以監督式學習演算法結合健保資料庫實現 憂鬱及躁鬱患者篩檢預測模式

Using Data Mining Techniques to Establish Prediction Model of Bipolar Disorder for Unipolar Depression Patients

吳俊儀(Chun Yi Wu)

指導教授 ： 胡雅涵

國立中正大學/管理學院/資訊管理學系/碩士(2014年)

摘要

憂鬱症在過去十幾年來為人類失能排名第二的疾病。憂鬱症及躁鬱症在臨床上的早期症狀並無明顯可區別的特徵。在這個領域中雖然尋求可靠的篩檢方式已做了相當多的研究，但是仍然無明確的方向將躁鬱症從憂鬱症篩檢出來。
抗憂鬱劑為憂鬱症患者在治療上常用藥物，因為憂鬱轉為躁鬱之情緒轉變較不易察覺，躁鬱症患者常被錯誤的施以單一抗憂鬱劑治療，反而導致躁鬱症病程加速惡化。
本研究目的為分析憂鬱症病患轉為躁鬱症時可能的共病現象。我們利用WEKA 軟體之決策樹、類神經網路、與邏輯斯迴歸做分析研究。將2003-2010年健保資料庫之診斷為憂鬱症病人共5830筆作為研究資料。
研究結果顯示5830位罹患憂鬱症病人中，有73筆是躁鬱症之病人。我們將只罹患憂鬱症病人5757位以隨機抽樣方式抽樣4%，抽出30組與73位轉躁鬱症患者合併。經決策樹分析實驗結果，其30組平均準確率為73.4%優於類神經網路與羅吉斯迴歸。研究結果顯示轉為躁鬱症之憂鬱症病人中，憂鬱症患者以患有人格障礙病史、藥物成癮、環境障礙病史、酒癮、焦慮症與其他分類之精神官能症等有轉為躁鬱症傾向，其次以病人就醫次數、醫院特約屬性、醫師專科年資對憂鬱症病人轉為躁鬱症具影響性。
總結上述憂鬱症患者篩檢為躁鬱症病患之共病現象，可以做為篩選出潛在的躁鬱症患者潛在症狀。另外精神病患者不健康的就醫行為模式會增加憂鬱症患者最終罹患為躁鬱症之風險，因此篩選躁鬱症病患時也應納入評估。

關鍵字

全民健康保險資料庫 ； 共病現象 ； 決策樹 ； 資料探勘 ； 躁鬱症 ； 憂鬱症

並列摘要

Over the last few decades, unipolar depression has been the second leading cause of disability. Clinical features of unipolar depression and bipolar disorder, do not readily differentiate the two illness trajectories in the early course of illness. Although a lot of research work has been done in this filed to seek for reliable measurement, there is no clear direction to distinguish bipolar disorder from unipolar depression.
Antidepressants are drugs used for the treatment of unipolar depression. Because the mood swings are less obvious from depression to manic episodes, many bipolar disorder patients are often wrongly treated with antidepressants alone. Treating these patients with antidepressants alone can actually increase the manic episodes and worsen the disorder.
Our purpose of this study is to explore the comorbidity symptoms of unipolar depression (UD) patients who are developing into bipolar disorder. The method to carry out this study is using data mining with WEKA decision trees、artificial neural network and logistic regression. The data consisted of 5830 patients with a history of depression from the National Health Insurance Research Database in Taiwan during 2003 to 2010.
The results show that 73 of 5830 patients who are diagnosed with depression actually suffering from bipolar disorder (BD). We extract 30 random sample sets from 5757 UD patients. Each set includes 4 percent of 5757 UD patients and is merged with 73 BD patients. We use 30 sets to run the WEKA classifier and the results show that decision tree is significantly superior to artificial neural network and logistic regression. We get the average accuracy rate of decision tree is 73.4%. We can accurately predict patients who have comorbidity symptoms, such as personality disorders, drug addiction, adjustment disorder, alcohol dependence syndrome, anxiety disorder and neurotic disorders, could have a greater chance of developing into bipolar disorder.
Therefore, the experimental result of this study proves that the comorbidity symptoms described above were beneficial to explore the potential patients who suffering from bipolar disorder. This study also demonstrated that unhealthy patient behaviors were also increased the risk of developing bipolar disorder.

並列關鍵字

Comorbidity symptoms ； Decision trees ； Data mining ； Bipolar disorder ； Unipolar depression ； National Health Insurance Research Database

參考文獻

曾淑芬(88)。從醫院管理角度論全民健保資料庫。中華公共衛生雜誌，18 (5)，363-372。

劉介宇、洪永泰、莊義利、陳怡如、翁文舜、劉季鑫 (民95)。台灣地區鄉鎮市區發展類型應用於大型健康調查抽樣設計之研究。健康管理學刊，4 (1)，1-22。

鄭守夏 (民88)。全民健保學術資料庫簡介。中華公共衛生雜誌，18 (3)，235-236。

Akiskal, H. S., & Benazzi, F. (2005). Atypical depression: A variant of bipolar ii or a bridge between unipolar and bipolar ii? Journal of affective disorders, 84 (2), 209-217.

Akiskal, H. S., & Benazzi, F. (2008). Continuous distribution of atypical depressive symptoms between major depressive and bipolar ii disorders: Dose-response relationship with bipolar family history. Psychopathology, 41 (1), 39-42.