Clinical Report: Enhanced Learning of Nasality Representation for Screening VPD
Overview
This study presents a two-stage framework to enhance the robustness of machine learning models for screening velopharyngeal dysfunction (VPD) in uncontrolled environments. The proposed method demonstrates performance in both in-domain and out-of-domain settings.
Background
Velopharyngeal dysfunction (VPD) significantly impacts speech intelligibility and communication, particularly in patients with cleft palate. Traditional screening methods require specialized expertise and controlled environments, which are often unavailable in low- and middle-income countries. This study addresses the need for VPD screening tools that can operate effectively in everyday acoustic conditions.
Data Highlights
Condition
Macro-F1
Accuracy
In-domain (best performance)
1.000
1.000
Out-of-domain (proposed method)
0.679
0.695
Out-of-domain (MFCC baseline)
0.612
0.641
Key Findings
The proposed method achieved ceiling performance in in-domain recordings (macro-F1 = 1.000, accuracy = 1.000).
In a nested 5-fold cross-validation, the best mean performance was with SVM (macro-F1 = 0.981 ± 0.022, accuracy = 0.985 ± 0.016).
In out-of-domain recordings, the proposed method outperformed the strongest baseline (MFCC) by +0.067 macro-F1 and +0.054 accuracy.
Machine learning models often degrade in performance when applied to uncontrolled environments due to domain shift.
Learning a nasality-focused representation can improve robustness against recording artifacts.
Clinical Implications
This study presents findings on the adaptation of machine learning models for VPD screening in real-world settings.
Conclusion
The study demonstrates that enhancing nasality representation through supervised learning can improve the performance of VPD screening tools in diverse acoustic environments.
by Weixin Liu, Bowen Qu, Amy Stone, Maria Powell, Shama Dufresne, Stephane Braun, Izabela Galdyn, Michael Golinko, Bradley Malin, Zhijun Yin, Matthew E. Pontell
