From Support Vector Machines to Neural Networks: Advancing Automated Velopharyngeal Dysfunction Detection in Patients With Cleft Palate.

BACKGROUND

The generation of intelligible speech is the single most important outcome after cleft palate repair. The development of velopharyngeal dysfunction (VPD) compromises the outcome, and the burden of VPD remains largely unknown in low- and middle-income countries (LMICs). To scale up VPD care in these areas, we continue to explore the use of artificial intelligence (AI) and machine learning (ML) for automatic detection of VPD from speech samples alone.

METHODS

An age-matched, single-institution cohort of 60 patients (30 control, 30 with VPD after cleft palate repair) generated approximately 8000 audio samples (4000 VPD and 4000 control). These samples were used to inform the development of a neural network-based, self-supervised deep learning ML model.

RESULTS

ML model testing with augmented and unaugmented data sets revealed accuracies of 1.0, macro precisions of 1.0, macro recalls of 1.0, and F1 scores of 1.0.

DISCUSSION: Although these results are promising and support the ability of ML models to detect VPD, the results likely indicate that the ML models are also picking up confounding data. Efforts are underway to address this problem while simultaneously employing disentanglement tactics to allow for multilingual speech analysis. The ability to clinically operationalize such a model could instantaneously enhance VPD care in LMICs for patients with cleft palate with little changes to existing healthcare infrastructure.