Machine Learning-Based Detection of Voice Disorders Using Large-Scale Clinical Data

Background and Objectives This study compared four machine learning models in differentiating normal from pathological voice using key acoustic indices (Cepstral Peak Prominence Smoothed [CPPS], Acoustic Voice Quality Index [AVQI], and Acoustic Breathiness Index [ABI]) and identified the most effective algorithm for this task.Materials and Method Voice samples from 11661 adults (normal: 11.0%, pathological: 89.0%) at a university hospital voice clinic were analyzed using Gradient Boosting, Logistic Regression, Neural Network, and Random Forest. Model performance was evaluated via 10-fold stratified cross-validation using area under the curve (AUC), accuracy, F1-score, recall, and specificity, with particular emphasis on specificity to minimize false-positive classifications in clinical screening.Results All four models achieved high discriminability (AUC ≥0.976), with notable differences observed in specificity across models. Gradient Boosting demonstrated the highest overall performance, followed by Neural Network and Random Forest, while Logistic Regression showed substantially lower specificity (0.624), misclassifying a considerable proportion of normal voices as pathological. Variable importance analysis identified CPPS as the most influential acoustic index across all models, with AVQI and ABI serving complementary roles.Conclusion Gradient Boosting most effectively differentiated normal from pathological voice in a large, imbalanced dataset, suggesting that ensemble and deep learning models demonstrate potential as useful supplementary tools for voice disorder diagnosis, with CPPS identified as the most informative acoustic index.