應用資料探勘技術於慢性腎臟病存活之研究

Applying Data Mining Techniques to the Prediction of Survival of Chronic Kidney Disease

李佩勳(Pei-Hsun Lee)

指導教授 ： 阮金聲

國立中正大學/管理學院/資訊管理學系/碩士(2014年)

摘要

根據衛生福利部2012年全國十大死因統計資料，腎臟疾病死亡率為每十萬人12.1人，排行第十位，相較於2011年略減，但國人仍不容忽視(衛生福利部國民健康署，民102a)。本研究收集南部某區域教學醫院2005年至2013年『初期慢性腎臟病醫療給付改善方案及Pre-ESRD預防性計畫』之196位慢性腎臟病患基本資料、檢驗數據資料、用藥資料、住院資料以及衛教資料，應用資料探勘建立預測模式影響慢性腎臟病患死亡之重要危險因子。盼能藉由醫療院所本身所累積的資料建立屬於醫療院所的預測模式，自慢性腎臟病患開始被診斷出相關症狀後，在接受腎臟內科專科醫師至少三個月專業治療後所產生的改變，尋找影響慢性腎臟病患死亡與存活之特徵差異，以相關變項建立最適合之預測模式，並試圖發覺慢性腎臟病照護計劃早期的危險因子且有助於改善病患3年後的狀態，延緩末期腎臟病的發生，提高慢性腎臟病患的存活，對臨床醫學上產生貢獻並降低全民健康保險醫療資源耗用。
研究結果發現，所建構模式以類神經網路預測慢性腎臟病存活最適合，其治療前病患資料平均正確率為87.24% (ROC曲線為89.6%)，其次為決策樹平均正確率為77.55% ( ROC曲線為81.3%)，最後為邏輯斯迴歸平均正確率為69.39% ( ROC曲線為74.3%)；而經過至少三個月治療後病患資料類神經網路平均正確率為88.27% (ROC曲線為92.7%)，其次為邏輯斯迴歸平均正確率為85.71% ( ROC曲線為84.5%)，最後為決策樹平均正確率為73.47% ( ROC曲線為80.8%)。因類神經網路為完全封閉的黑箱結構，較無法解釋，且無法了解運作過程；而決策樹分類器分析結果則較符合實際狀況且分類規則易懂，可作為臨床醫護人員對病患診療及預後參考之依據。

關鍵字

資料探勘 ； 末期腎臟病 ； 慢性腎臟病 ； 存活分析 ； 透析治療

並列摘要

According to the data from Health Promotion Administration, Ministry of Health and Welfare, kidney disease ranked the 10th of the leading causes of death in 2012, and the mortality rate was 12.1 per 100,000 person-years. Although the mortality rate was slightly less than that in 2011, the importance of care for chronic kidney disease(CKD) still can’t be ignored. This study was aimed to use the data, 2005-2013, which included biochemical data, medicine data, inpatient data, and outpatient care education from "Early-CKD improvement program & Pre-ESRD preventive Program"of a regional hospital in Southern Taiwan, to predict the important risk factors leading to dialysis or death in CKD patients. With CKD care program of the insitution, patient would be followed in outpatient department at least every 3 months. A forecast model was supposed to be built by analyzing the data of the first 3-month follow up by data-mining. We tried to find out the early factors in CKD care program, and those contributed the status of patients 3 years later. The Futher aim was to see if these early factors were reversible and to remind doctors and nurses to change health-care strategies improving the survival of CKD patients and delay of dialysis. Finally, it contributed to the clinical medicine, and reduced to the burden of National Health Insurance.
In this research, the best model established by data mining for predicting the survival of CKD is neural network. Before treatment, its accuracy and ROC curves are 87.24% and 89.6%, respectively. The second and last model are decision trees and logistic regression, whose accuracies are 77.55% and 69.39%, and ROC curves are 81.3% and 74.3%. After therapy, its accuracy and ROC curves are 88.27% and 92.7%. The second and last model are logistic regression and decision trees, whose accuracies are 85.71% and 73.47%, and ROC curves are 84.5% and 80.8%. Because A neural network is perceived as being a black box, it is extremely difficult to document how specific classification decisions are reached. However, the results of decision tree are more suitable with practical situation and models are easy to understand for diagnosis and prognosis of reference for Healthcare workers.

並列關鍵字

Data mining ； End-stage renal disease( ESRD) ； Chronic Kidney Disease (CKD) ； Survival Analysis ； Dialysis

參考文獻

王守玠、楊得政 (民96)。慢性腎臟病患者常見之心血管疾病。腎臟與透析，19(2)，89-94。

王舒民、楊雅斐、黃秋錦 (民96)。慢性腎臟病與高血壓。腎臟與透析，19(2)，64-70。

林明彥、黃尚志 (民96)。臺灣慢性腎臓病/末期腎臓病流行病學過去、現在與未來。腎臟與透析，19(1)，1-5。

楊芝青 (民96)。慢性腎臟病危險因子的探討。北市醫學雜誌，4(9_S)，880-889。

郭宜瑾、洪啟智、陳鴻鈞 (民102)。慢性腎臟病貧血診斷與治療的新進展。內科學誌，24(4)，278-287。