🫁 Respiratory Disease Diagnosis Using Cough Sound Analysis

A deep learning framework for non-invasive, automated respiratory disease detection through cough audio analysis.

📌 Overview

This project presents an automated two-stage pipeline for diagnosing respiratory diseases from cough sounds — without any invasive procedures or specialized medical equipment.

Stage 1 — Cough Detection: Distinguishes cough sounds from background noise using MFCC features + CNN classifier.
Stage 2 — Disease Classification: Classifies detected coughs into specific respiratory conditions using a multi-class CNN model.
Stage 3 — Wavelet Enhancement: Uses Continuous Wavelet Transform (CWT) scalograms to improve time-frequency representation and disease discrimination.

---

🎯 Diseases Classified

• Asthma
• Bronchitis
• Chronic Obstructive Pulmonary Disease (COPD)
• Pneumonia
• Bronchiolitis
• Bronchiectasis
• Upper Respiratory Tract Infection (URTI)
• Healthy (control)

---

🏗️ System Architecture

Raw Audio Input
      │
      ▼
┌─────────────────────────────┐
│  Preprocessing              │
│  • Noise reduction          │
│  • Amplitude normalization  │
│  • 6-second segmentation    │
└─────────────┬───────────────┘
              │
     ┌────────┴────────┐
     ▼                 ▼
┌─────────┐     ┌──────────────┐
│  MFCC   │     │  Wavelet CWT │
│ Features│     │  Scalograms  │
└────┬────┘     └──────┬───────┘
     │                 │
     ▼                 ▼
┌─────────┐     ┌──────────────┐
│ CNN     │     │ CNN          │
│ Phase 1 │     │ Phase 2      │
│ (Cough  │     │ (Disease     │
│ Detect) │     │ Classify)    │
└────┬────┘     └──────┬───────┘
     │                 │
     └────────┬────────┘
              ▼
       Disease Prediction
       + Class Probabilities

---

📊 Results

Phase 1 — Cough Detection (MFCC + CNN)

Metric	Value
Detection Accuracy	97.4%
Confidence Score (sample)	0.9931
F1-Score	0.9709
Accuracy Score	0.9815

Phase 2 — Disease Classification (MFCC + CNN)

Sampling Rate	Precision	Recall	F1-Score	Accuracy
12 kHz	0.49	0.48	0.46	0.93
16 kHz	0.40	0.53	0.44	0.87
22.05 kHz	0.48	0.52	0.48	0.86
40 kHz	0.60	0.45	0.49	0.93

Overall validation accuracy: 92%

Phase 3 — Wavelet-CNN Classification

Class	Precision	Recall	F1-Score
Bronchiectasis	0.92	0.92	0.92
Bronchiolitis	0.94	0.92	0.93
COPD	0.96	0.90	0.91
Healthy	0.94	0.94	0.95
Pneumonia	0.92	0.94	0.94
URTI	0.92	0.92	0.92
Overall Accuracy			0.90

---

🛠️ Methodology

Feature Extraction — MFCC

1. Pre-emphasis — First-order FIR filter (α = 0.95–0.97)
2. Framing — 20–30 ms overlapping frames
3. Windowing — Hamming window
4. STFT — Short-Time Fourier Transform for spectral analysis
5. Mel Filter Bank — Perceptual frequency scaling
6. DCT — Discrete Cosine Transform for cepstral coefficients (12–20 coefficients)

Feature Extraction — Wavelet Transform

1. CWT — Continuous Wavelet Transform using Morlet wavelet
2. Scalogram — 2D time-frequency energy map
3. Log Compression — Enhances dynamic range
4. CNN Input — Scalogram images fed to CNN for classification

CNN Architecture (MFCC-based)

Input (MFCC: 40 coefficients × Time Frames)
  → Conv2D (32 filters, 3×3) + ReLU + BatchNorm + MaxPooling
  → Conv2D (64 filters, 3×3) + ReLU + BatchNorm + MaxPooling
  → Conv2D (128 filters, 3×3) + ReLU + BatchNorm
  → Global Average Pooling
  → Dense (128) + ReLU + Dropout (0.4)
  → Output (Softmax, n_classes)

---

📁 Dataset

ICBHI 2017 Respiratory Sound Database

• 126 participants (children, adults, elderly)
• 920 labeled recordings (10–90 seconds each)
• ~5.5 hours of audio
• 6,898 respiratory cycles
  • 886 with wheezes
  • 1,864 with crackles
  • 506 with both

---

🔧 Tech Stack

• Language: Python
• Audio Processing: Librosa, NumPy, SciPy
• Feature Extraction: MFCC, CWT (Morlet Wavelet)
• Deep Learning: TensorFlow / Keras (CNN)
• Visualization: Matplotlib
• Data Augmentation: Noise addition, time stretching, pitch shifting

---

🚀 Key Findings

• Cough detection CNN achieves 97.4% accuracy with MFCC features
• 12 kHz sampling rate offers the most balanced performance for disease classification — most clinically relevant cough features lie below 6–8 kHz
• Wavelet (CWT) scalograms outperform MFCCs for non-stationary transient events (crackles, wheezes), improving accuracy from ~83% to 90%
• COPD and Pneumonia achieve the highest per-class accuracy due to distinct spectral signatures
• Minority classes (URTI, LRTI) suffer from limited data — augmentation and GAN-based synthesis recommended for future work

---

🔭 Future Scope

• Expand dataset with more diverse cough samples
• GAN-based data synthesis for rare disease classes
• Transfer learning and attention mechanisms for subtle acoustic variations
• Real-time mobile/web-based screening tool
• Explainable AI (XAI) for clinical trust
• Multi-modal integration (audio + patient metadata)

---

📚 References

1. Al-Anazi et al., "Artificial intelligence in respiratory care," Annals of Thoracic Medicine, 2024.
2. Albawi et al., "Understanding of Convolutional Neural Network (CNN)," ICET, 2017.
3. Ghrabli, Elgendi & Menon, "A spectrogram-based framework for identifying spectral fingerprints in cough sounds," 2024.
4. Author et al., "AI-assisted COVID-19 cough screening using MFCC and log-Mel features," 2023.
5. Sriram, Kumar & Singh, "AI-based cough sound analysis: A survey of ML and DL methods," 2022.
6. Author et al., "Systematic review of AI-based biomedical sound analysis," 2024.
7. Bhattacharya et al., "Wavelet-enhanced CNN for chronic cough detection," 2022.
8. Aykanat et al., "Automatic classification of adventitious respiratory sounds using CNNs," 2017.

---

"By combining signal processing and deep learning, this framework provides a scalable and efficient solution for automated respiratory disease diagnosis — supporting remote health monitoring and early screening."