应用机器学习实现听性脑干反应波形自动识别  

Machine Learning in Automatic Recognition of Auditory Brainstem Response Waveforms

作　　者：梁思超 许嘉 叶佐昌[2] 刘海旭 梁仁和 郭振平 卢曼林 高娟娟 伊海金 LIANG Sichao;XU Jia;YE Zuochang;LIU Haixu;LIANG Renhe;GUO Zhenping;LU Manlin;GAO JuanJuan;YI Haijin(Department of Otolaryngology-Head and Neck Surgery,Beijing Tsinghua Changgeng Hospital,Beijing 102218,China;不详)

机构地区：[1]北京清华长庚医院耳鼻咽喉头颈外科,北京102218 [2]清华大学集成电路学院 [3]北京精仪天和智能装备有限公司

出　　处：《中华耳科学杂志》2025年第1期59-64,共6页Chinese Journal of Otology

摘　　要：目的训练多种机器学习模型用于听性脑干反应(auditory brainstem response,ABR)波形的自动识别,并确定准确率最高的模型,使ABR自动识别技术更好地应用于临床实践。方法选取2021年6月至2022年6月北京清华长庚医院收治的100例听力正常和伴有听力损伤人群的受试者(200耳)为研究对象,根据年龄和听力水平将受试者分为组1(年龄18~59岁,500、1000、2000、4000 Hz频率平均听阈≤25 dB HL)、组2(年龄≥60岁,500、1000、2000、4000 Hz频率平均听阈≤25 dB HL)、组3(年龄18~59岁,500、1000、2000、4000 Hz频率平均听阈>25 dB HL)、组4(年龄≥60岁,500、1000、2000、4000 Hz频率平均听阈>25 dB HL),每组25例。收集受试者纯音测听和ABR数据,提取ABR信号时域和频域特征,与受试者年龄、性别、纯音听阈,刺激声强度以及原始信号序列拼接得到特征向量。分别使用逻辑回归、支持向量机分类、伯努利朴素贝叶斯分类、高斯朴素贝叶斯分类、高斯过程分类、决策树、随机森林、表格网络、轻量化梯度提升框架、极致梯度提升框架和局部级联集成。等机器学习模型对ABR波形进行识别训练,并对整体数据和分组数据分别计算不同模型下波形识别的准确率。结果高斯过程分类模型的整体准确率达到了94.89%,超过了其他机器学习模型。其中95.62%为<60岁听力正常受试者、92.19%为≥60岁听力正常受试者、92.92%为<60岁伴有听力损失受试者、92.50%为≥60岁且伴有听力损失受试者。结论机器学习技术在ABR波形的自动识别方面具有良好的应用前景,高斯过程分类模型优于其他机器学习模型。Objective The study aims to train various machine learning models for automated recognition of auditory brainstem response(ABR)waveforms to identify the model with the highest accuracy,thereby facilitating the application of automated ABR recognition technology in clinical practice.Methods The study included 100 participants(200 ears)recruited from Beijing Tsinghua Changgeng Hospital between June 2021 and June 2022,including individuals with normal hearing and those with hearing impairment.Pure-tone audiometry and ABR data were collected,and participants were divided into four groups based on age and hearing levels:i.e.18~59 years of age with normal hearing(average hearing threshold at 500,1000,2000,4000 Hz≤25 dB HL)(Group 1),≥60 years with normal hearing(Group 2),18~59 years with abnormal hearing(average hearing threshold>25 dB HL)(Group 3),and≥60 years with abnormal hearing(Group 4),with 25 subjects in each group.Time-domain and frequency-domain features of ABR signals were extracted and combined with participant demographics,pure-tone threshold,stimulus intensity and raw signal sequence to form comprehensive feature vectors.Eleven machine learning models were used for ABR waveform recognition,including logistic regression,support vector classification,Bernoulli Naive Bayes(BNB),Gaussian Naive Baye,Gaussian process classification,decision tree,random forest,tabular network,light gradient boosting machine,extreme gradient boosting,and local cascade ensemble.The recognition accuracy of each model was assessed for overall data and each group data.Results The Gaussian process classification model demonstrated the highest overall recognition accuracy(94.89%),which outperformed all other models.The subgroup accuracies were as follows:95.62%for normal hearing participants under 60 years old,92.19%for normal hearing participants aged 60 and above,92.92%for hearing-impaired participants under 60 years,and 92.50%for healing-impaired participants aged 60 and above.Conclusion Machine learning models demonstrates subs

关 键 词：听觉脑干反应 波形识别 机器学习 高斯过程分类模型 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] R318[自动化与计算机技术—控制科学与工程]