🫁 Respiratory Sound Classification: 8-Class Pulmonary Disease Detection

A deep learning pipeline for multiclass classification of pulmonary diseases from respiratory sounds. Published in IEEE WCONF 2023.

---

📌 Project Overview

This research project implements an 8-class respiratory disease classification pipeline using real-world auscultation data. It applies audio preprocessing, mel-spectrogram conversion, and CNN-based classification using ResNet50, EfficientNet-B0, and VGG16.

📝 Published Paper: "A Deep Learning Framework for Multiclass Categorization of Pulmonary Diseases"

🔗 IEEE Xplore Link

📚 DOI: 10.1109/WCONF58270.2023.10235057

---

🩺 Dataset

The project is based on the ICBHI 2017 Respiratory Sound Database, a real-world, expert-annotated dataset collected from 126 patients across various age groups.

• 📈 6,898 labeled audio segments
• 🧓 Age groups: Children, Adults, Elderly
• ⏱ Recording length: 10s–90s
• 🧬 Classes: Pneumonia, COPD, Asthma, LRTI, URTI, Bronchiectasis, Bronchiolitis, Healthy

🔗 Dataset Access

Dataset Version	Link
📦 Official ICBHI Site	bhichallenge.med.auth.gr
📁 Kaggle Mirror	Kaggle Link
🎧 Preprocessed Audio	Google Drive
🖼️ Mel-Spectrograms	Google Drive

---

🧠 Pipeline Architecture

1. Preprocessing – Segment audio into 6s slices, denoise, normalize
2. Feature Extraction – Convert to mel-spectrograms
3. Model Training – CNN-based classification (EfficientNet, ResNet50, VGG16)
4. Evaluation – Confusion matrix, accuracy, ROC curves

---

🧪 Model Performance

Model	Task	Accuracy (%)
ResNet50	Binary (Healthy vs Others)	87.1
EfficientNet-B0	Binary (Healthy vs Others)	94.1
EfficientNet-B0	Multiclass (8-Class)	82.1
ResNet50	Multiclass (8-Class)	79.8
VGG16	Multiclass (8-Class)	0.14 (failed)

---

📂 Repository Contents

.
├── 01_audio_preprocessing_subslices.ipynb
├── 02_train_test_split.ipynb
├── 03_feature_extraction_spectrograms.ipynb
├── 04_resnet50_binary_classification.ipynb
├── 05_efficientnet_binary_classification.ipynb
├── 06_efficientnet_multiclass_classification.ipynb
├── 07_resnet50_multiclass_classification.ipynb
├── 08_vgg16_multiclass_classification.ipynb
├── patient-disease-labels.csv
├── pulmonary.png
└── README.md

---

🧾 Notebook Descriptions

Notebook	Description
01_audio_preprocessing_subslices.ipynb	Splits raw audio into 6-second segments
02_train_test_split.ipynb	80/20 stratified data split
03_feature_extraction_spectrograms.ipynb	Mel-spectrogram generation using librosa
04_resnet50_binary_classification.ipynb	Binary classification using ResNet50
05_efficientnet_binary_classification.ipynb	Binary classification using EfficientNet
06_efficientnet_multiclass_classification.ipynb	8-class classification using EfficientNet
07_resnet50_multiclass_classification.ipynb	8-class classification using ResNet50
08_vgg16_multiclass_classification.ipynb	Failed training example (for comparison)

---

⚙️ Setup Instructions

🛠️ Requirements

• Python 3.8.3+
• TensorFlow 2.x
• Librosa
• NumPy, Pandas, scikit-learn
• Jupyter, Matplotlib

🚀 Quick Start

git clone https://github.com/your-username/respiratory-sound-classification.git
cd respiratory-sound-classification
pip install -r requirements.txt

Then run notebooks in order from 01_ to 08_.

---

📖 Use Case & Clinical Scope

⚠️ This project is intended for research and educational purposes only. It is not approved for clinical use by any regulatory authority (e.g. FDA, CE, CDSCO).

Potential Applications:

• Decision support tools
• AI-assisted stethoscope design
• Remote respiratory screening systems

---

📚 Citation

Please cite this work as:

@INPROCEEDINGS{10235057,
  author={Khanaghavalle, G R and Manoj, Allen and Karthikeyan, Janani and Murali, Ritunjay},
  booktitle={2023 World Conference on Communication & Computing (WCONF)}, 
  title={A Deep Learning Framework for Multiclass Categorization of Pulmonary Diseases}, 
  year={2023},
  pages={1-6},
  keywords={Deep learning;Pulmonary diseases;Lung;Air pollution;Convolutional neural networks},
  doi={10.1109/WCONF58270.2023.10235057}
}

---

🧠 Skills Demonstrated

• Audio signal processing (librosa)
• CNN-based image classification
• Model evaluation & statistical validation
• Reproducible research practices
• Clinical AI design awareness
• Research publication & reporting