This project implements a hybrid classical-quantum neural network to detect respiratory diseases using lung sound signals. We used the ICBHI 2017 dataset and applied advanced audio preprocessing and data augmentation, followed by classification using quantum circuits integrated with deep learning.

Early diagnosis of respiratory illnesses is crucial due to their high global mortality rate. Traditional methods often rely on manual auscultation. This project automates the classification of lung diseases based on sound signals using quantum machine learning.

• Source: ICBHI 2017 Challenge Dataset
• Patients: 126
• Audio Files: 920
• Diagnosis Classes: 6

Each audio file ranges from 10–90 seconds and has varying sampling rates. Main problem with the dataset: class imbalance

• Resampling to 4 KHz
• Snippet Generation: Break long files into smaller time segments or snippets
• Baseline Wandering Removal using Discrete Fourier Transform (0–1 Hz noise eliminated)
• Amplitude Normalization to range [-1, 1]

To balance classes, we applied:

• Time Stretching: 0.7x and 0.9x speed
• Pitch Shifting: −2 and +1 semitones
• Noise Addition: White noise to minority class signals

• Converted processed audio to mel-spectrogram images (RGBA format)
• Resized to 4×4 to match quantum circuit input

A 3-stage hybrid model:

1. Feature Extraction

   • Convolutional + MaxPooling layers extract spatial features

2. Quantum Processing

   • 4-qubit quantum circuit (16 output probabilities)
   • Features encoded as rotation angles (θ₀ to θ₇) for quantum gates

3. Classification

   • Fully connected layer maps quantum output to disease probabilities using softmax

• Training Accuracy: 97.68%
• Testing Accuracy: 95.66%
• Confusion matrices for both datasets included in the notebook

• Languages: Python
• Libraries: NumPy, SciPy, TorchAudio, Matplotlib, pandas
• Deep Learning framework: PyTorch
• Quantum ML: Qiskit
• Notebook: Google Colab