Predicting Vocal Pathologies: A Multi-Dataset Machine Learning Framework for Healthcare Applications

Vocal pathologies affect vocal fold dynamics, altering pitch, loudness, and other voice traits. Early detection requires accurate prediction methods, but limited datasets and class imbalances pose challenges. To address these, we use data augmentation and synthetic minority oversampling to enhance dataset diversity and balance, ensuring model generalizability across demographics. Our study employs a machine learning paradigm for diagnosing vocal disorders using sustained vowel /a/ samples from four datasets in German, Spanish, English, and Arabic. Extracted features include mel-frequency cepstral coefficients, delta, delta-delta, and log-energy. Multiple classifiers were tested, with support vector machine, stochastic gradient descent, and long short-term memory models achieving accuracy over 97% and area under the curve scores near 1. These findings highlight the method's ability to uncover clinical voice patterns, offering a reliable and efficient tool for assessing vocal disorders.