藉由人工智慧深度學習技術辨識透析血管通路阻塞狀況

Identification of Obstruction Status of Dialysis Vascular Access Using the Artificial Intelligence Deep Learning

李尉鈞(Wei-Jun Li)

指導教授 ： 王家鍾

義守大學/醫學院/生物醫學工程學系/碩士(2020年)

摘要

近年來國內洗腎人口比例屢次創新高，以達到9萬人口，發生率是世界第一，且醫療健保對慢性腎臟病之給付為居國內之冠。針對末期腎臟病患者，血液透析為主要選擇的治療方式，因此血管通路順暢與否對於洗腎病人是很重要的。而當透析瘻管發生狹窄與阻塞時，主要是透過經皮血管造型術來加以處理。
本論文主要以非侵入式的方法來擷取手術前後透析瘻管的血流聲音，並以人工智慧(AI)深度學習技術來分類血管通路與否，以期不須經由X光透視攝影就能區分是否有嚴重狹窄情況。本論文所測量之病人共有119例，皆由高雄榮民總醫院心臟血管中心就診案例中篩選，利用已開發之裝置分別擷取術前及術後透析瘻管動脈癒合端(A點)、瘻管中央(B點)及靜脈癒合端(C點)之17.3855秒左右的血流聲音訊號，在取樣頻率為2k Hz之下，各位置共獲得34,771個資料點。手術前後成功取樣的分別有116例及117例。本論文所建立的AI深度學習模型，是以血流聲音訊號作為輸入層的輸入資料，其中案例的70%做為訓練之用及30%作為測試之用，並由輸出層的單一神經元來分類瘻管阻塞與否。
結果顯示，分別以A點、B點及C點的血流聲音訊號作為輸入層資料時，所建立AI深度學習模型之準確度分別達到61%、58%及63%。若將A、B、和C三點之血流聲音訊號合併成一筆資料點（共52.1565秒，104,313資料點），作為深度學習模型輸入層的輸入資料，則準確度達到65%；同時, 本論文以輸入案例數加倍(Double)方式來提高樣本數，則分辨瘻管是否阻塞的準確度達到82%。
由結果歸納得出，僅藉由透析瘻管血流聲音訊號作為輸入資料，本論文所建立的AI深度學習模型就能夠分辨瘻管是否暢通或者阻塞，故將能協助醫事人員簡易、快速診斷出洗腎病患透析瘻管是否可進行洗腎處置或需立即進行瘻管重塑手術，以利延長瘻管的壽命。

關鍵字

人工智慧 ； 血液透析 ； 血管通路 ； 血管音圖 ； 血流聲音 ； 瘻管 ； 深度學習

並列摘要

As the number of patients undergoing hemodialysis increases, the same increment has also been observed in the incidence of hemodialysis access dysfunction. Whether it is an arteriovenous fistula or an artificial fistula, obstruction may occur after a period of use. Therefore, the purpose of this thesis is to record the phonoangiogram (PAG) signals of the vascular access and to find out its patency situation.
The study included 119 patients who were treated with routine hemodialysis. All of them were screened from the Cardiac Vascular Center of Kaohsiung Veterans General Hospital. The thesis aimed to use non-invasive methods to record blood flow sound (PAG signal) from the fistula before and after surgery, and to establish an artificial intelligence (AI) deep learning model to classify if the vascular access had stenosis without X-ray applications.
The blood flow sound signals of about 17.3855 seconds were measured at the fistula artery healing end (position A), the fistula center (position B) and the vein healing end (position C), respectively. A total of 34,771 data points was obtained with a sampling frequency of 2k Hz. Also, 116 cases and 117 cases were successfully sampled before and after the surgery operation. The AI deep learning model in this thesis used the blood flow sound signals as the input data of the input layer. 70% of all cases were used for training and 30% for testing of the proposed model. One single neuron of the output layer was used to classify whether the fistula was blocked or not.
The results showed that the accuracies of the AI deep learning model reached 61%, 58%, and 63% when the blood flow sound signals at positions A, B, and C were used as the input data, respectively. If the blood flow sound signals recorded from the positions A, B, and C were merged into one data file (52.165 seconds, 104,313 data points) as the input data of the input layer of the model, an accuracy of 65% was found. Furthermore, when the number of samples was doubled, the accuracy of distinguishing whether the fistula is blocked reached 82%.
In summary, the AI deep learning model established in this thesis may help medical personnel diagnose the patency of the vascular access in patients with renal dialysis.

並列關鍵字

Artificial intelligence ； Hemodialysis ； Vascular access ； Phonoangiogram ； Blood flow sound ； Fistula ； Deep learning

參考文獻

[1] L. R. Rabiner and R. W. Schafer, Digital Processing of speech Recognition Signal, Prentice Hall (1978).

[2] H. A. Mansy, S. J. Hoxie, N. H. Patel, and R. H. Sandler, “Computerised analysis of auscultatory sounds associated with vascular patency of haemodialysis access,” Med Biol Eng Comput, 43(1):56-62, 2005.

[3] K. Kian and A. Asif, “Status of research in vascular access for dialysis,” Nephrol Dial Transplant, 25(11):3682-6, 2010.

[4] C. Lomonte, F. Casucci, M. Antonelli, B. Giammaria, N. Losurdo, G. Marchio, and C. Basile, “Is there a place for duplex screening of the brachial artery in the maturation of arteriovenous fistulas?,” Semin Dial, 18(3):243-6, 2005.

[5] M. Surlan and P. Popovic, “The role of interventional radiology in management of patients with end-stage renal disease,” Eur J Radiol, 46(2):96-114, 2003.