Currently, routine imaging examinations for dialysis access include ultrasound and X-ray angiography, which not only require expensive medical imaging equipment but also demand a high level of clinical experience from medical staff. This paper proposes a time-frequency domain feature analysis method for audio signals based on a self-developed portable phonoangiogram (PAG) acquisition device, which simulates the auscultation process of professional doctors to intelligently detect dialysis access stenosis. The method was tested on a dataset of 62 patients, achieving accuracy, sensitivity, specificity, and F1 score of 0.905, 0.911, 0.902, and 0.850, respectively. The results indicate that the algorithm's overall performance surpasses existing algorithms, providing robust support for clinical detection of dialysis access stenosis.
在时域内,本文计算了声形图分段信号的峰值(Peak,Pk)、均方根值(Root mean square,RMS)和短时能量(Short-time energy,ST-E)三个统计值作为时域特征。峰值这种声音强度的增加通常与血液流速的增加或血管狭窄的程度有关[25],峰值表现了该段信号的最大声音强度。在处理声形图信号时,均方根值可以提供关于信号能量或强度的信息。例如,较高的均方根值表明该段时间血液流经测量部位的声音强度变化较大,可能有狭窄发生。短时能量是一种音频信号处理中常用的特征提取方法[26],可以用来检测音频信号的能量变化。与时域统计特征不同的是,计算短时能量需要分帧。本文中帧长为20ms,帧移为8 ms,使用汉明窗截断。第一步得到的声形图分段信号将在这里被分成6帧并逐帧计算能量,最后取该段信号的能量均值作为特征。峰值、均方根值、短时能量的公式表示为:
式中:为t时刻的输入向量;ht-1为t-1时刻的隐藏状态; ft 、 it 和 ot 分别为遗忘门、输入门和输出门的激活向量;为当前时刻生成的候选单元状态;Ct-1和Ct 分别为t-1时刻和t时刻的单元状态;ht 为t时刻的隐藏状态输出; Wf 、 Wi 、 Wc 和 Wo 分别为遗忘门、输入门、候选单元状态和输出门对应的权重矩阵; bf 、 bi 、 bc 和 bo 分别为对应的偏置向量;σ为sigmoid激活函数。
DaugirdasJ T, BlakeP G, IngT S. Handbook of Dialysis[M]. New York: Lippincott Williams & Wilkins, 2015.
[3]
ChenT K, KnicelyD H, GramsM E. Chronic kidney disease diagnosis and management: a review[J]. Jama, 2019, 322(13): 1294-1304.
[4]
RonnebergerO, FischerP, BroxT. U-net: convolutional networks for biomedical image segmentation[C]∥Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, 2015: 234-241.
[5]
LokC E, HuberT S, LeeT, et al. KDOQI clinical practice guideline for vascular access: 2019 update[J]. American Journal of Kidney Diseases, 2020, 75(4): 1-164.
[6]
SchmidliJ, WidmerM K, BasileC, et al. Editor's choice-vascular access: 2018 clinical practice guidelines of the European society for vascular surgery (ESVS)[J]. European Journal of Vascular and Endovascular Surgery, 2018, 55: 757-818.
[7]
QuencerK B, AriciM. Arteriovenous fistulas and their characteristic sites of stenosis[J]. American Journal of Roentgenology, 2015, 205: 726-734.
[8]
Roy-ChaudhuryP, SukhatmeV P, CheungA K. Hemodialysis vascular access dysfunction: a cellular and molecular viewpoint[J]. Journal of the American Society of Nephrology, 2006, 17: 1112-1127.
[9]
LeeT, Roy-ChaudhuryP. Advances and new frontiers in the pathophysiology of venous neointimal hyperplasia and dialysis access stenosis[J]. Advances in Chronic Kidney Disease, 2009, 16: 329-338.
[10]
AbreoK, BuffingtonM, SachdevaB. Angioplasty to promote arteriovenous fistula maturation and maintenance[J]. Journal of Vascular Access, 2018, 19: 337-340.
[11]
KatsanosK, KarnabatidisD, KitrouP, et al. Paclitaxel-coated balloon angioplasty vs. plain balloon dilation for the treatment of failing dialysis access: 6-month interim results from a prospective randomized controlled trial[J]. Journal of Endovascular Therapy, 2012, 19: 263-272.
[12]
LokC E, RajanD K, ClementJ, et al. Endovascular proximal forearm arteriovenous fistula for hemodialysis access: results of the prospective, multicenter novel endovascular access trial (NEAT)[J]. American Journal of Kidney Diseases, 2017, 70: 486-497.
[13]
LeesR S, Dewey JrC F. Phonoangiography: a new noninvasive diagnostic method for studying arterial disease[J]. Proceedings of the National Academy of Sciences, 1970, 67(2): 935-942.
[14]
SeoJ H, MittalR. A coupled flow-acoustic computational study of bruits from a modeled stenosed artery[J]. Medical & Biological Engineering & Computing, 2012, 50: 1025-1035.
[15]
ChenC H, TaoT H, ChouY H, et al. Arteriovenous fistula flow dysfunction surveillance: early detection using pulse radar sensor and machine learning classification[J]. Biosensors, 2021, 11(9): 11090297.
[16]
GrochowinaM, LeniowskaL, Gala-BłądzińskaA. The prototype device for non-invasive diagnosis of arteriovenous fistula condition using machine learning methods[J]. Scientific Reports, 2020, 10(1): 16387.
[17]
WangH Y, WuC H, ChenC Y, et al. Novel noninvasive approach for detecting arteriovenous fistula stenosis[J]. IEEE Transactions on Biomedical Engineering, 2014, 61(6): 1851-1857.
[18]
ZhouG, ChenY, ChienC, et al. Deep learning analysis of blood flow sounds to detect arteriovenous fistula stenosis[J]. NPJ Digital Medicine, 2023, 6(1): 37658233.
[19]
OtaK, NishiuraY, IshiharaS, et al. Evaluation of hemodialysis arteriovenous bruit by deep learning[J]. Sensors, 2020, 20(17): 20174852.
[20]
MøllerH, PedersenC S. Hearing at low and infrasonic frequencies[J]. Noise & Health, 2004, 6(23): 37-57.
[21]
MünzelT, HeitzerT, HarrisonD G. Physiologie und pathophysiologie des stickstoffmonoxid-superoxid-systems[J]. Herz, 1997, 22: 158-172.
[22]
JiangC, YangD, ChenX. Similarity learning for cover song identification using cross-similarity matrices of multi-level deep sequences[C]∥ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing, Barcelona, Spain, 2020: 26-30.
[23]
PandaB, MandalS, MajerusS J A. Flexible, skin coupled microphone array for point of care vascular access monitoring[J]. IEEE Transactions on Biomedical Circuits and Systems, 2019, 13(6): 1494-1505.
[24]
LeesR S, Dewey JrC F. Phonoangiography: a new noninvasive diagnostic method for studying arterial disease[J]. Proceedings of the National Academy of Sciences, 1970, 67(2): 935-942.
[25]
ZubairS, YanF, WangW. Dictionary learning based sparse coefficients for audio classification with max and average pooling[J]. Digital Signal Processing, 2013, 23(3): 960-970.
[26]
BenetosE, StylianouY. Auditory spectrum-based pitched instrument onset detection[J]. IEEE Transactions on Audio, Speech, and Language Processing, 2010, 18(8): 1968-1977.
[27]
TzanetakisG, CookP. Musical genre classification of audio signals[J]. IEEE Transactions on Speech and Audio Processing, 2002, 10(5): 293-302.
[28]
HeK, ZhangX, RenS, et al. Deep residual learning for image recognition[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 770-778.
[29]
GravesA, GravesA. Long short-term memory[J]. Supervised Sequence Labelling with Recurrent Neural Networks, 2012, 1: 37-45.
[30]
XuX, ZhaoH, LiuH, et al. LSTM-Gan-xgboost based anomaly detection algorithm for time series data[C]∥11th International Conference on Prognostics and System Health Management (PHM-2020 Jinan), Jinan, China, 2020: 334-339.
[31]
GéronA. Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow[M]. Beijing: O'Reilly Media, 2019.
[32]
GoodfellowI, BengioY, CourvilleA. Deep Learning[M]. Cambridge: MIT Press, 2016.