Deep Learning-Based Early COPD Detection and Severity Classification
Using Respiratory Sound Analysis
VTU Major Project · Department of Computer Science · VCET Puttur · 2025–26
- Overview
- Motivation & Problem Statement
- Team
- Objectives
- Dataset — ICBHI 2017
- Methodology & Architecture
- Repository Structure
- Tech Stack
- Getting Started
- Notebooks Guide
- Expected Results
- Project Timeline
- Applications
- Contributing
- License
- References
- Acknowledgements
Chronic Obstructive Pulmonary Disease (COPD) is a progressive, life-threatening lung disease affecting over 300 million people worldwide and is the third leading cause of death globally (WHO, 2023). Despite its prevalence, COPD is frequently diagnosed only in advanced stages, largely because early symptoms — chronic cough, mild breathlessness — are subtle and often dismissed.
This project develops an AI-powered, non-invasive COPD screening system that analyzes respiratory (lung) sounds recorded via digital stethoscopes to:
- Detect the presence of COPD (binary classification: COPD / Normal)
- Classify its severity into four clinical grades: Normal → Mild → Moderate → Severe
The system leverages the ICBHI 2017 Respiratory Sound Database and a hybrid CNN-LSTM deep learning architecture, which jointly captures spectral (spatial) and temporal patterns from audio signals — making it more robust than traditional CNN or LSTM-only approaches.
- COPD is irreversible but manageable — early intervention significantly slows progression and improves quality of life.
- Conventional diagnosis (spirometry, CT scans) is expensive, invasive, requires specialist access, and is largely unavailable in rural or resource-limited settings.
- Acoustic-based screening using lung sounds (wheezing, crackles, rhonchi) is cost-effective, portable, and repeatable.
- Lung sounds contain rich acoustic biomarkers that correlate with structural airway changes in COPD.
- Traditional signal processing approaches (hand-crafted features + classical ML) lack generalization.
- CNNs capture local spectral patterns in spectrograms; LSTMs capture long-range temporal dependencies in breathing cycles — making a CNN-LSTM hybrid ideal for this domain.
- Most existing work focuses only on binary classification (COPD / Non-COPD) or single-disease detection (asthma, pneumonia). This project extends to 4-class severity grading aligned with GOLD (Global Initiative for Chronic Obstructive Lung Disease) clinical stages.
| USN | Name | Role |
|---|---|---|
| 4VP23CS084 | Sanath K | Team Leader — Project coordination, model development |
| 4VP23CS070 | Rajath Kiran A | Team Member — Data preprocessing, feature engineering |
| 4VP23CS076 | Rithesh | Team Member — Model training, hyperparameter tuning |
| 4VP23CS093 | Sheethal D Rai | Team Member — Evaluation, reporting, documentation |
Project Guide: Prof. Pramod Kumar PM Department of Computer Science & Engineering, VCET Puttur (VTU Affiliated)
| # | Objective | Phase |
|---|---|---|
| 1 | Acquire and understand the ICBHI 2017 respiratory sound dataset | Phase I |
| 2 | Implement audio preprocessing: noise reduction, resampling to 22050 Hz, respiratory cycle segmentation | Phase I |
| 3 | Extract audio features: MFCC (40 coefficients), Mel-spectrograms (128 bands), Chroma, Spectral Contrast | Phase I |
| 4 | Build and validate a Baseline CNN model for binary COPD classification (COPD / Normal) | Phase I |
| 5 | Extend to a CNN-LSTM Hybrid for 4-class severity classification: Normal / Mild / Moderate / Severe | Phase II |
| 6 | Achieve binary classification accuracy > 90% and severity F1-score > 80% | Phase II |
| 7 | Evaluate models using Accuracy, Precision, Recall, F1-Score, ROC-AUC, and Confusion Matrix | Phase II |
| 8 | Document findings in a research report and target a journal submission (IEEE / Springer) | Phase III |
ICBHI 2017 Respiratory Sound Database
Rocha et al., "An Open Access Database for the Evaluation of Respiratory Sound Classification Algorithms", Physiological Measurement, 2019.
| Property | Details |
|---|---|
| Source | ICBHI 2017 Challenge |
| Total Recordings | 920 audio files (.wav format) |
| Total Duration | ~5.5 hours |
| Subjects | 126 patients |
| Recording Devices | 7 different digital stethoscopes |
| Recording Locations | Trachea, anterior/posterior chest, lateral |
| Breath Cycle Annotations | 6,898 annotated respiratory cycles |
| Sound Labels | Normal, Crackle, Wheeze, Crackle+Wheeze |
| Patient Demographics | Age, sex, weight, height, BMI, diagnosis included |
The dataset covers patients with: COPD, Healthy, URTI (Upper Respiratory Tract Infection), Bronchiectasis, Pneumonia, Bronchiolitis, and LRTI (Lower Respiratory Tract Infection).
For this project, patient diagnosis metadata is used to derive COPD severity labels aligned with GOLD Staging:
- GOLD 0 / Normal — No obstruction
- GOLD I / Mild — FEV1 ≥ 80% predicted
- GOLD II / Moderate — 50% ≤ FEV1 < 80%
- GOLD III–IV / Severe — FEV1 < 50%
Note: The raw dataset audio files are not tracked in this repository due to size. See Dataset Setup for download instructions.
┌─────────────────────────────────────────────────────────────┐
│ COPD Detection Pipeline │
└─────────────────────────────────────────────────────────────┘
[RAW AUDIO INPUT] (.wav files, 920 recordings)
│
▼
┌───────────────────────────────────┐
│ 1. DATA PREPROCESSING │
│ • Noise reduction (noisereduce) │
│ • Resampling → 22050 Hz │
│ • Channel normalization (mono) │
│ • Respiratory cycle segmentation │
│ using ICBHI annotation files │
└───────────────────────────────────┘
│
▼
┌───────────────────────────────────┐
│ 2. FEATURE EXTRACTION │
│ • Mel-Spectrogram (128 bands) │
│ • MFCC (40 coefficients) │
│ • Delta & Delta-Delta MFCCs │
│ • Chroma Features │
│ • Spectral Contrast │
│ • Zero Crossing Rate │
└───────────────────────────────────┘
│
▼
┌───────────────────────────────────┐
│ 3. LABEL ASSIGNMENT │
│ • Binary: COPD / Normal │
│ • 4-Class: Normal / Mild / │
│ Moderate / Severe (GOLD) │
└───────────────────────────────────┘
│
┌───┴──────────────┐
│ │
▼ ▼
┌──────────┐ ┌────────────────────┐
│ MODEL 1 │ │ MODEL 2 │
│ Baseline │ │ CNN-LSTM Hybrid │
│ CNN │ │ (Sequential + │
│ (Binary) │ │ Temporal) │
└──────────┘ └────────────────────┘
│ │
▼ ▼
┌──────────┐ ┌────────────────────┐
│ COPD / │ │ Normal / Mild / │
│ Normal │ │ Moderate / Severe │
└──────────┘ └────────────────────┘
│
▼
┌───────────────────────────────────┐
│ 4. EVALUATION │
│ • Accuracy, Precision, Recall │
│ • F1-Score (macro + weighted) │
│ • ROC-AUC (OvR) │
│ • Confusion Matrix │
│ • Learning Curves │
└───────────────────────────────────┘
File: src/preprocess.py
The preprocessing pipeline transforms raw .wav recordings into clean, segmented audio clips ready for feature extraction.
Steps:
- Load audio using
librosa.load()at a target sample rate of 22,050 Hz (mono) - Noise reduction using
noisereducelibrary (spectral gating method) - Normalization — peak amplitude normalized to [-1, 1]
- Segmentation — ICBHI annotation
.txtfiles provide exact start/end times for each respiratory cycle. Each cycle is extracted as an independent sample. - Padding/Truncation — cycles are zero-padded or truncated to a fixed length (e.g., 5 seconds) for uniform input size.
Output: Segmented .npy audio arrays stored in data/processed/
File: src/features.py
Features are extracted per respiratory cycle and stored as NumPy arrays.
| Feature | Description | Shape (per sample) |
|---|---|---|
| Mel-Spectrogram | Frequency-time representation using mel scale | (128, T) |
| MFCC | 40 Mel-Frequency Cepstral Coefficients + Δ + ΔΔ | (120, T) |
| Chroma | 12 pitch class energy distributions | (12, T) |
| Spectral Contrast | Energy difference across 7 frequency bands | (7, T) |
| Zero Crossing Rate | Rate of signal sign changes | (1, T) |
For the CNN models, Mel-Spectrograms (and optionally stacked MFCC features) are used as 2D image-like inputs. For the CNN-LSTM model, the temporal axis T forms the sequence dimension fed into the LSTM layers.
Output: Feature arrays stored in data/features/ as .npy files with matching label arrays.
File: src/models/baseline_cnn.py
A standard 2D Convolutional Neural Network treating Mel-Spectrograms as images.
Input: Mel-Spectrogram (1, 128, T)
│
├── Conv2D(32, 3×3) → BatchNorm → ReLU → MaxPool(2×2)
├── Conv2D(64, 3×3) → BatchNorm → ReLU → MaxPool(2×2)
├── Conv2D(128, 3×3) → BatchNorm → ReLU → MaxPool(2×2)
├── Conv2D(256, 3×3) → BatchNorm → ReLU → GlobalAvgPool
│
├── Flatten
├── Dropout(0.4)
├── Dense(256) → ReLU
├── Dropout(0.3)
└── Dense(2) → Softmax (Binary: COPD / Normal)
Training Config:
- Optimizer: Adam (lr=1e-4)
- Loss: Cross-Entropy
- Batch Size: 32
- Epochs: 50 with EarlyStopping (patience=10)
- Data Augmentation: Time stretching, pitch shifting, additive noise
File: src/models/cnn_lstm.py
The hybrid model first applies convolutional layers to capture local spectral patterns, then feeds the resulting feature maps as a time sequence into LSTM layers to model temporal breathing dynamics.
Input: Mel-Spectrogram (1, 128, T)
│
├── [CNN Block — Feature Extraction]
│ Conv2D(64, 3×3) → BN → ReLU → MaxPool (freq axis only)
│ Conv2D(128, 3×3) → BN → ReLU → MaxPool (freq axis only)
│ Conv2D(256, 3×3) → BN → ReLU
│ → Reshape: (Batch, T', Features) ← time steps preserved
│
├── [LSTM Block — Temporal Modeling]
│ BiLSTM(256 units) → Dropout(0.3)
│ LSTM(128 units) → Dropout(0.3)
│
├── Attention Layer (optional, Phase II)
│
├── Dense(256) → ReLU → Dropout(0.4)
├── Dense(128) → ReLU
└── Dense(4) → Softmax (4-class: Normal/Mild/Moderate/Severe)
Key Design Decisions:
- Pooling is applied only along the frequency axis in CNN blocks so that the time dimension
Tis preserved and passed intact to the LSTM. - Bidirectional LSTM is used so that each time step has context from both past and future respiratory cycle segments.
- An optional attention mechanism highlights the most diagnostically relevant time windows.
File: src/evaluate.py
| Metric | Why It Matters |
|---|---|
| Accuracy | Overall correctness; baseline measure |
| Precision | Minimize false COPD positives (avoid unnecessary anxiety) |
| Recall (Sensitivity) | Minimize missed COPD cases (critical for screening) |
| F1-Score | Harmonic mean; handles class imbalance |
| ROC-AUC | Discrimination ability across thresholds |
| Confusion Matrix | Per-class error analysis |
| Learning Curves | Overfitting / underfitting diagnosis |
For multi-class severity, both macro-averaged and weighted F1-scores are reported. Given the clinical context (where missing a true COPD case is more costly than a false positive), Recall is treated as the primary optimization metric.
COPD-Detection/
│
├── 📁 data/
│ ├── raw/ # Original ICBHI 2017 .wav files + .txt annotations
│ │ └── (download separately — see Dataset Setup)
│ ├── processed/ # Segmented respiratory cycle arrays (.npy)
│ └── features/ # Extracted feature arrays
│ ├── mel_spectrograms.npy
│ ├── mfcc_features.npy
│ └── labels_{binary,4class}.npy
│
├── 📁 notebooks/
│ ├── 01_eda.ipynb # Exploratory Data Analysis
│ │ # - Dataset distribution, demographics
│ │ # - Audio waveform & spectrogram visualization
│ │ # - Class imbalance analysis
│ ├── 02_preprocessing.ipynb # Preprocessing walkthrough with visualizations
│ ├── 03_baseline_cnn.ipynb # Baseline CNN training, evaluation, ablation
│ └── 04_cnn_lstm_hybrid.ipynb # CNN-LSTM training, severity classification
│
├── 📁 src/
│ ├── preprocess.py # Full preprocessing pipeline (CLI + importable)
│ ├── features.py # Feature extraction module
│ ├── dataset.py # PyTorch Dataset class for ICBHI data
│ ├── augment.py # Audio augmentation utilities
│ ├── train.py # Training loop with logging & checkpointing
│ ├── evaluate.py # Evaluation script (metrics, plots)
│ ├── utils.py # Helper functions (seed setting, plotting, etc.)
│ └── models/
│ ├── __init__.py
│ ├── baseline_cnn.py # Baseline CNN architecture (PyTorch nn.Module)
│ └── cnn_lstm.py # CNN-LSTM Hybrid architecture
│
├── 📁 results/
│ ├── plots/
│ │ ├── confusion_matrix_cnn.png
│ │ ├── confusion_matrix_cnn_lstm.png
│ │ ├── roc_curves.png
│ │ └── loss_accuracy_curves.png
│ └── metrics.json # Serialized evaluation results (all models)
│
├── 📁 reports/
│ ├── synopsis.pdf # Approved project synopsis
│ └── progress_reports/
│ ├── PR1_introduction.pdf
│ ├── PR2_requirements.pdf
│ ├── PR3_design.pdf # (upcoming)
│ └── PR4_implementation.pdf # (upcoming)
│
├── 📁 checkpoints/ # Saved model weights (.pth)
│ ├── baseline_cnn_best.pth
│ └── cnn_lstm_best.pth
│
├── requirements.txt # Python dependencies
├── .gitignore
├── LICENSE
└── README.md
Note:
data/raw/,data/processed/,data/features/, andcheckpoints/are listed in.gitignoreand must be generated locally following the Getting Started instructions.
| Category | Library / Tool | Version | Purpose |
|---|---|---|---|
| Language | Python | 3.10+ | Core development language |
| Deep Learning | PyTorch | ≥ 2.0 | Model building, training, GPU acceleration |
| Audio Processing | Librosa | ≥ 0.10 | Feature extraction, audio I/O |
| Audio I/O | SoundFile | ≥ 0.12 | Reading/writing WAV files |
| Noise Reduction | noisereduce | ≥ 3.0 | Spectral gating-based denoising |
| Numerical | NumPy | ≥ 1.24 | Array operations |
| Data Manipulation | Pandas | ≥ 2.0 | Metadata handling, CSVs |
| Visualization | Matplotlib | ≥ 3.7 | Plots, spectrograms, curves |
| Visualization | Seaborn | ≥ 0.12 | Statistical visualizations |
| ML Utilities | Scikit-learn | ≥ 1.3 | Metrics, splits, class balancing |
| Notebooks | Jupyter | ≥ 7.0 | Interactive development & EDA |
| Experiment Tracking | TensorBoard | ≥ 2.13 | Loss curves, metric logging |
| Dataset | ICBHI 2017 | — | Respiratory sound recordings |
- Python 3.10 or higher
pippackage manager- Git
- A CUDA-capable GPU is recommended for training (CPU training is possible but slow)
- ~3 GB free disk space for the dataset + features
Verify your Python version:
python --version
# Should output: Python 3.10.x or higher# 1. Clone the repository
git clone https://github.com/Rajath2005/COPD-Detection.git
cd COPD-Detection
# 2. Create a virtual environment (recommended)
python -m venv venv
# Activate on Linux/macOS:
source venv/bin/activate
# Activate on Windows (Command Prompt):
venv\Scripts\activate.bat
# Activate on Windows (PowerShell):
venv\Scripts\Activate.ps1
# 3. Install all dependencies
pip install --upgrade pip
pip install -r requirements.txtThe ICBHI 2017 dataset must be downloaded separately from the official source.
# Step 1: Visit the official ICBHI 2017 Challenge page
# https://bhichallenge.med.auth.gr/ICBHI_2017_Challenge
# Register if required and download the dataset archive.
# Step 2: Extract the archive
unzip ICBHI_2017_Challenge.zip -d data/raw/
# OR
tar -xvf ICBHI_2017_Challenge.tar.gz -C data/raw/
# Expected structure after extraction:
# data/raw/
# ├── 101_1b1_Al_sc_Meditron.wav
# ├── 101_1b1_Al_sc_Meditron.txt ← annotation file
# ├── 102_1b1_Al_sc_Meditron.wav
# ├── 102_1b1_Al_sc_Meditron.txt
# ├── ...
# └── ICBHI_Challenge_diagnosis.txt ← patient diagnosesTip: The annotation
.txtfiles contain tab-separated rows:start_time end_time crackle_label wheeze_label. These are parsed bysrc/preprocess.pyto segment recordings into individual respiratory cycles.
Run each step sequentially:
# Step 1: Preprocess raw audio → segmented cycles
python src/preprocess.py \
--input_dir data/raw/ \
--output_dir data/processed/ \
--target_sr 22050 \
--cycle_duration 5.0
# Step 2: Extract features from processed cycles
python src/features.py \
--input_dir data/processed/ \
--output_dir data/features/ \
--feature_type mel_mfcc
# Step 3a: Train the Baseline CNN (binary classification)
python src/train.py \
--model baseline_cnn \
--feature_dir data/features/ \
--task binary \
--epochs 50 \
--batch_size 32 \
--lr 1e-4 \
--checkpoint_dir checkpoints/
# Step 3b: Train the CNN-LSTM Hybrid (severity classification)
python src/train.py \
--model cnn_lstm \
--feature_dir data/features/ \
--task severity \
--epochs 100 \
--batch_size 32 \
--lr 5e-5 \
--checkpoint_dir checkpoints/
# Step 4: Evaluate a trained model
python src/evaluate.py \
--model cnn_lstm \
--checkpoint checkpoints/cnn_lstm_best.pth \
--feature_dir data/features/ \
--task severity \
--output_dir results/
# Step 5: Launch TensorBoard to monitor training
tensorboard --logdir runs/# Start the Jupyter server
jupyter notebook notebooks/
# Or launch JupyterLab for a better UI
pip install jupyterlab
jupyter lab notebooks/Open notebooks in this order for a complete walkthrough:
01_eda.ipynb→ Understand the dataset02_preprocessing.ipynb→ See preprocessing steps visually03_baseline_cnn.ipynb→ Train and evaluate the baseline04_cnn_lstm_hybrid.ipynb→ Full hybrid model
- Dataset overview: number of recordings, patients, total duration
- Audio waveform visualization of Normal vs COPD samples
- Mel-Spectrogram comparison across disease types
- Class distribution and imbalance analysis
- Patient demographics (age, sex, BMI) by diagnosis
- Respiratory cycle duration statistics
- Step-by-step noise reduction demonstration (before/after spectrograms)
- Resampling effect on audio quality
- Respiratory cycle segmentation walkthrough using annotations
- Fixed-length padding/truncation strategy comparison
- Feature preparation and data loaders
- Model architecture summary and parameter count
- Training loop with live loss/accuracy curves
- Validation performance: accuracy, precision, recall, F1
- Confusion matrix visualization
- Error analysis: common misclassification patterns
- Ablation study: MFCC vs Mel-Spectrogram as input
- Architecture design rationale
- Training on 4-class severity labels
- Multi-class metrics: macro F1, per-class precision/recall
- ROC curves (One-vs-Rest for each severity class)
- Attention weight visualization (if attention is enabled)
- Comparison against Baseline CNN
| Metric | Target | Expected Range |
|---|---|---|
| Accuracy | > 90% | 90–94% |
| Precision (COPD) | > 88% | 87–92% |
| Recall (COPD) | > 88% | 88–93% |
| F1-Score (COPD) | > 88% | 88–92% |
| ROC-AUC | > 0.92 | 0.92–0.96 |
| Metric | Target | Expected Range |
|---|---|---|
| Accuracy | > 80% | 80–87% |
| Macro F1-Score | > 78% | 78–85% |
| Weighted F1-Score | > 82% | 82–88% |
| ROC-AUC (OvR, macro) | > 0.88 | 0.88–0.93 |
Note: These are projected targets based on prior literature on ICBHI 2017 benchmarks. Actual results will be reported and updated here during Phase II.
| Paper | Model | Task | Accuracy |
|---|---|---|---|
| Perna & Tagarelli (2019) | CNN | Binary | 83.7% |
| Nguyen et al. (2020) | CNN-RNN | Binary | 86.5% |
| Ma et al. (2021) | CNN-LSTM + Attention | Binary | 89.2% |
| This Work (Target) | CNN-LSTM Hybrid | Binary + 4-Class | > 90% |
| # | Milestone | Deadline | Status |
|---|---|---|---|
| 1 | Synopsis submission | 27 Feb 2026 | ✅ Done |
| 2 | Synopsis presentation | 03–10 Mar 2026 | ✅ Done |
| 3 | Progress Report 1 — Introduction & Background | 14 Apr 2026 | 🔄 In Progress |
| 4 | Progress Report 2 — Requirements & Dataset | 14 Apr 2026 | 🔄 In Progress |
| 5 | Presentation 1 (EDA + Preprocessing + Baseline CNN) | 2nd Week May 2026 | ⏳ Upcoming |
| 6 | Progress Report 3 — System Design & Architecture | 4th Week Aug 2026 | ⏳ Upcoming |
| 7 | Progress Report 4 — Implementation Details | 4th Week Aug 2026 | ⏳ Upcoming |
| 8 | Presentation 2 (CNN-LSTM Results) | 2nd Week Sep 2026 | ⏳ Upcoming |
| 9 | Progress Report 5 — Testing & Validation | 2nd Week Oct 2026 | ⏳ Upcoming |
| 10 | Progress Report 6 — Results & Analysis | 2nd Week Oct 2026 | ⏳ Upcoming |
| 11 | Final Report + Journal Paper Draft | 20 Nov 2026 | ⏳ Upcoming |
| 12 | Internal Viva / Project Exhibition | 4th Week Nov 2026 | ⏳ Upcoming |
- Early COPD Screening Tool — deployable as a mobile/web app integrated with a digital stethoscope for point-of-care screening
- Rural & Remote Healthcare — works on low-cost devices, no specialist required, no imaging equipment needed
- Telemedicine Integration — patients record lung sounds at home; AI flags abnormalities for remote physician review
- AI-Assisted Diagnosis — reduces clinician workload in high-volume respiratory clinics
- Longitudinal Monitoring — track severity progression over time using repeated recordings
- Drug Trial Endpoints — objective acoustic biomarkers for COPD severity in clinical trials
- Medical Training — visualize and study respiratory acoustic patterns for student training
- Open-Source Benchmark — reproducible ICBHI 2017 pipeline for the research community
We welcome contributions from the community! Whether you're fixing a bug, improving documentation, or proposing a new model variant — all contributions are appreciated.
# 1. Fork the repository on GitHub
# 2. Clone your fork
git clone https://github.com/<your-username>/COPD-Detection.git
cd COPD-Detection
# 3. Create a feature branch
git checkout -b feature/your-feature-name
# 4. Make your changes and commit
git add .
git commit -m "feat: add attention mechanism to CNN-LSTM"
# 5. Push to your fork
git push origin feature/your-feature-name
# 6. Open a Pull Request on GitHub- Follow PEP 8 style for Python code
- Add docstrings to all functions and classes
- Write unit tests for new features in
tests/ - Update relevant documentation / notebook if the pipeline changes
- Use clear, descriptive commit messages (preferably Conventional Commits)
Please use GitHub Issues to report bugs or suggest improvements. Include:
- A clear description of the problem
- Steps to reproduce (with code snippets if applicable)
- Expected vs. actual behaviour
- Environment details (OS, Python version, PyTorch version)
This project is licensed under the MIT License — see the LICENSE file for details.
You are free to use, modify, and distribute this code for personal, academic, or commercial purposes, provided proper attribution is given.
-
Rocha, B.M., et al. (2019). An Open Access Database for the Evaluation of Respiratory Sound Classification Algorithms. Physiological Measurement, 40(3), 035001. DOI
-
Perna, D., & Tagarelli, A. (2019). Deep Auscultation: Predicting Respiratory Anomalies and Diseases via Recurrent Neural Networks. IEEE CBMS 2019.
-
Nguyen, T., et al. (2020). Lung Sound Classification Using Co-tuning and Stochastic Normalization. IEEE TNSRE.
-
Ma, Y., et al. (2021). LungBRN: A Smart Digital Stethoscope for Detecting Respiratory Disease Using bi-ResNet Deep Learning Algorithm. IEEE BioCAS 2021.
-
Global Initiative for Chronic Obstructive Lung Disease (GOLD). (2024). Global Strategy for the Diagnosis, Management, and Prevention of COPD. goldcopd.org
-
WHO. (2023). Chronic obstructive pulmonary disease (COPD) Fact Sheet. who.int
- ICBHI 2017 Challenge organizers and contributors for making the respiratory sound database publicly available
- Prof. Pramod Kumar PM — for continuous guidance, mentorship, and technical feedback throughout the project
- VCET Puttur & VTU — for academic infrastructure and support
- PyTorch, Librosa, and scikit-learn open-source communities
- Prior researchers on the ICBHI 2017 benchmark whose published results guide our baselines
Built with 🫁 + 💻 by Team COPD-Detection — VCET Puttur, 2025–26
