🤖 AI

A Hierarchical Feature Engineering Framework for Automated Classification of Phonotraumatic and Non-Phonotraumatic Vocal Hyperfunction

This study proposes a hierarchical feature engineering framework that leverages coupling features to effectively distinguish phonotraumatic and non-phonotraumatic vocal hyperfunction from healthy controls using ambulatory neck-surface acceleration data, achieving strong classification performance particularly for the non-phonotraumatic subtype through non-linear feature interactions.

Original authors: June-Woo Kim, Kangwook Jang, Minu Kim, Hyunju Lee

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: June-Woo Kim, Kangwook Jang, Minu Kim, Hyunju Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your voice is like a car engine. Sometimes, the engine runs rough because of a physical scratch or dent (like a stone in the tire). Other times, the engine runs rough because the driver is gripping the steering wheel too tightly or driving with too much stress, even though the engine parts look fine.

In the medical world, these two problems are called Phonotraumatic Vocal Hyperfunction (PVH) and Non-Phonotraumatic Vocal Hyperfunction (NPVH).

PVH is like the physical damage (e.g., vocal nodules).
NPVH is like the tension or misuse without physical damage (e.g., muscle tension dysphonia).

The researchers in this paper wanted to build a "smart mechanic" (a computer program) that could listen to a person's voice while they go about their normal day and tell the difference between these two problems and a healthy voice. They used a special sensor worn on the neck that vibrates when you talk, kind of like a fitness tracker for your voice.

Here is how they did it, explained simply:

1. The "Feature" Recipe

Instead of just listening to the voice, the computer had to break it down into tiny ingredients, which the researchers called features. They built a "hierarchical" (layered) recipe with four types of ingredients:

Static Ingredients (The Snapshot): These are simple averages. "On average, how loud was the voice?" or "How deep was the pitch?"
Dynamic Ingredients (The Movie): These look at how the voice changes over time. "Did the pitch jump up and down wildly?" or "Did the volume get shaky?"
Ratio Ingredients (The Balance): These compare two things. "Is the pitch changing faster than the volume?"
Coupling Ingredients (The Teamwork): This is the secret sauce. It looks at how different parts of the voice system work together. For example, it checks if the effort to push air out matches the sound coming out. It's like checking if the driver's foot pressure on the gas pedal matches the car's speed.

2. The Two Challenges

The researchers tested their "smart mechanic" on two different jobs:

Job 1: Distinguish PVH (Physical damage) from healthy voices.
Job 2: Distinguish NPVH (Tension/No damage) from healthy voices.

3. The Results: A Tale of Two Tasks

Job 1 (The Physical Damage) was easy to spot.
The computer found that people with physical damage (PVH) had very obvious, loud "tells" in their voice data. Even simple measurements (like just the average pitch) were enough to spot them. When the researchers added the complex "Coupling" ingredients (the teamwork checks), the computer got even better, reaching a 89% success rate (AUC of 0.891).

Analogy: It's like spotting a car with a flat tire. You can see it from far away, and checking the tire pressure confirms it.

Job 2 (The Tension) was much harder.
The computer struggled to find the "tension" type of voice problem. When they just looked at simple averages, the computer was basically guessing (55% success). Even with the complex "Coupling" ingredients, it only reached 73% success (AUC of 0.728).

Analogy: This is like trying to tell if a driver is stressed just by looking at the car's speedometer. The car might look fine, but the driver is gripping the wheel too hard. The "tells" are very subtle and hidden.

4. What They Learned

The study found a big difference between the two conditions:

PVH leaves a clear, loud fingerprint on the voice data. You can find it with simple math.
NPVH is much sneakier. It doesn't leave a single loud fingerprint. To find it, the computer needs to look at how different voice parts interact with each other (the "Coupling" features) and use complex math to find patterns that simple averages miss.

5. The Final Test

The researchers tested their best model on a brand-new set of data it had never seen before (the "Held-Out Test Set").

For the Physical Damage (PVH), it was excellent, scoring 91%.
For the Tension (NPVH), it dropped significantly to 58%, which is barely better than a coin flip.

The Bottom Line

The paper concludes that while we have built a very good tool for spotting physical voice damage using neck sensors, spotting "tension" voice problems is still a mystery. The "tension" doesn't leave a clear, single signal; it hides in the complex, subtle relationships between different parts of the voice. The researchers suggest that in the future, we might need to look at the raw sound waves directly (like listening to the engine's hum) rather than just breaking it down into numbers, to finally crack the code on the tension problem.

Technical Summary: A Hierarchical Feature Engineering Framework for Automated Classification of Phonotraumatic and Non-Phonotraumatic Vocal Hyperfunction

Problem Statement
Vocal hyperfunction (VH) is a chronic condition involving voice misuse, clinically categorized into phonotraumatic VH (PVH, e.g., nodules) and non-phonotraumatic VH (NPVH, e.g., muscle tension dysphonia). While ambulatory neck-surface acceleration (ACC) offers a non-invasive method for monitoring daily vocal behavior, robust biomarkers for distinguishing these subtypes from healthy controls remain limited. Existing approaches often rely on simple time-averages of individual acoustic or aerodynamic features, potentially overlooking the dynamic, interdependent nature of speech production and the physiological coupling between aerodynamic effort and acoustic output. The challenge lies in the high variability of daily acoustic and aerodynamic measures, which makes separating PVH and NPVH subtypes from controls difficult using standard univariate metrics.

Methodology
The study utilizes the NeckVibe Challenge dataset, comprising 3,278 samples from 468 subjects, to address two binary classification tasks: PVH vs. Control (Task 1) and NPVH vs. Control (Task 2). The core methodology involves a hierarchical feature engineering framework that transforms 14 core physical quantities (6 acoustic measures and 8 Impedance-based Inverse Filtering [IBIF] derived aerodynamic measures) into subject-level representations.

The framework constructs four distinct feature categories:

Static Features: Seven statistical descriptors (mean, SD, percentiles, skewness, kurtosis, IQR) applied to the 14 base variables plus vocal dose percentage, totaling 99 features.
Dynamic Features: First-order ( $\Delta$ ) and second-order ( $\Delta\Delta$ ) difference statistics, including delta means, SDs, upper-tail magnitudes, and linear trends, adding 197 features.
Ratio-based Features: Normalized variability measures (e.g., $\Delta$ SD/mean, $|slope|$/mean) to quantify temporal variability relative to absolute levels, resulting in 253 features.
Coupling Features: Six physiologically motivated interaction terms capturing source-filter relationships (e.g., cppall/spectralTiltall) and stability-effort couplings, extending the set to 203 features.

Experimental Pipeline:

Statistical Analysis: Univariate comparisons were performed using Welch's t-tests with Benjamini-Hochberg False Discovery Rate (FDR) correction to identify significant group differences.
Machine Learning: A stratified 10-fold group cross-validation (CV) was employed, ensuring all recordings from a single subject remained within either the training or validation set.
Feature Selection: Recursive Feature Elimination with Cross-Validation (RFECV) using an XGBoost estimator was applied within training folds to mitigate redundancy.
Classification: Five models (Logistic Regression, SVM, Random Forest, XGBoost, LightGBM) were evaluated.
Interpretability: SHapley Additive exPlanations (SHAP) were used to analyze feature contributions.

Key Results
The study reveals a significant asymmetry between the two classification tasks:

Task 1 (PVH vs. Control):
- Statistical: Robust separability was observed. After FDR correction, a substantial number of features remained significant, particularly dynamic and higher-order descriptors, with large effect sizes ( $|d| \approx 0.9–1.0$ ).
- Performance: The model achieved an AUC of 0.891 (Logistic Regression with coupling features) on the cross-validation set and 0.917 on the held-out test set.
- Insight: PVH appears near-linearly separable, with performance improving progressively as feature complexity (dynamic and coupling) increases. SHAP analysis indicated distributed contributions across multiple features.
Task 2 (NPVH vs. Control):
- Statistical: No features survived FDR correction. Only features with nominal significance ( $p < 0.05$ ) were observed, with small effect sizes ( $|d| \approx 0.3$ ).
- Performance: Performance was significantly lower. Static features performed near chance (AUC 0.556). The best result was achieved with LightGBM using coupling features, yielding an AUC of 0.728 (CV) and 0.579 (held-out test).
- Insight: NPVH discrimination benefits from modeling non-linear feature interactions (coupling features), but the overall separability remains limited. SHAP analysis showed concentrated contributions from dynamic variability measures but higher variance across folds, suggesting a less stable discriminative structure.

Key Contributions

Hierarchical Framework: Proposes a structured feature engineering approach (static, dynamic, ratio, coupling) specifically designed to capture the complex physiological mechanisms of vocal hyperfunction.
Statistical Rigor: Provides a rigorous univariate analysis demonstrating that while PVH exhibits strong, distributed statistical differences, NPVH shows minimal separability at the daily summary level.
ML Pipeline Validation: Demonstrates that while PVH is largely linearly separable, NPVH classification relies on capturing non-linear interactions, though even advanced modeling struggles to achieve high accuracy for NPVH.
Benchmarking: Establishes a competitive baseline for the NeckVibe Challenge, highlighting the practical utility of hierarchical features for PVH biomarker identification while acknowledging the current limitations in NPVH detection.

Significance and Claims
The paper claims that the proposed hierarchical feature engineering framework is effective for identifying biomarkers in phonotraumatic vocal hyperfunction, where distinct physiological signatures are captured by coupling and dynamic features. However, the authors modestly conclude that non-phonotraumatic VH remains challenging to classify using current aerodynamic and acoustic measures. They attribute this difficulty to the lack of structural lesions in NPVH, leading to substantial distributional overlap with healthy controls.

The study suggests that while multivariate modeling partially mitigates the limitations for NPVH, the current feature set may not capture the subtle, non-stationary prosodic variations or micro-tremors characteristic of NPVH. Consequently, the authors posit that future advancements for NPVH may require self-supervised speech foundation models trained on raw waveforms to uncover hidden latent representations, rather than relying solely on manually engineered features.