Abstract
Common voice disorders are often chronic conditions resulting from inefficient or abusive vocal patterns, known as vocal hyperfunction. Traditional clinical assessments provide only a "snapshot" of vocal behavior, failing to capture the complexities of daily life. This research update details an ambulatory monitoring system using a neck-worn accelerometer to quantify detrimental vocal behaviors in real-world settings. By applying biological acoustic modeling and machine learning to a large dataset of patients and controls, the study demonstrates the potential for improved diagnosis and treatment of hyperfunctional voice disorders.
Introduction: The Challenge of Vocal Hyperfunction
Voice disorders affect nearly 30% of adults at some point, frequently caused by "vocal hyperfunction"—inefficient or abusive patterns of vocal behavior such as chronic yelling or strained speaking. A significant hurdle in clinical management is that laboratory assessments are often unrepresentative of a patient's typical behavior. Patients may exhibit healthy vocal patterns during a brief clinical exam but revert to damaging habits in their daily environments. Objective, unobtrusive monitoring during "activities of daily life" is essential for accurate diagnosis and effective long-term therapy.
Methods: A Wearable Approach to Vocal Health
The researchers developed a wearable system consisting of a miniature accelerometer placed on the neck, below the larynx, tethered to a smartphone. This configuration offers a critical privacy advantage: the sensor captures vocal fold vibrations without recording intelligible speech, allowing for use in sensitive environments like schools or offices.
The study utilizes three primary analysis frameworks:
- Ambulatory Voice Measures: Standard metrics including Sound Pressure Level (SPL), fundamental frequency ($f_0$), and "vocal dose" (the cumulative duration of vocal fold vibration).
- Subglottal Impedance-Based Inverse Filtering (IBIF): A biologically inspired model that estimates glottal airflow from neck-surface acceleration. This allows for the calculation of aerodynamic properties, such as the speed of vocal fold closure, without the need for intrusive airflow masks.
- Machine Learning Classification: Utilizing LASSO (Least Absolute Shrinkage and Selection Operator) logistic regression to distinguish between patients and healthy controls based on the statistical distribution of vocal features collected over a week.
Datasets and Study Design
The research involves one of the largest cohorts in this field, comprising 142 subjects. The primary analysis focused on two main groups:
- Phonotraumatic Hyperfunction: 51 female pairs (patient and matched healthy control) involving conditions like nodules or polyps.
- Non-phonotraumatic Hyperfunction: 20 female pairs involving Muscle Tension Dysphonia (MTD).
Participants were predominantly in high-vocal-demand occupations, such as teachers, singers, and consultants.
Key Insights: Beyond Simple Averages
A major finding of the study is that simple averages of pitch or volume are often insufficient to distinguish patients from healthy speakers. Instead, the differences reside in the statistical "tails" of the data. Patients with vocal fold lesions exhibited "right-shifted" distributions, indicating significant deviations from their baseline as the day progressed—likely a reflection of increased vocal effort due to fatigue.
Furthermore, the validation of the IBIF model in ambulatory settings proved that internal laryngeal physics can be accurately estimated using skin-surface sensors. The relationship between estimated airflow and sound pressure in the field (1.13 dB/dB) closely mirrored laboratory findings, providing a robust "sanity check" for the ambulatory aerodynamic model.
Results and Clinical Potential
The machine learning model achieved a classification accuracy of 72.5% (74 out of 102 subjects) for the phonotraumatic group. The most significant predictors for identifying patients were the skewness of SPL and the 95th percentile of fundamental frequency. These results suggest that patients with nodules or polyps do not necessarily speak more than controls, but they speak with higher effort and greater variability in pitch and volume.
Future Directions: Real-Time Biofeedback
Current models assume "strong labeling," where all data from a patient is treated as disordered. Future work aims to refine these labels to identify specific bursts of vocal abuse. The ultimate goal is the implementation of real-time biofeedback, where the smartphone provides haptic or visual alerts when a user exhibits damaging vocal patterns, facilitating immediate behavioral change and bridging the gap between clinical therapy and daily life.
Summary updated