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

Researchers developed a hierarchical feature engineering framework to classify vocal hyperfunction subtypes using non-invasive neck-surface acceleration monitoring. The machine learning approach achieved 89.1% AUC for phonotraumatic cases and 72.8% for non-phonotraumatic cases, with coupling features proving crucial for distinguishing both conditions from healthy controls.

This research addresses a clinical gap in automated vocal health monitoring by leveraging machine learning to differentiate vocal hyperfunction subtypes from acceleration biomarkers. The study's significance lies in its methodological approach—rather than relying solely on traditional statistical metrics, the researchers constructed a comprehensive feature hierarchy incorporating static measurements, dynamic patterns, ratio-based calculations, and source-filter coupling interactions. This multi-layered architecture reveals fundamental differences in how the two vocal conditions manifest in biometric data.

The performance disparity between phonotraumatic and non-phonotraumatic hyperfunction classifications reflects underlying physiological complexity. Phonotraumatic cases show near-linear separability, suggesting the condition produces distinct, measurable acoustic signatures. Non-phonotraumatic hyperfunction requires nonlinear modeling, indicating more subtle biomechanical patterns. This finding has broader implications for medical AI development—different pathological conditions demand tailored feature engineering strategies rather than one-size-fits-all algorithms.

For healthcare technology developers and voice pathology specialists, this framework demonstrates how wearable accelerometer data can enable objective, continuous monitoring outside clinical settings. The hierarchical approach provides a blueprint for feature selection in other vocal or musculoskeletal conditions. However, the NPVH's lower AUC (0.728) highlights current limitations in distinguishing functional disorders from healthy states, suggesting future work should explore additional sensor modalities or longer-duration monitoring.

The research positions automated vocal monitoring as a viable clinical tool, pending validation on larger, diverse populations. Success here could enable earlier intervention for voice disorders, reducing long-term tissue damage and improving treatment outcomes.

• →Hierarchical feature engineering combining static, dynamic, ratio, and coupling features improves vocal hyperfunction classification accuracy
• →Phonotraumatic hyperfunction achieves 89.1% AUC, suggesting near-linear separability in biometric data
• →Non-phonotraumatic hyperfunction requires nonlinear feature interactions, achieving 72.8% AUC and indicating more subtle pathological patterns
• →Neck-surface acceleration monitoring enables non-invasive, continuous vocal health assessment outside traditional clinical settings
• →Coupling features capturing source-filter interactions prove essential for distinguishing both vocal hyperfunction subtypes from healthy controls