基于人工智能的GRBASZero与临床医生对土耳其语发音障碍者嗓音的感知评估及多参数嗓音测量之比较

文献信息

中文摘要

目的： 本研究旨在探讨临床医生基于听觉感知的GRBAS评级与人工智能驱动的GRBASZero应用程序在发音障碍个体中生成的输出之间的关系，并考察两种方法与客观声学、空气动力学及患者报告结局测量（PROMs）的相关性。
方法： 研究纳入60名发音障碍个体（排除健康对照以防止分布性地板效应）。录制持续元音/a/、/i/、/s/、/z/发声及朗读段落样本，并使用Praat软件分析空气动力学参数（最长发声时间和s/z比值）以及声学参数。由13名言语语言病理学家（SLPs）使用GRBAS量表对录音进行感知评估，显示出极佳的评分者间信度（ICC值范围为0.93至0.97；Koo and Li, 2016），并与GRBASZero应用程序生成的评分进行比较。此外，参与者完成了土耳其语改编的患者报告结局测量，包括嗓音障碍指数-10（VHI-10）、嗓音相关生活质量（VRQOL）、嗓音疲劳指数（VFI）和声道不适量表（VTDS）。
结果： 临床医生与GRBASZero评估在参数G（总嘶哑度）、R（粗糙度）和B（气息声）之间发现显著的中等相关性，而在参数A（无力感）和S（紧张度）之间观察到弱相关性。两种评估方法均与客观声学参数表现出理论上一致的关系；然而，与临床医生不同，人工智能与空气动力学测量无显著相关性。此外，虽然临床评估与患者报告结局测量（PROMs）显示显著相关性，但人工智能评估很少与这些自我感知量表相关，强化了听觉感知判断与患者感知生活影响之间的概念区分。
结论： 基于人工智能的感知评估与临床医生评级表现出中至弱的 agreement。这些发现表明，像GRBASZero这样的工具有潜力作为临床嗓音评估中的支持性辅助工具，而非绝对的客观替代品。此外，人工智能输出与患者报告结局测量之间有限的相关性在理论上是预期的，反映了基于声学的感知判断与患者感知症状负担之间的概念区分。同样，人工智能无法与空气动力学测量相关，突显了其在捕捉声学信号之外潜在生理缺陷方面的局限性。

英文摘要

OBJECTIVES: This study aimed to investigate the relationship between clinician-based auditory-perceptual GRBAS ratings and the outputs generated by the AI-powered GRBASZero application in individuals with dysphonia and to examine how both methods correlate with objective acoustic, aerodynamic, and patient-reported outcome measures (PROMs).
METHODS: The study was conducted on 60 dysphonic individuals (following the exclusion of healthy controls to prevent distributional floor effects). Sustained /a/, /i/, /s/, /z/ phonations and reading passage samples were recorded, and aerodynamic parameters (maximum phonation time and s/z ratio), along with acoustic parameters, were analyzed using Praat software. Perceptual evaluations of the recordings were performed by 13 Speech and Language Pathologists (SLPs) using the GRBAS scale, exhibiting excellent inter-rater reliability (ICC values ranging from 0.93 to 0.97; Koo and Li, 2016), and were compared with the scores generated by the GRBASZero application. Additionally, participants completed Turkish-adapted patient-reported outcome measures, including the Voice Handicap Index-10 (VHI-10), Voice-Related Quality of Life (VRQOL), Voice Fatigue Index (VFI), and Vocal Tract Discomfort Scale (VTDS).
RESULTS: Significant moderate correlations were found between clinician and GRBASZero assessments for the parameters G (Grade), R (Roughness), and B (Breathiness), while weak correlations were observed for the parameters A (Asthenia) and S (Strain). Both evaluation methods demonstrated theoretically consistent relationships with objective acoustic parameters; however, unlike clinicians, the AI showed no significant correlation with aerodynamic measures. Furthermore, while clinical assessments showed significant correlations with patient-reported outcome measures (PROMs), AI assessments rarely correlated with these self-perceptual scales, reinforcing the conceptual distinction between auditory-perceptual judgments and patient-perceived life impact.
CONCLUSION: AI-based perceptual assessments demonstrate moderate-to-weak agreement with clinician ratings. These findings suggest that tools like GRBASZero have the potential to serve as supportive adjuncts rather than absolute objective replacements in clinical voice evaluation. Furthermore, the limited correlation between AI outputs and patient-reported outcome measures is theoretically expected, reflecting the conceptual distinction between acoustic-based perceptual judgments and patient-perceived symptom burden. Similarly, the AI's inability to correlate with aerodynamic measures highlights its limitation in capturing underlying physiological deficits beyond the acoustic signal.