基于机器学习及音频技术监测动静脉内瘘功能的可行性分析  

Feasibility of arteriovenous fistula function monitoring based on machine learning and audio technology

作　　者：王凡立 杨艳丽 徐元恺 阮琳 李文 刘永亮 孙利军[1] 雷营 刘小明 赵培楠 张丽红[1] WANG Fan-li;YANG Yan-li;XU Yuan-kai;RUAN Lin;LI Wen;LIU Yong-liang;SUN LI-jun;LEI Ying;LIU Xiao-ming;ZHAO Pei-nan;ZHANG Li-hong(Department of Nephrology,First Hospital of Hebei Medical University,Shijiazhuang 050000,China;Department of Nephrology,Zhejiang Hospital Affiliated to Zhejiang University School Of Medicine,Hangzhou 310030,China;Department of Nephrology,Tianjin Third Central Hospital,Tianjin 300170,China)

机构地区：[1]河北医科大学第一医院肾内科,石家庄050000 [2]浙江大学医学院附属浙江医院肾内科,杭州310030 [3]天津市第三中心医院肾内科,天津300170

出　　处：《中国血液净化》2024年第9期701-705,共5页Chinese Journal of Blood Purification

基　　金：河北省卫生健康创新专项(22377794D)。

摘　　要：目的探讨采用基于机器学习及音频技术分析自体动静脉内瘘(arteriovenous fistula,AVF)的听诊数据从而实现监测AVF功能的可行性。方法选择河北医科大学第一医院肾内科收治的AVF狭窄且需要行经皮腔内血管成形术(percutaneous transluminal angioplasty,PTA)治疗的患者共50名,分别在PTA前和PTA后使用无线电子听诊器记录AVF的听诊数据,利用音频技术将音频文件转换为频谱图并提取出声学特征,然后利用MATLAB软件中的30余种分类学习器对提取出的声学特征进行分析构建出评估AVF功能的模型,比较这些模型的性能。结果将从50名患者中获得的100个音频文件纳入研究。在PTA前后AVF的频谱图对比中,PTA前的频谱图显示出更大的高频振幅(Z=-4.721,P<0.001)。PTA前最高频与最低频的差值更大(Z=-6.169,P<0.001)。在利用30余种分类学习器建立的模型中高效线性支持向量机(support vector machine,SVM)、粗略高斯SVM建立的模型性能最好(AUC=0.892、0.882),5折交叉验证后准确度达81.11%;而二次判别、三次k近邻法(k-nearest neighbor,kNN)、中神经网络、双层神经网络构建的模型测试准确度最高可达90%。结论基于频谱的机器学习模型可以预测AVF的显著性狭窄,因此将其用于AVF功能的监测是可行的。且高效线性SVM、粗略高斯SVM构建的模型,在本次可行性研究中性能最好。Objective To explore the feasibility of using machine learning and audio technology and analyzing the auscultation data from arteriovenous fistula(AVF)to monitor AVF function.Methods A total of 50 patients with AVF stenosis which required percutaneous transluminal angioplasty(PTA)were recruited in this study.Digital sound of the AVF shunt was recorded using a wireless electronic stethoscope before(pre-PTA)and after PTA(post-PTA).The audio files were subsequently converted to spectrograms,and the acoustic signatures were extracted.More than 30 classification learners in MATLAB software were used to analyze the extracted acoustic features,and models to evaluate the AVF function were constructed.Finally,the performance of these models was compared.Results One hundred audio files were obtained from the 50 recruited patients and were pooled for the study.In the spectrogram comparison of AVF before PTA and after PTA,the spectrogram before PTA showed a larger high-frequency amplitude.There was a significant difference between the files before and after PTA(Z=-4.721,P<0.001).The highest and lowest frequencies in a cardiac cycle were significant different before and after PTA,and the difference was greater before PTA than after PTA(Z=-6.169,P<0.001).Among the models established by more than 30 kinds of classification learners,the model established by efficient linear support vector machine(SVM)and coarse Gaussian SVM had the best performance,with 81.11%accuracy after 5-fold cross validation.The test accuracy of the model constructed by quadratic discrimination,cubic kNN,medium neural network,and double-layer neural network can reach up to 90%.Conclusions Spectrum-based machine learning models can predict the saliency narrowness of AVF,so it is feasible to use it for the monitoring of AVF function.The model constructed by efficient linear SVM and rough Gaussian SVM had the best performance in this feasibility study.

关 键 词：动静脉内瘘 血管成形术 机器学习 音频 人工智能 

分 类 号：R318.16[医药卫生—生物医学工程]