Unlocking Voice Health: How Quantum Machine Learning is Revolutionizing Dysphonia Assessment

Quantum Approaches for Dysphonia Assessment

0:00

Speech is a fundamental part of human communication, but for many, conditions like dysphonia can make it incredibly challenging. Dysphonia, a prevalent medical condition affecting about 10% of the general population and up to 50% of professional voice users, leads to symptoms like hoarseness, a weak or breathy voice, strained speech, or even complete voice loss. These symptoms are typically caused by malformations or dysfunctions of the vocal cords or the larynx.

Traditional dysphonia assessment methods, while effective, often involve medical history reviews, symptom assessments, laboratory examinations, and equipment like laryngoscopes or imaging tests (CT/MRI scans). However, these methods can be time-consuming, uncomfortable for patients, and expensive, with annual diagnostic and management costs potentially ranging from $577 to $953 USD per patient.

To address these challenges, researchers have increasingly turned to machine learning (ML), which has made significant strides in analyzing voice signals for dysphonia assessment. While traditional ML algorithms like Random Forest and Support Vector Machines have shown high accuracy, they often require significant expertise in feature selection.

The Deep Learning Promise and its Dataset Problem Deep learning methods, particularly Convolutional Neural Networks (CNNs), have demonstrated great promise in speech classification and assessing pathological speech, achieving accuracies like 82.33%. However, a major hurdle for deep neural networks in medical applications is the limited availability of high-quality medical data. Rare conditions and privacy concerns make obtaining large datasets difficult, often leading to performance challenges for CNNs. While techniques like oversampling or combining pre-trained CNNs with other classifiers have been explored to mitigate this, the search for more robust solutions continues.

Enter Quantum Machine Learning: A Novel Approach In recent years, quantum machine learning (QML) has emerged as a powerful field, offering potential solutions to computational bottlenecks faced by classical ML algorithms, especially when dealing with small datasets. Quantum algorithms have shown an ability to perform well with limited training samples.

This new research represents a novel and significant step: it is the first prior work on investigating the benefits of quantum approaches specifically for detecting voice and speech disorders.

Introducing Quanvolutional Neural Networks (QNNs) This study introduces a novel hybrid quantum-classical architecture called Quanvolutional Neural Networks (QNNs) for dysphonia assessment. QNNs combine the strengths of both quantum and classical technology.

Key advantages of QNNs:

• Enhanced feature extraction: QNNs leverage quantum convolutional (quanvolutional) layers which transform input data using quantum circuits, extracting more complex features than classical methods.

• Suitability for small datasets: QNNs are particularly well-suited for situations with limited data, a common challenge in medical research.

• Near-term quantum hardware compatibility: QNNs operate effectively in the current "Noisy Intermediate-Scale Quantum (NISQ) era" because they require only shallow quantum circuits operating on small subsections of data, avoiding the need for large-scale quantum resources.

How a Quanvolutional Layer Works: The quanvolutional layer, which replaces a standard convolutional layer in a classical CNN, transforms patches of input data using quantum circuits instead of traditional matrix multiplication. This process involves three main parts:

1. Encoding: Classical input data (like a 2x2 section of a Mel spectrogram image) is transformed into quantum data. This study uses angle encoding, which employs parameterized rotations to map the image's intensity values to quantum states. Angle encoding is known for its higher expressivity.

2. Random quantum circuit: A specially designed quantum circuit, composed of randomly chosen 1-qubit and 2-qubit gates with random parameters, processes the encoded data. These multi-qubit gates are crucial as they generate entanglement, allowing the circuit to exploit quantum correlations that are difficult for classical systems to replicate.

3. Decoding: The output of the quantum circuit is converted back into classical data by measuring the quantum states. These measurements yield expectation values that are mapped to output features, forming a multi-channel output image for subsequent classical layers.

   
   

The Research in Action The study utilized the Perceptual Voice Qualities Database (PVQD), focusing on the sustained /a/ vowel sound from dysphonia patients and healthy individuals. The audio data was preprocessed into Mel spectrograms, which are visual representations of frequencies over time, commonly used in speech recognition and audio classification. The dataset comprised 243 training samples and 61 testing samples.

The researchers developed six models across two scenarios to compare performance:

• Two QNN models (QNN1 and QNN2).

• Two standard CNN models (CNN1 and CNN2).

• Two RANDOM models (RANDOM1 and RANDOM2), which mimic the randomness of quantum circuits with non-trainable random non-linear convolutional layers, serving as a control to understand if observed gains are truly due to quantum properties.

QNN2, RANDOM2, and CNN2 models incorporated an additional classical convolutional layer to assess performance with increased complexity.

Groundbreaking Results The experimental results demonstrated clear advantages for the QNN models:

• Superior Accuracy: QNN models consistently outperformed CNN models in classification accuracy across all four experiments and varying training sample sizes. For example, with only 60 training samples, QNN1 achieved 82.22% ± 4% accuracy, outperforming RANDOM1 (80.00% ± 4%) and CNN1 (76.67% ± 3%). When additional layers were added in Scenario 2, QNN2 maintained the best performance, reaching 89.26% ± 2% accuracy averaged over 120 to 240 training samples.

• Faster Convergence: QNN models also showed faster convergence speed compared to CNN models, stabilizing and reducing loss more efficiently.

• Effectiveness with Limited Data: The results highlight QNNs' particular strength with limited training samples, suggesting they can better leverage quantum properties for learning.

• The Power of Angle Encoding: Unlike previous work, the use of angle encoding in the quanvolutional layer contributed to the enhanced performance of QNNs over the RANDOM models, indicating the benefit of quantum properties beyond just random non-linear transformations.

• Hybrid Approach Success: The study also affirmed the viability of combining classical convolutional and quanvolutional layers to further enhance classification accuracy.

Looking Ahead This research underscores the immense potential of integrating quantum approaches into medical data classification, particularly for conditions like dysphonia. Future work will explore diversifying training data using techniques like adding noise, time-stretching, shifting, and pitch alteration. Experimenting with different encoding and decoding approaches and integrating quantum layers with other classical models will also be key areas of investigation to further optimize the QNN framework.

This study paves the way for more efficient, accurate, and accessible dysphonia assessment, offering a glimpse into the quantum future of healthcare.