TL;DR
This paper introduces GeHirNet, a novel two-stage hierarchical deep learning model for classifying voice pathologies. By first distinguishing between male and female pathological voices and then classifying specific diseases within each gender, GeHirNet significantly improves accuracy and reduces gender bias, outperforming existing methods. This approach is particularly effective in addressing class imbalance and gender-specific acoustic variations in voice data.
What Problem Does This Paper Address?
Accurate classification of voice pathologies is crucial for diagnosis and treatment, but current AI-based methods face significant challenges. One major hurdle is the inherent acoustic differences between male and female voices, which can lead to biased performance. Another critical issue is the class imbalance in datasets, where rare pathologies are underrepresented, making it difficult for models to learn effectively. Existing approaches often treat gender as a secondary factor or fail to adequately address data imbalance, leading to suboptimal performance, especially for less common conditions. This paper aims to tackle these limitations by proposing a model that explicitly incorporates gender awareness into the classification pipeline.
Methods
The core of this work is the GeHirNet architecture, a two-stage hierarchical model.
- Stage 1: Gender-Based Pathology Detection: The first stage uses a ResNet-50 backbone to classify audio samples into four categories: male healthy control (HCM), female healthy control (HCF), male pathology (PM), and female pathology (PF). This initial step aims to separate gender-specific pathological patterns.
- Stage 2: Gender-Specific Pathology Classification: In the second stage, separate classifiers (also based on ResNet-50) are trained for male and female pathology groups. These classifiers then distinguish between six specific diseases (COVID-19, Parkinson's Disease, Dysphonia, Vocal Cord Paresis, Laryngitis, and ALS).
To address class imbalance, the authors employ two data augmentation techniques:
- Multi-scale Resampling: This technique adjusts the sampling rate of audio data to equalize the representation of minority classes.
- Time Warping: Applied directly to audio segments, this method involves shuffling segments of a recording to create variations, helping to generalize the model to unseen data.
The model utilizes Mel spectrograms as input features, which are known to capture spectral information effectively. Preprocessing steps include silence removal, crossfade for smooth transitions, outlier removal, and min-max normalization.
Datasets
The study utilizes a merged dataset from four publicly available sources, focusing on sustained vowel /a/ recordings:
- Coswara: Primarily for COVID-19 classification.
- ALS (Russian): For ALS diagnosis.
- PC-GITA (Spanish): For Parkinson's disease research.
- SVD (German): For a variety of voice disorders, including Laryngitis, Vocal Cord Paresis, and Dysphonia.
The paper does not introduce a new benchmark dataset but rather consolidates existing ones to create a more comprehensive and diverse dataset for their gender-aware analysis.
Key Insights and Ideas
- Gender as a First-Class Citizen: The most significant insight is the effectiveness of treating gender as a primary classification factor before delving into specific pathologies. This hierarchical approach directly tackles gender-specific acoustic variations that confound single-stage classifiers.
- Synergy of Hierarchy and Augmentation: The combination of the two-stage hierarchical architecture with data augmentation techniques (especially time warping) yields substantial performance gains, highlighting that addressing both structural bias (gender) and data scarcity (imbalance) is crucial.
- Deep Layers Capture Gender Specificity: Analysis using Centered Kernel Alignment (CKA) suggests that while shallow layers of the ResNet capture universal acoustic features, deeper layers learn more gender-specific patterns, justifying the gender-aware approach.
- Time Warping's Efficacy: Time warping, applied directly to audio, proved to be a highly effective augmentation strategy, leading to the best overall performance, suggesting its suitability for sustained vowel recordings where semantic context is less critical.
Key Results
- State-of-the-Art Performance: The proposed GeHirNet** (two-stage with time warping) achieved the highest performance, with an accuracy of 97.63%, F1 score of 97.61%, and MCC of 95.25%.
- Significant Improvement over Baseline: GeHirNet** improved MCC by 5% over the single-stage baseline (0.9525 vs. 0.9041).
- Gender-Specific Advantages: The two-stage architecture (GeHirNet) alone outperformed the baseline by 3% in MCC (0.9363 vs. 0.9041).
- High Healthy Control Detection: GeHirNet** achieved an impressive 99.3% accuracy in detecting healthy controls across datasets, demonstrating strong generalization.
- Competitive Disease-Specific Accuracy: The model matched or exceeded prior state-of-the-art (SOTA) performance for all six target diseases.
Limitations and Future Work
- Data Augmentation Focus: While data augmentation was effective, the paper notes that further exploration of advanced techniques could yield even better results.
- Computational Cost: Training separate classifiers in the second stage, while effective, increases computational complexity. Future work could explore more integrated or efficient ways to handle gender-specific classification.
- Generalizability to Diverse Speech: The study focused on sustained vowel /a/ recordings. Future work could investigate the applicability of GeHirNet to more complex speech tasks, such as continuous speech or conversational audio.
- Clinical Validation: While the model shows high accuracy, further validation in real-world clinical settings would be necessary to confirm its practical utility.
Summary updated