Automatic Detection of Neurodevelopmental Disorders Using CryAcoustic Features

Abstract

Prompt detection of neurodevelopmental disorders (NDDs) is vital for launching timely interventions that can markedly improve developmental outcomes. In our study, we investigate whether infant cry sounds could serve as non-invasive biomarkers for NDD detection. We focus on CryAcoustic Features, which combine prosodic elements with selected voice-quality parameters. Our approach began with the extraction of a broad set of acoustic characteristics from infant cry recordings using the eGeMAPS feature set. We then applied a rigorous feature selection process to identify the most informative descriptors. For the classification task, we assessed several machine learning techniques, including Support Vector Machine (SVM) and Random Forest (RF). In addition, we developed a hybrid model that fuses the outputs of SVM and RF via logistic regression acting as a meta-classifier. This integrated approach yielded the best performance, achieving an accuracy of 75.52%, a recall of 80.05%, and an AUC of 83.00%. Our findings highlight the effectiveness of CryAcoustic Features in distinguishing between typically developing infants and those with NDDs, suggesting that cry-based acoustic analysis is a promising, noninvasive strategy for initial NDD screening.