A Deep Learning Algorithm for Objective Assessment of Hypernasality in Children With Cleft Palate.

Area of Science:

• Speech-language pathology
• Biomedical engineering
• Artificial intelligence in healthcare

Background:

• Evaluating hypernasality necessitates extensive clinician training, limiting global accessibility and exacerbating health disparities for children with clefts.
• Current diagnostic methods are resource-intensive and difficult to scale internationally.

Purpose of the Study:

• To introduce and validate an objective hypernasality measure (OHM), an automated speech-based algorithm.
• To assess the OHM's performance against trained clinicians' perceptual ratings.

Main Methods:

• A deep neural network (DNN) was trained on healthy speech to identify nasal acoustic cues.
• The model aggregated posterior probabilities to compute the OHM, without requiring clinical data for training.
• Validation involved correlation with established hypernasality databases (Americleft, NMCPC).

Main Results:

• The OHM demonstrated significant correlation with expert perceptual ratings (r = 0.797 for Americleft, r = 0.713 for NMCPC).
• The study evaluated OHM's sensitivity for mild hypernasality and its internal reliability.
• OHM performance was comparable to a DNN regression model trained directly on hypernasal speech.

Conclusions:

• The objective hypernasality measure (OHM) effectively quantifies hypernasality severity.
• OHM performance is on par with trained clinicians, offering a scalable and accessible diagnostic tool.
• This technology has the potential to reduce health disparities in cleft care.