以類神經網路預測躁鬱症患者使用鋰鹽的藥物反應

Prediction of Lithium Response in Bipolar Disorder Patients using Artificial Neural Network

蔡宛玲(Wan-Ling Tsai)

指導教授 ： 郭柏秀

國立臺灣大學/公共衛生學院/流行病學與預防醫學研究所/碩士(2013年)

https://doi.org/10.6342/NTU.2013.00868

摘要

研究背景:
躁鬱症(雙極性情感疾患)是一種致病機轉複雜的精神疾病，而目前認為藥物的治療對於躁鬱症的預後是一個重要的因素。鋰鹽是一種情緒穩定劑，目前被視為治療躁鬱症的第一線藥物，它可以有效的預防躁鬱症的復發以及降低自殺的風險性。儘管鋰鹽只能對於部分病人有效果，約有三分之二的病人服用鋰鹽後是沒有治療效果的。在過去的文獻中，預測鋰鹽效果的變項，常常會受限於各式各樣的測量方法，測量後的結果無法相互比較。在傳統的邏輯式迴歸只考慮到變項間的線性關係，但在現實生活中，變項間有可能存在非線性的關係。類神經網路可以處理臨床預測變項與結果的非線性關係。所以本篇研究的目的是要建立一個應用在台灣族群的服用鋰鹽反應的預測模型以及相關的預測變項。
材料與方法:
我們的病人從台北五家醫院收集，並經過精神疾病診斷分類準則(DSM-IV)判斷為躁鬱症。總共有117名有服用鋰鹽藥物的病人，包含了54位男性以及63位女性。我們使用Alda準則去測量鋰鹽的反應效果。病人將會依據Alad切分等於2分成有鋰鹽反應(≧2)以及沒有鋰鹽反應(<2)。病人資訊的取得是藉由問卷以及病歷翻閱取得，包含了臨床變項(例如:發病年齡，快速周期，家族精神病史)，社會變項(神經質人格)以及人口學變項(例如:社經地位和職業)。雖然資料是藉由病歷和問卷取得，依然有遺失值，我們以K nearest neighbor(KNN)來替補遺失值。我們使用類神經網路和邏輯式迴歸比較鋰鹽反應的預測準確度。另一方面，我們使用類神經網路中的連結權重方法和逐步邏輯式迴歸來挑選重要的預測變項。並使用正確分類比率、敏感度、精確度以及AUC來比較不同模型預測的準確性。
結果:
在我們的研究中，不管是在放入全部的變項、類神經網路連結權重挑出的9個變項和逐步邏輯式挑出的變項的模型，類神經網路的分類準確性比邏輯式迴歸佳。以放入全部的變項得到，邏輯式迴歸和類神經網路的AUC分別是0.56和0.85。敏感度分別是0.40和0.94，精確度是0.65和0.70。另外在變項選擇上，邏輯式迴歸利用逐步回歸法，挑出了菸草的使用和混合發作次數為預測變項，而類神經網路最後的模型選擇了9個預測變項，分別是快速週期,菸草使用，酒精及藥物的使用，性別，先鬱後躁的發作類型(DMI)，自殺，神經質人格。最後我們利用類神經網路的連結權重方法挑出的9個變項作為最後外部預測的預測模型型應用在4個規律服用鋰鹽的病人身上，準確率高達100%，皆能正確分類有無鋰鹽反應。
結論:
在本研究中類神經網路與邏輯式迴歸相比較後，類神經網路因為可以處理變項間非線性的關係而提供較準確的預測效果，一個好的預測模型對於臨床的使用是很重要的，它可以給予病人正確的用藥處方。未來的研究更需要這樣的預測模型幫助躁鬱症病人找到屬於自己的個人用藥處方，以減少不必要的藥物測試。

關鍵字

躁鬱症 ； 類神經網路 ； 邏輯式迴歸 ； 鋰鹽 ； Alda 準則

並列摘要

Bipolar affective disorder (BAD) is a severe psychiatric disorder and pharmacologic maintenance treatment is crucial factors for patients' prognosis. Lithium is a mood-stabilizer, which is considered as a first-line drug for treating BAD. It has also been proven to be effective in long-term prevention of recurrence and reduce the risk of suicidal behavior. Although lithium works very well in some patients, about two- thirds of lithium-treated patients are non-responders. In the past, the searching for relevant predictors for lithium response is hindered by varying measurement of lithium response, and the results are not easily compared across studies. The goal of the present study is to build prediction models for lithium response in Taiwanese samples and select important input variables for prediction. Traditional logistic regression models consider only linear effect between dependent and independent variables, which is not always held in some cases. On the contrary, an artificial neural network (ANN) method can deal with non-linear relationship between response outcome and clinical and psychosocial predictors. Thus, we compared the prediction accuracy using both logistic regression and ANN models..
BAD patients who met the diagnostic criteria of DSM-IV were recruited from five hospitals in Taipei. There were 117 patients who took lithium regularly, including 54 males and 63 females (mean age 45±12 years). We used Alda scale to assess lithium response. Patients were grouped into responders, whose score >= 2 and non-responders (<2). Information regarding the predictors was acquired from semi-structured interview and chart review, including clinical features (e.g. age onset, rapid cycling, family history and so on ), psychosocial features (e.g. neuroticism) and demographic (e.g. social status and employment) variables. Prediction models were constructed using ANN and logistic regression for lithium response. We used stepwise logistic regression and connection weight of ANN to select important variables. We also compared the prediction accuracy across different models using sensitivity, specificity and area under the curve (AUC).
We found that the ANN model yielded a much higher level of prediction rate (0.85) for lithium response than did it with multiple logistic regression model (0.56). The sensitivity in logistic regression and ANN model were 0.40 and 0.94, and specificity were 0.65 and 0.70, respectively. The results of variable selection via ANN and logistic regression are not all consistent. In the logistic regression analysis, number of mixed episodes and tobacco use were selected as predictors. On the other hand, rapid cycling, substance problems of tobacco, alcohol and drug, sex, depression-mania-interval, number of depressive episodes, suicide and neuroticism were important variables in the ANN model.
In conclusion, a timely prediction for lithium response is special clinical importance to assist guiding patients’ treatment strategy. The use of ANN model provides accurate prediction and proves to be an attractive solution for this daily problems. Information regarding to tobacco use especially plays an important roles in predictive accuracy. Further research is needed to bring us closer to“achieve personalized medicine”for the use of lithium in BAD patients.

並列關鍵字

bipolar disorder ； lithium ； artificial neural network ； logistic regression ； Alda scale

參考文獻

Abou-Saleh, M. & Coppen, A. 1986. Who responds to prophylactic lithium? Journal of affective disorders, 10, 115-125.

Baldessarini, R. J. & Tondo, L. 2000. Does lithium treatment still work? Evidence of stable responses over three decades. Arch Gen Psychiatry, 57, 187-90.

Baldessarini, R. J., Tondo, L., Floris, G. & Hennen, J. 2000. Effects of rapid cycling on response to lithium maintenance treatment in 360 bipolar I and II disorder patients. Journal of affective disorders, 61, 13-22.

Bauer, M. S., Calabrese, J. R., Dunner, D. L. & Post, R. 1994. Multisite data reanalysis of the validity of rapid cycling as a course modifier for bipolar disorder in DSM-IV. The American journal of psychiatry.

Coryell, W., Akiskal, H., Leon, A. C., Turvey, C., Solomon, D. & Endicott, J. 2000. Family history and symptom levels during treatment for bipolar I affective disorder. Biological psychiatry, 47, 1034-1042.