Notes on Respiratory Disease Classification Project

Introduction

• Overview of the project on respiratory disease classification using lung sounds.
• Project published in the I Transaction Journal (2023).
• Focus on identifying lung-related diseases by analyzing respiratory sounds.

Key Concepts

• Lung Sounds/Respiratory Sounds:
  • Sounds heard by doctors using a stethoscope to diagnose diseases.
  • Stethoscopes may not capture all lung sounds accurately, leading to missed diagnoses.

Project Aim

• Develop a machine learning framework to classify lung sounds into specific disease categories.
• Address shortcomings of traditional stethoscopes by recognizing and analyzing sounds using algorithms.

Data Collection

• Need for a diverse dataset of lung sounds from various patients.
• Example diseases include:
  • Pneumonia
  • Healthy lung sounds
  • Coronary lung sounds
  • Lung infection signals
  • URTI (Upper Respiratory Tract Infection)
• Existing dataset: Biomedical Health Information dataset (2017).
• Aim to enhance the dataset with additional data.

Machine Learning Framework

• Focus on creating a model that includes:
  • Preprocessing
  • Feature extraction
• Proposed system utilizes CNN (Convolutional Neural Networks) and LSTM (Long Short-Term Memory) networks.
• Current papers often use Inception Network, but this project proposes using CNN and LSTM.

Feature Extraction

• Two types of features to be extracted:
  • GTCC (Gaussian Mixture Model-based Time-frequency Coefficients)
  • STFT (Short-Time Fourier Transform)

Addressing Existing System Limitations

• Existing systems often struggle with:
  • High computational time
  • Complex algorithms (e.g., Inception Network)
  • Loss of critical signal features leading to lower accuracy
• Proposed CNN and LSTM algorithms aim to improve model performance and accuracy.

System Architecture

• Workflow includes:
  1. Resampling lung sound signals.
  2. Preprocessing to remove noise and enhance signal quality.
  3. Feature extraction using GTCC and STFT.
  4. Training CNN and LSTM for classification.
  5. Output classification for various diseases (e.g., asthma, COPD, pneumonia).

Advantages of Proposed System

• Lower resource requirements compared to existing systems.
• High accuracy rates (achieved around 98% accuracy).
• Machine learning and deep learning integration.

Applications

• Can be applied in various settings including wearable devices and smartwatches for disease identification.

Conclusion

• The proposed system utilizing CNN and LSTM provides better performance compared to existing systems.
• A significant tool for the health industry in diagnosing respiratory diseases.

Project Execution

• Libraries used: Librosa for audio processing.
• Demonstration of the project web application, allowing users to upload lung sound recordings for analysis.
• Users can see results, including extracted features and disease classification with probabilities.

Final Notes

• Contact details for project acquisition and support offered.
• Encouragement to subscribe for further updates and project information.