数据驱动的噪声中言语表现预测：机器学习模型的系统综述与荟萃分析。

文献信息

中文摘要

目的： 噪声中言语理解困难（SIN）是沟通能力和生活质量的主要决定因素，影响听力损失患者以及部分听力正常者。数据驱动的预测方法为建模言语感知背后的复杂多维因素提供了有前景的途径。本系统综述与荟萃分析评估了预测SIN相关结局的计算方法，在定量合成可行时比较了不同模型类别的预测性能，并识别了与模型选择、泛化性和临床转化相关的方法学因素。
方法： 遵循PRISMA 2020指南，我们对使用机器学习（ML）、神经网络、自动语音识别（ASR）或基于听觉模型的方法来预测、模拟或评估SIN相关结局的研究进行了系统综述与荟萃分析。对报告支持向量机（SVM）或随机森林（RF）性能的研究，采用随机效应模型进行定量合成，并评估了研究间异质性。
结果： 符合条件的研究包括传统ML、神经网络、基于ASR以及机制性听觉建模方法。SVM准确率合并为0.71，SVM曲线下面积合并为0.84；RF准确率合并为0.86，RF召回率合并为0.81。Bootstrap估计与主要随机效应估计大体一致，尽管各分析中研究间异质性从中等到极高不等。
结论： 数据驱动的预测方法在预测SIN相关结局方面显示出前景，但高异质性阻碍了确定单一最优模型。模型选择应反映预测目标、数据类型、样本量、可解释性和预期用途。当前证据对概念验证性能的支持强于对临床效用的支持，凸显了开展更大规模、标准化、外部验证研究的必要性。

英文摘要

PURPOSE: Difficulty understanding speech in noise (SIN) is a major determinant of communication ability and quality of life, affecting individuals with hearing loss as well as some with normal hearing. Data-driven prediction methods offer a promising approach for modeling complex, multidimensional factors underlying speech perception. This systematic review and meta-analysis evaluated computational approaches for predicting SIN-related outcomes, compared predictive performance across model classes where quantitative synthesis was feasible, and identified methodological factors relevant to model selection, generalizability, and clinical translation.
METHODS: Following PRISMA 2020 guidelines, we conducted a systematic review and meta-analysis of studies using machine learning (ML), neural networks, automatic speech recognition (ASR), or auditory model-based approaches to predict, simulate, or evaluate SIN-related outcomes. Quantitative synthesis was performed using random-effects models for studies reporting support vector machine (SVM) or random forest (RF) performance, and between-study heterogeneity was assessed.
RESULTS: Eligible studies included traditional ML, neural network, ASR-based, and mechanistic auditory modeling approaches. SVM accuracy pooled at 0.71 and SVM area under the curve at 0.84, while RF accuracy pooled at 0.86 and RF recall at 0.81. Bootstrap estimates were broadly consistent with the primary random-effects estimates, although between-study heterogeneity ranged from moderate to very high across analyses.
CONCLUSION: Data-driven prediction approaches show promise for predicting SIN-related outcomes, but high heterogeneity precludes identifying a single optimal model. Model selection should reflect the prediction target, data type, sample size, interpretability, and intended use. Current evidence supports proof-of-concept performance more strongly than clinical utility, highlighting the need for larger, standardized, externally validated studies.