| 摘要: | 一、研究背景與目的
背景：在台灣，末期腎臟病患者選擇血液透析治療佔九成。動靜透析脈廔管（Arteriovenous fistula (AVF) 及Arteriovenous graft(AVG)）即是患者的生命線，若透析廔管忽然無預警阻塞或失效，將是病患與醫護人員最大夢靨。血管通路功能障礙是血液透析患者發病率和死亡率最重要的原因之一，其佔住院人數的三分之一，佔這些患者醫療保健費用的很大一部分。目前臨床上是透過每次血液透析前，以生理檢查評估廔管功能狀況或者當次透析發現廔管無功能才緊急安排皮氣球血管擴張術(percutaneous transluminal angioplasty, PTA)介入治療。有研究發現當透析廔管發生功能障礙時，應立即在 48 小時內進行治療，若超過5天，將增加手術提取困難。一個優秀的監測方法應該是快速、簡單、準確、非侵入性、不依賴於操作者並且具有成本效益。本研究運用機器學習做預測血管功能障礙，以降低因延遲發現、延遲治療導致管路重置風險。
目的：其一提早發現血液透析患者之動靜脈廔管功能不良問題，其二針對影響動靜脈廔管障礙重要因子分析。
二、研究材料與方法
本研究運用機器學習模型作為預測透析廔管功能障礙。我們研究了在臺北醫學大學附設醫院血液透析室之患者。於2020年1月到2021年10月，總共22個月。患者需接受血液透析大於一年且使用動靜脈廔管（AVF或AVG）做為長期透析通路之患者。血管通路功能障礙被定義為需要進行血栓切除術或經皮血管成形術。以曾經做經皮氣球血管擴張術PTA(percutaneous transluminal angioplasty)的病患為陽性患者，而無發生動靜脈廔管功能障礙的患者為陰性患者。以二值變量（binary variable）輸入代碼0 或1。收集病患的基本生物資料、血液檢驗資料及血液透析中透析機所測得的數據資料，包括靜脈壓、過濾率…等。運用XGBoost (極限梯度提升)找出影響廔管功能障礙重要特徵。資料分成訓練集（80% training dataset）及測試集(20% testing dataset)。運用8種機器模型做訓練並經過5折交叉驗證(5-fold cross validation)作為模型評估。
三、分析與結果
我們總共納入了 216 名患者。 ABC三組不同參數進行機器模型預測。發現C組參數最少但不影響預測校能。模型前三名分別是XGBoost、Random Forest及Decision Tree。ROC 曲線下面積 (AUC-ROC) 分別為0.99/0.99/0.95。模型顯示了觀察到的廔管障礙之風險因子。對風險估計影響最大的變量是糖尿病史，其次是血管通路種類和實驗室數值的指標。
四、討論
本研究為預測病人1個月內會發生動靜脈廔管功能不良問題，讓高風險廔管問題早期發現早期治療。此外也針對重要的影響因子在臨床上加以預防。另外提早做好血管通路讓血管較成熟再上針使用以及改善營養狀況都能降低廔管阻塞風險。
Background: In Taiwan, 90% of patients with end-stage renal disease choose hemodialysis. Arteriovenous fistula (AVF) and Arteriovenous graft (AVG) are patients’ lifelines. If the dialysis catheter occludes without warning, it is associated with significant distress for patients and medical staff. Vascular access dysfunction is one of the most important causes of morbidity and mortality in hemodialysis patients, accounting for one-third of hospitalizations and a significant portion of the healthcare costs of these patients. The assessment of AVF/AVG by physical examination before hemodialysis to evaluate the functional status of the canal is a standard in clinical practice. Percutaneous transluminal angioplasty (PTA) interventional therapy is urgently arranged when the AVF/AVG is found to be non-functioning during dialysis. Some studies have found that when the dialysis tube becomes dysfunctional, it should be managed within 48 hours. If it exceeds five days, it will increase the difficulty of surgical intervention. An excellent monitoring method should be fast, simple, accurate, non-invasive, operator-independent, and cost-effective. This study uses machine learning to predict vascular dysfunction and reduce the risk of circuit replacement due to delayed detection and treatment.
Objectives: The first aim is to detect the dysfunction of arteriovenous ducts early in hemodialysis patients, and the second is to analyze the crucial factors affecting the arteriovenous ductal disorders.
Methods: We used machine learning as a prediction model for dialysis catheter dysfunction. This study recruited patients in the hemodialysis center of the Taipei Medical University Hospital from January 2020 to October 2021(a total of 22 months). Patients need to receive hemodialysis for more than one year and use an arteriovenous catheter (AVF or AVG) for long-term dialysis access. Vascular access dysfunction was defined as the need for thrombectomy or percutaneous angioplasty. Patients who had undergone percutaneous transluminal angioplasty (PTA) were regarded as positive cases, and those without arteriovenous duct dysfunction were regarded as negative cases. We then treat the problem as a binary classification between positive and negative data. All the collecting variables included the patient's demographic data, blood test data, and data measured by the dialysis machine in hemodialysis, including venous pressure, filtration rate, etc. The important features that affected duct dysfunction were identified via XGBoost (Extreme Gradient Boosting) machine learning algorithm. The data is divided into a training set (80% of samples) and a testing set (20% of samples). Finally, eight machine learning algorithms were used for model training (via 5-fold cross-validation) and model evaluation.
Results: In total, 216 patients are included in our study. Three sets of parameters (A, B, and C) were used for machine model prediction. Group C was found to have the fewest parameters but did not affect the predicted calibration performance. The top three models are XGBoost, Random Forest, and Decision Tree. The area under the ROC curve (AUC-ROC) was 0.99/0.99/0.95. The model shows the observed risk factors for hemodialysis shunt. The variables that had the most significant impact on risk estimates were the history of diabetes, followed by indicators of vascular access type and laboratory values.
Conclusion: This study aims to predict the occurrence of arteriovenous dysfunction in patients within one month so that high-risk vascular problems can be detected and treated early. In addition, the clinical prevention of important factors is also analyzed. |
