MFCC-Based Machine Learning for Noninvasive Detection of Vocal Fold Nodules in Children

Vocal fold nodules are a common pediatric voice disorder that can adversely affect both phonation and psychosocial development. Currently, diagnosis primarily relies on invasive laryngoscopy evaluation, which limits routine application in young children. Acoustic analysis offers a promising, noninvasive alternative for identifying and monitoring vocal fold nodules. This study aimed to evaluate the feasibility of using mel-frequency cepstral coefficient (MFCC)-derived features for the objective, automated detection of vocal fold nodules in children.

A total of 100 children were included, comprising 50 patients with laryngoscopically confirmed vocal fold nodules and 50 age-matched controls. Sustained vowel /u/ phonations were analyzed using MFCC-based acoustic extraction. Group differences in MFCC parameters were assessed. Five machine learning models were developed to classify vocal fold nodules, and feature importance was examined using Shapley Additive exPlanations (SHAP).

Children with vocal fold nodules showed significantly lower MFCC-Std and MFCC-Var values, as well as notable changes in MFCC-Min and MFCC-Skew compared with controls. Among the five machine learning algorithms, the XGBoost model demonstrated the highest classification performance, with an area under the curve of 0.7936. SHAP analysis identified MFCC-Var, MFCC-Min, and MFCC-Skew as the most influential features.

The integration of MFCC-derived acoustic features with an XGBoost classifier enables accurate, noninvasive identification of pediatric vocal fold nodules. These findings support the potential application of voice-based modeling in scalable screening and therapeutic monitoring of pediatric vocal fold nodules.