Decoupled GA feature selection for robust voice pathology detection in data constrained settings

Iranian Journal of Science and Technology, Transactions of Electrical Engineering · Springer Nature · 2026-11-19 · DOI: 10.1007/s44163-026-02597-y

Аннотация

Automatic detection of voice pathologies from acoustic signals provides a non-invasive and cost-effective solution for early clinical screening. However, developing feature-based detection systems remains challenging under limited labeled data, high intra-class variability, and heterogeneous acoustic manifestations of voice disorders. This study proposes a signal-processing-driven machine-learning framework for voice pathology detection using the Hospital Universitario Príncipe de Asturias (HUPA) corpus, which contains sustained /a/ phonations from healthy individuals and patients with organic laryngeal disorders. A heterogeneous acoustic representation is constructed by integrating complementary cepstral, perturbation, and glottal descriptors to capture multiple aspects of vocal function. To address feature redundancy and high dimensionality, a genetic algorithm (GA)-based feature selection strategy is applied within a strict leakage-free stratified 5-fold cross-validation framework. The methodological design decouples feature subset discovery from final classifier selection: SVM-RBF, Random Forest, and XGBoost are used as complementary wrapper estimators during GA optimization, while the downstream classifier is optimized independently. This design allows the selected acoustic representation to be evaluated across different classification models rather than being tied to a single optimization estimator. Among the evaluated classifiers, XGBoost achieves the highest observed performance, with an accuracy of 89.86% ± 2.35 and a ROC-AUC of 0.9569, compared with 88.02% ± 2.95 and 0.9478, respectively, using the original 88-feature representation. The GA reduces the representation to 25 features. The observed performance differences were not statistically significant according to Wilcoxon signed-rank tests across the five outer folds and are therefore interpreted as empirical differences rather than evidence of statistically established superiority or robustness. Overall, the results provide an empirical assessment of combining heterogeneous acoustic representations, GA-based feature selection, and decoupled classifier evaluation for voice pathology detection under data-constrained conditions.

Читать у издателя →

Похожие статьи