Related Experiment Video
Updated: Jan 10, 2026
Targeted Training of Ultrasonic Vocalizations in Aged and Parkinsonian Rats
Published on: August 8, 2011
Vocal Biomarkers for Parkinson's Disease Classification through Hybrid Feature Selection with Beluga Whale
Umashankar Ganesan1, Lumen Christy V2, Sheeba Santhosh3
1Department of Biomedical Engineering, GRT Institute of Engineering & Technology, Tiruttani 631209, India.
Introduction:
Parkinson's disease (PD) is a progressive neurodegenerative disorder that often causes vocal impairments. As a result, voice analysis is a promising avenue for early and objective diagnosis. While voice is a well-established biomarker, the high dimensionality of acoustic feature sets can lead to model overfitting. This necessitates effective feature selection strategies.
Objective:
The objective is to validate a hybrid feature selection pipeline that combines statistical filters with a metaheuristic wrapper algorithm. The aim is to identify a minimal, yet highly discriminative, set of acoustic features for PD classification. This approach aims to strike a balance between dimensionality reduction and high diagnostic accuracy.
Methods:
This study validates a hybrid feature selection pipeline. It combines statistical filters such as Spearman correlation (Corr) and mutual information (MI) with a metaheuristic beluga whale optimization (BWO) wrapper. The framework's efficacy was benchmarked across a diverse suite of machine learning classifiers. Evaluation was conducted using 10-fold stratified group cross-validation on a dataset of 40 PD patients and 40 healthy controls. The process dynamically identifies parsimonious feature subsets within each validation fold.
Results:
When paired with a random forest classifier, the Corr+MI+BWO pipeline achieved a peak F1-score of 0.966 using an average of only 21.2 features. This represents a 73% reduction in dimensionality with a minimal 2.4% performance trade-off compared to a standard filter-based approach. The most consistently selected features proved to be both statistically significant (P < 0.05) and part of a clinically coherent signature of Parkinsonian dysarthria.
Conclusion:
The proposed hybrid framework is an effective methodology that successfully balances model simplicity with high predictive accuracy. This research provides a strong foundation for the development of objective, noninvasive tools for the early detection and monitoring of PD.
Data Availability:
This study utilized a publicly available dataset of voice recordings from individuals with Parkinson's disease and healthy controls obtained from Figshare, available at the following link: https://figshare.com/articles/dataset/Voice_Samples_for_Patients_with_Parkinson_s_Disease_and_Healthy_Controls/23849127.
