人工智能在鼾声分析中的应用:OSA检测与阻塞部位分类的系统综述
该系统性综述评估机器学习与人工智能模型分析鼾声以检测阻塞性睡眠呼吸暂停(OSA)及基于VOTE系统分类阻塞部位的应用。检索2011年至2024年11月30日五个数据库,纳入42项研究,采用PROBAST评估偏倚风险。ML模型在鼾声分类和OSA检测中准确率高达98.6%,VOTE分类的UAR为49%至75%,使用靶向特征后可达87.5%至94.7%或准确率92.2%至95.5%。
要点
- ML模型如SVM、CNN和混合深度学习架构在鼾声分类和OSA检测中准确率高达98.6%。
- VOTE分类基线UAR为49%至75%,靶向特征联合SVM或kNN可达87.5%至94.7% UAR。
- 挑战包括外部验证有限、数据集不平衡和噪声干扰。
核心结论
- AI驱动的鼾声分析在无创OSA筛查和解剖部位分类方面具有潜力。
- 未来工作应关注多模态数据整合、真实世界验证和模型泛化性以支持临床转化。
研究设计与样本
- 系统性综述,遵循PRISMA 2020指南和PICO框架,检索PubMed、Scopus、Ovid、Cochrane和Web of Science数据库2011年至2024年11月30日的文献。
- 共识别591条记录,经筛选和全文审查后纳入42项研究,使用PROBAST工具评估方法学质量。
主要发现
- 支持向量机、卷积神经网络和混合深度学习架构等ML模型在鼾声分类和OSA检测中达到高准确率(高达98.6%)。
- 关键特征包括梅尔频率倒谱系数、小波变换和深度频谱表示。
- 多项研究尝试VOTE分类,基线流程报告未加权平均召回率(UAR)为49%至75%,使用靶向特征联合SVM或k近邻时UAR达87.5%至94.7%或准确率92.2%至95.5%。
- 挑战包括外部验证有限、数据集不平衡和噪声干扰。
关键数值与推荐
- ML模型在鼾声分类和OSA检测中准确率高达98.6%。
- VOTE分类基线UAR为49%至75%,靶向特征联合SVM或kNN时UAR为87.5%至94.7%或准确率92.2%至95.5%。
临床怎么用
- AI驱动的鼾声分析可作为无创OSA筛查和阻塞部位分类的辅助工具。
- 临床应用中需注意外部验证有限、数据集不平衡和噪声干扰等挑战。
局限性
- 外部验证有限。
- 数据集不平衡。
- 噪声干扰。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: The aim of the systematic review is to evaluate the application of machine learning (ML) and artificial intelligence (AI) models in the analysis of snoring sounds for the detection of obstructive sleep apnea (OSA) and the classification of obstruction sites based on the Velum, Oropharynx, Tongue, and Epiglottis (VOTE) system. DATA SOURCES: We conducted a comprehensive literature search in PubMed, Scopus, Ovid, Cochrane, and Web of Science databases from 2011 to November 30, 2024. The search included terms related to snoring, machine learning, artificial intelligence, and acoustic analysis. REVIEW METHODS: The review was structured using the PICO framework and adhered to Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) 2020 guidelines. A total of 591 records were identified, and 42 studies were included after screening and full-text review. Methodological quality was assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST) tool, and data extraction focused on ML models, data sets, feature extraction methods, and performance metrics. RESULTS: ML models such as support vector machine (SVMs), convolutional neural networks (CNNs), and hybrid deep learning architectures achieved high accuracy (up to 98.6%) in snore classification and OSA detection. Key features included mel-frequency cepstral coefficients (MFCCs), wavelet transforms, and deep spectral representations. Several studies attempted VOTE classification, reporting unweighted average recalls (UARs) of 49% to 75% with baseline pipelines, and up to 87.5% to 94.7% UAR or 92.2 to 95.5% accuracy when using targeted features with SVM or k-nearest neighbor (kNN). Challenges included limited external validation, data set imbalance, and noise interference. CONCLUSION: AI-driven snore analysis holds potential for noninvasive OSA screening and anatomical site classification. Future work should focus on multimodal data integration, real-world validation, and model generalizability to support clinical translation.
引用
- Tartaglia FC, Pace GM, Giombi F, et al. Artificial Intelligence in Snoring Sound Analysis: OSA Detection and Obstruction Site Classification, a Systematic Review. Otolaryngol Head Neck Surg. 2026. PMID: 41521859. doi:10.1002/ohn.70114
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。