以人工智慧預測第二代抗精神病藥物之療效及副作用 - 一種個別化精神藥理學的新方法

Artificial Intelligence Prediction of Response and Adverse Effects of Second-Generation Antipsychotics - A New Approach for Individualized Psychopharmacology

林朝誠(Chao-Cheng Lin)

指導教授 ： 邱泓文

共同指導教授 ： 李友專

臺北醫學大學/醫學院/醫學科學研究所/博士(2009年)

https://doi.org/10.6831/TMU.2009.00055

摘要

精神分裂症是一種極為複雜的精神疾病，曾被稱為影響人類最嚴重的疾病，非典型或第二代抗精神病藥物代表治療精神分裂症的進步，其中Clozapine是目前對難治型精神分裂症效果最好的藥物，然而過去之研究報告顯示Clozapine只對大約三分之一到二分之一的難治型精神分裂症患者有效，且Clozapine產生的副作用很多，且有產生致命的agranulocytosis之風險，然而目前沒有一個明確指引可以幫助醫師決定那個病人會對那個抗精神病藥物有效，造成醫師用藥的困難。

第二代抗精神病藥物於臨床上逐漸被廣泛使用的時候，其引起的副作用中，體重增加及代謝症候群因機率高達三分之一左右而逐漸受到注意。然而最近的多數研究仍指出，大多數的病人其代謝症候群仍然未被偵測出來，因此如何發展出預測服用第二代抗精神病藥物的病人是否產生療效或嚴重副作用具有臨床的重要性。

由於精神分裂症為一多因子之疾病，受生物、心理、環境之綜合影響，以致於目前難以用單一或少數幾個因子來預測抗精神病藥物之藥效或副作用，因此針對單一因子的臨床或基因研究的結果並無法應用於個案臨床治療的選擇。利用多基因或多臨床變項的統計分析雖然也能用來預測，然而多基因及多臨床變項間的交互作用常牽涉複雜的非線性關係，因此需要使用較複雜的非線性方法來處理分析資料。人工智慧，如類神經網路，具有從現有資料中學習並預測未知資料的能力，類神經網路可運用非線性數學模式來模擬人腦解決問題的過程，因而能處理一般演算法難以處理的問題，也許對於我們目前要預測複雜的精神分裂症的治療反應或副作用是一種新而可行的工具，然而目前文獻上沒有人以類神經網路來預測分析第二代抗精神病藥物之療效及副作用。

本研究論文的主要目標為：(1) 以臨床及基因變項預測clozapine之療效；(2)識別服用第二代抗精神病藥物之精神分裂症病人是否產生代謝症候群之副作用。本研究建立類神經網路模型與迴歸模型，並測試其預測臨床療效及副作用之準確性，並比較人工智慧與迴歸模型之鑑別力何者較佳，我們假設人工智慧模型之預測準確度可能會相當或高於迴歸模型。

結果研究I包括服用clozapine的93名精神分裂症病人，結果發現最佳的類神經網路模型為標準的feed-forward, fully-connected, back-propagation, multilayer perceptron模型，其預測clozapine藥效的準確度為83.3% (AUC=0.821). Sensitivity 及specificity 分別為100% and 76.5%，以area under the receiver operator characteristic curve來比較模型之鑑別率，發現類神經網路模型優於迴歸模型(0.821±0.054 vs. 0.579±0.068; p<0.001) ，僅含基因變項之類神經網路優於僅含臨床變項之類神經網路模型(0.805±0.056 vs. 0.647±0.066; p=0.046)，可見基因多型性在預測clozapine藥效上具有重要角色。

研究II包括400位服用第二代抗精神病藥物超過六個月的精神分裂症病人，類神經網路模型及迴歸模型之輸入變項僅包括人口學及人體測量學資料，結果顯示類神經網路模型及迴歸模型均具有高度的準確率(86.6% vs. 83.6%)、sensitivity (89.6% vs. 87.9%)、及specificity (85.9% vs. 82.2%) 可辨別代謝症候群，用內部效度資料測得的類神經網路模型及迴歸模型之area under the receiver operator characteristic curve分別為0.936±0.030及0.921±0.033 (p=0.49)，用外部效度資料檢測類神經網路模型及迴歸模型結果仍然很好(0.908 ± 0.041 vs. 0.899 ± 0.043, p=0.739)。

雖然抗精神病藥物作用極為複雜， 但用人工智慧來預測第二代抗精神病藥物之療效及副作用似乎可行，本研究結果將可以在學術上提供一種新的預測模式，讓醫師在未來能依病患其個別體質及需求來決定如何選用第二代抗精神病藥物。

關鍵字

人工智慧 ； 抗精神病藥物 ； 預測 ； 反應 ； 副作用

並列摘要

Schizophrenia is a complex mental illness and has been described as the “worst disease affecting mankind”. Atypical or second-generation antipsychotic drugs (SGAs) represent an important advance in the treatment of schizophrenia. Among them, clozapine is regarded as the most effective antipsychotic for treating refractory schizophrenia (Kane et al. 1988; Fleishman 1999). However, clozapine has clinical response rates of only 30–50% in treatment-refractory schizophrenia patients and several adverse effects are often complained of. Moreover, the use of this antipsychotic carries significant morbidity from seizure and serious blood disorders such as potentially fatal agranulocytosis. However, there is no guideline to assist psychiatrists to decide which patient would response to an SGA.

As the SGA use has been increased, it has been recognized that SGAs are associated with a higher risk of excessive weight gain and metabolic syndrome. However, most patients with antipsychotic-related metabolic syndrome were not detected in recent reports. Therefore, it is important to develop a method to predict efficacy and severe side effects for patients treated with SGA.

Although some factors were found to be associated with antipsychotic response and adverse effects, no single factor can predict the efficacy or side effects. Only a few genetic studies tried to predict clozapine response with combinations of many genetic polymorphisms. The epistasis or multiple gene-gene interaction may play a role in drug response. However, the combined analysis of multiple gene polymorphisms and clinical variables may require the use of novel nonlinear methods other than traditional methods of analysis. Artificial intelligence, like artificial neural network (ANN), can learn from existing data and extend its prediction to unknown data. ANN uses nonlinear mathematical models to mimic the human brain’s own problem-solving process. Therefore, it may be a new and possible tool to predict the complex response and side effects of schizophrenia treatment. However, as we started our studies, there was no published paper regarding the prediction of efficacy and SGA side effects by ANN.

The objectives of this study are (1) to predict therapeutic response to clozapine with clinical and genetic factors; (2) to identify metabolic syndrome in schizophrenic patients taking second-generation antipsychotics. We established artificial intelligence and regression models for the predictions and compare the accuracy and discriminatory power between the models. We conducted 2 studies and hypothesized that the performance of artificial neural network models would be as high as or higher than those with logistic regression.

Study I included 93 schizophrenic patients taking clozapine. The results showed that the best artificial neural network model was a standard feed-forward, fully-connected, back-propagation, multilayer perceptron model. The overall accuracy for clozapine response was 83.3%. The sensitivity and specificity of the prediction are 100% and 76.5% respectively. The resultant area under receiver operating characteristic curve is 0.821. By using the area under the receiver operating characteristics curve as a measure of performance, the artificial neural network model outperformed the logistic regression model (0.821±0.054 vs. 0.579±0.068; p<0.001). The artificial neural network with only genetic variables outperformed that with only clinical variables (0.805±0.056 vs. 0.647±0.066; p=0.046). The gene polymorphisms should play an important role in the prediction.

Study II included 400 patients with second-generation antipsychotic treatment for more than 6 months. The input variables of artificial neural network and logistic regression were limited to demographic and anthropometric data only. The results demonstrated that both the final artificial neural network and Logistic regression models had high accuracy (86.6% vs. 83.6%), sensitivity (89.6% vs. 87.9%), and specificity (85.9% vs. 82.2%) to identify metabolic syndrome in the internal validation set. The areas under the receiver operator characteristic curve were high for both the artificial neural network and logistic regression models (0.936±0.030 vs. 0.921±0.033, p=0.49). During external validation, high AUC was still obtained for both models (0.908 ± 0.041 vs. 0.899 ± 0.043, p=0.739).

Although the mechanism of antipsychotic reaction is complex, it is possible to predict the efficacy and SGA side effects by artificial intelligence. Our results establish a new model that makes the choice of antipsychotics be tailored to individual needs in the future.

並列關鍵字

artificial intelligence ； antipsychotic ； predict ； response ； side effect

參考文獻

1. Allison DB, Mackell JA, McDonnell DD (2003) The impact of weight gain on quality of life among persons with schizophrenia. Psychiatr. Serv. 54, 565-567.

2. Amendolia SR, Brunetti A, Carta P et al. (2002) A real-time classification system of thalassemic pathologies based on artificial neural networks. Med. Decis. Making 22, 18-26.

3. American Psychiatric Association (1994) Diagnostic and Statistical Manual of Mental Disorders. 4th ed., Washington (DC).

4. Arana E, Marti-Bonmati L, Paredes R, Bautista D (1998) Focal calvarial bone lesions. Comparison of logistic regression and neural network models. Invest Radiol. 33, 738-745.

5. Arranz M, Collier D, Sodhi M et al. (1995) Association between clozapine response and allelic variation in 5-HT2A receptor gene. Lancet 346, 281-282.