features: MFCC (13 coefficients)
• Input Shape: (100, 13) - padded/truncated to 100 time steps.

Neural Network Structure

• Type: Sequential LSTM
• Layers:
  1. LSTM (128 units)
  2. Dropout (0.3)
  3. Dense (64 units, ReLU)
  4. Dropout (0.3)
  5. Output Dense (5 units, Softmax)

Training Configuration

• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Metics: Accuracy
• Epochs: 50
• Batch Size: 32

Model Conversion

The trained model is saved as model.h5 and converted to TensorFlow.js format for web deployment using tensorflowjs_converter.

Note: Ensure compatible versions of numpy and tensorflowjs are installed to avoid AttributeError: module 'numpy' has no attribute 'bool' during conversion.

Technologies Used

• Core: Python 3.x
• ML Framework: TensorFlow / Keras
• Audio Processing: Librosa
• Web Deployment: TensorFlow.js
• API Backend: FastAPI (Dockerized)

FastAPI Backend (fastapi/)

This folder contains a Dockerized FastAPI application that serves the trained model.h5 for audio classification.

Structure

• main.py: The FastAPI application code.
• Dockerfile: Docker configuration for the API.
• docker-compose.yml: Docker Compose configuration.
• requirements.txt: Python dependencies.
• model.h5: The trained Keras model.

Prerequisites

• Docker and Docker Compose installed.

How to Run

1. Navigate to the fastapi directory:

   cd fastapi

2. Build and run the container:

   docker-compose up -d --build

   The API will be available at http://localhost:8001.

API Usage

Endpoint: /predict [POST]

Upload an audio file to classify it.

Request:

• Key: file
• Type: File
• Value: (Select your audio file, e.g., .wav)

Response Example:

{
    "filename": "audio_sample.wav",
    "prediction": "MR",
    "class_id": 1,
    "confidence": 0.9999,
    "probabilities": {
        "AS": 0.0001,
        "MR": 0.9999,
        "MS": 0.0000,
        "MVP": 0.0000,
        "N": 0.0000
    }
}

Documentation

Visit http://localhost:8001/docs for the interactive Swagger UI.