我國現今洗腎盛行率和每年新增加洗腎病患比率,均居世界雙料冠軍。透析瘻管可說是血液透析患者的生命線,不論是自體瘻管或是人工瘻管,經使用一段時間後,均可能發生阻塞的現象。本研究改良一套非侵入性擷取透析瘻管血流聲音的裝置,將所測得一段時間之血流音訊號,利用時域分析,將三個位置的原血流聲音訊號,以3秒為一單位截取訊號,並將每筆3秒之血流音訊號及其一階和二階微分訊號合併成一個影像圖檔,經篩選後共獲得術前1891(阻塞)及術後1930(暢通)樣本數。本研究利用Python撰寫所需程式,並使用人工智慧深度殘差網路(ResNet50、ResNet101)來分類是否為阻塞的或暢通的血流聲音圖檔。本文最終採用ResNet50之模型進行深度學習訓練,根據結果顯示,此模型於第29 Epoch 時,分類準確率高達95.6%;而損失值也僅剩下0.13。利用非侵入的方式量測不同瘻管部位的血流聲音,間接評估病患之人工或自體瘻管是否已有阻塞的現象,不僅可以降低檢測人員進行物理檢查方式時,因觸感、經驗等因素造成程度上之誤判,也以利提醒患者在尚未完全阻塞時,便到醫院進行血管檢查,延長瘻管的使用壽命,及減少健保支出。
Recently, the prevalence of dialysis in Taiwan ranked the world highest. The dialysis fistula can be seen as the lifeline of hemodialysis patients. However, a stenotic phenomenon may occur in the fistula after a period of use, usually a couple of months. Therefore, the study aimed to develop a device for measuring the blood flow sound generated by the fistula or graft. Then, the sound signals were analyzed in time domain and re-captured to yield 3-second intervals of the sound signals. Also, we designed a software program using “Python” language to perform the first-order and the second-order differentiation of the blood flow sound signals generated by the fistula. Then, we created image files, containing the 3-second sound signal, and its first and second derivatives, as the input of the artificial intelligence (AI) model. After removing low-quality images, we collected 1891 images before the surgery and 1930 images after the surgery. Furthermore, this study employed two deep residual networks (ResNet50 & ResNet101) to classify whether the vascular access is stenotic or unblocked. The results showed that the ResNet50 model exhibited good learning performance with classification accuracy of 95.6% and lose value of 0.13 at the 29-training epoch of the learning process. In summary, the proposed AI method demonstrates its potential to be a useful diagnosis way for early and non-invasive evaluation of the stenosis situation in the hemodialysis patients. Also, the outcome of the study will help to reduce the degree of misjudgment by touch perception and physical examination.