The Rise in Artificial Intelligence and Machine Learning Models to Screen for Cleft-Related Velopharyngeal Dysfunction: A Systematic Review.

In 30 seconds

This PRISMA-guided systematic review examined AI and machine-learning models for detecting velopharyngeal dysfunction (VPD) in people with cleft palate. Of 455 articles screened, 34 studies were included, representing 3967 participants and 92,323 training samples; participant ages ranged from 1 to 93 years. Across studies reporting performance, midpoint estimates showed mean accuracy of 82.9%, sensitivity of 80.5%, and specificity of 82.2%. External validation was reported in only 3 of 34 studies, and the review concluded that no clinically deployable model has yet been achieved.

Key findings

• Across studies reporting metrics, midpoint estimates were mean accuracy 82.9%, precision 86.7%, F1-score 0.88, sensitivity 80.5%, specificity 82.2%, and PCC 0.58.
• Support vector machines were used most often (16/34, 47.1%), followed by convolutional neural networks (6/34, 17.6%).
• Only 3 studies (3/34, 8.8%) reported external validation.

Why it matters

Automated analysis of speech features may support VPD screening, and the reported performance suggests potential for further development. However, the limited external validation and inconsistent reporting make it difficult to know whether results will generalize across populations and settings. The review’s conclusion that no model is yet clinically deployable is important when interpreting these performance estimates.

What to keep in mind

The review identifies inconsistent reporting, reliance on engineered features, and limited external validation as constraints on generalizability.