Artificial Intelligence-Based Hypernasality Diagnosis Using CAPS-A-AM Rated Speech Samples in Pediatric Velopharyngeal Dysfunction

ObjectivePerceptual evaluation by speech-language pathologists (SLPs) is essential for initial evaluation of velopharyngeal dysfunction (VPD). Machine learning (ML) offers a promising avenue for developing accessible speech assessment tools when SLP expertise is limited. We aimed to develop a workflow and preliminary algorithm for AI-based hypernasality detection based on the Cleft Audit Protocol for Speech-Augmented-Americleft Modification (CAPS-A-AM), a standardized framework for auditory perceptual speech assessment.DesignIn this prospective, single-center study, speech samples were collected during SLP-guided evaluation, with consensus CAPS-A-AM ratings established. Mel spectrograms for high vowels (/i/ and /u/) were generated from sustained vowels, isolated words, and sentences for model development. Three ML approaches were evaluated: logistic regression (LR), Convolutional Neural Network (CNN) Attention-Multiple Instance Learning (MIL) (EfficientNet-V2-S), and a CNN-Extreme Gradient Boosting Hybrid (XGBoost).Patients/ParticipantsForty pediatric participants aged 2 to 17, including individuals with VPD, conditions associated with VPD, and healthy participants.Main Outcome Measure(s)Model performance was tested in binary hypernasality classification compared to SLP consensus at two CAPS-A-AM thresholds: absent (0) versus any hypernasality (1-4) and absent/borderline (0-1) versus mild-to-severe hypernasality (2-4).ResultsMultiple independent modeling approaches were able to detect clinically rated hypernasality. EfficientNet-V2-S achieved the highest observed performance estimates in this cohort (F1 score = 0.786-0.900). However, overlapping confidence intervals precluded demonstration of model superiority.ConclusionML-based hypernasality classification using mel spectrograms demonstrated promising feasibility. Future work will expand sample acquisition and refine model development with increasingly diverse speech inputs for training and sample rating.