應用資料探勘技術探討遵循診治指引預測慢性阻塞性肺病之預後

Applying Data Mining Techniques To Exploring The Guideline Adherence, And Predicting The Prognosis Of Chronic Obstructive Pulmonary Disease

許鎂(Chia-Mei Hsu)

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(2012年)

摘要

研究背景：慢性阻塞性肺病(COPD)是目前全世界死亡率持續明顯上升之主要疾病。世界衛生組織估計到2020年，COPD將成為全球第三大死因。GOLD自2001年開始制定COPD診治指引，作為全球臨床醫師診治COPD之準則；但在很多國家，包括美國及臺灣，COPD的死亡率及醫療費用仍呈現上升之趨勢。

研究目的：了解臺灣醫師遵循COPD診治指引之狀況，及COPD患者發生共存疾病之情況；進而應用資料探勘技術，建立最佳存歿預測模式、找出存歿特徵，探討共存疾病及診治指引遵循度對COPD患者預後存歿之關聯規則。

研究對象與方法：本研究採用國家衛生研究院全民健保資料庫40萬人承保抽樣歸人檔，研究對象為40歲以上，於2003年以COPD國際疾病分類碼前三碼為491、492、496就醫之個案，收集其2003~2005年之就醫資料，進行描述性分析，並使用Weka軟體進行資料探勘分類技術、關聯分析及群集分析。

研究結果：研究對象總計5,211人，男性3,298人(63.3％)，女性1,913人(36.7％)，平均年齡67歲；達88.2％有發生共存疾病，前四項依序為：高血壓(66％)，心臟血管疾病(57.6％)，肺炎(49.6％)，第二型糖尿病(30.9％)；個案六年死亡率為24.8％，死因包括肺部疾病(28.05％)，癌症(21.33％)，心臟血管疾病(12.44％) ，腦血管疾病(11.44％)及其他(26.74％)。臺灣臨床醫師具有較佳的COPD診治指引遵循項目為：預防性抗生素的使用(95.8％)、常規性全身性類固醇的使用(88.8％)、每次就醫給予支氣管擴張劑使用(80.1％)；遵循度較差的項目為戒菸衛教、肺部復健訓練、肺功能檢查。以資料探勘決策樹、多層感知器、羅吉斯迴歸，三種演算法建立分類預測存歿，研究結果顯示三者之正確率介於89.2~89.5，AUC介於0.9~0.91，三者均具有良好的預測力。關聯分析結果發現，當個案肺功能檢查遵循度為中度(三年內檢查1至2次)，且每年都有接種流感疫苗，則其存活機率高(正確率0.87)。群集分析存歿族群特徵差異發現，死亡族群特徵為：年齡大於78歲、發生共存疾病數≧2、有發生心臟血管疾病、有發生肺炎；存活特徵為：年齡小於58歲，且無發生共存疾病、無發生心臟血管疾病、無發生肺炎。

研究結論：臺灣地區確實忽視了COPD此全球唯一仍快速成長的致死疾病。研究個案達70.9％於三年內完全無接受肺功能檢查；疾病的評估與監測並未落實執行，也可能因此對疾病的控制及治療並不充足，進而影響共存疾病的發生及其預後存歿。建議臺灣衛生主管機關及醫療界都應重視慢性阻塞性肺病COPD。

關鍵字

慢性阻塞性肺病 ； 資料探勘 ； 診治指引 ； 共存疾病

並列摘要

Background：Chronic Obstructive Pulmonary Disease is the only major disease that continues to increase in both prevalence and mortality. The WHO forecasts that COPD will become the 3rd leading cause of mortality worldwide by 2020. Since 2001 the GOLD has been developing guidelines on COPD diagnosis and treatment for the clinicians in worldwide. However, in many countries including the United States and Taiwan, the mortality of COPD and cost is still showing an upward trend.

Objective：In this study, we seek to understand Taiwanese clinicians’ implementation for diagnosis and treatment guidelines of COPD patients and the occurrences of COPD comorbidities. The goal is to apply data mining techniques to find the best prediction model for COPD mortality, while also identifying mortality characteristics, COPD comorbidity, and guideline adherence’s association rules with survival prognosis of COPD patients.

Materials and methods：The study was conducted in reference to the database of the National Health Research Institute .The recruit patients are 40 years of age and over in 2003 and were diagnosed with COPD (ICD-9 codes 491, 492, and 496). Patient information had been collected for three years since 2003, shown by descriptive analysis, and used with the Weka software for data mining classification techniques, association analysis and cluster analysis.

Results：A total of 5,211 patients, 3,298 (63.3％) males and 1,913 (36.7％) females, and averaged in 67 years of age, were included in the study. Of the 5,211 patients, 88.2％ were found with comorbidity. The top four diseases were: Hypertension (66％), CVD (57.6％), Pneumonia (49.6％), Type 2 Diabetes (30.9％). Patient mortality is at 24.8％ with the four leading causes : pulmonary diseases (28.05％), cancer (21.33％), CVD (12.44％), cerebrovascular disease (11.44％). The better complied procedure for COPD treatment by clinicians in Taiwan are: preventive antibiotics (95.8％), systemic corticosteroids (88.8％), and bronchodilator (80.1％). The lesser complied items are education for smoking cessation, pulmonary rehabilitation, and spirometry examination. The study also compared the survival prediction results calculated with Decision Tree, Multilayer Perceptron, and Logistic Regression methods. The study shows that there are no major differences in prediction accuracies among the three algorithms, with accuracies between 89.2~89.5％, and AUC between 0.9~0.91. Further association analysis discovers that when the patient participates in 1 to 2 spirometry examinations in a three year period, while being inoculated with influenza vaccine on an annual basis, the patient will have a high chance of survival (accuracy 0.87). The characteristics shown by the cluster analysis of fatal cases are: Age over 78 years, 2 or more comorbidities, occurrence of CVD, occurrence of pneumonia. Survival is characterized by: younger than 58 years, and without the occurrence of comorbidity, CVD, and pneumonia.

Conclusions：Taiwan had been remained to ignore the only fast-growing lethal disease. Of the patients studied, 70.9％ haven’t had a spirometry examination over a period of three years. The assessment and monitoring of COPD has not been implemented sufficiently, therefore the current measures may not be sufficient enough to control and treat disease, thereby affecting the incidence of comorbidities and survival prognosis. We recommend that Taiwan's health authorities and the medical field to afford more emphasis and attention on this fast-growing disease.

並列關鍵字

COPD ； data mining ； guideline ； comorbidity

參考文獻

6.蘇經凱(2009)。慢性阻塞性肺病臨床診療指引遵從度與療效及醫療資源耗用的關係。高雄醫學大學健康科學院醫務管理學研究所碩士在職專班論文。

1.Anthonisen, N. R., Skeans, M. A., Wise, R. A., Manfreda, J., Kanner, R. E., & Connett, J. E. (2005). The effects of a smoking cessation intervention on 14.5-year mortality: a randomized clinical trial. Ann Intern Med, 142(4), 233-239.

3.Barbarito, N., Vaghi, A., & De Mattia, E. (2011). Prevalence of airflow obstruction according GOLD, ATS and ERS criteria in symptomatic ever-smokers referring to a pulmonary rehabilitation department. Monaldi archives for chest disease, 75(3), 157-161.

4.Best, N., & Hansell, A. L. (2009). Geographic Variations in Risk. Epidemiology, 20(3), 400-410.

5.Borsini, E., Broggini, G., Fasana, A., Baldassarri, C., Manzo, A. M., & Perboni, A. D. (2011). Access to pyrrolo-pyridines by gold-catalyzed hydroarylation of pyrroles tethered to terminal alkynes. Beilstein journal of organic chemistry, 7, 1468-1474.