應用類神經網路於 Guillain-Barré 症候群預後之預測

Applying Artificial Neural Network on the Prognosis of Guillain-Barré Syndrome

謝恩綺(En-Chi Hsieh)

指導教授 ： 邱泓文

臺北醫學大學/醫學科技學院/醫學資訊研究所/碩士(2015年)

https://doi.org/10.6831/TMU.2015.00029

摘要

背景:Guillain-Barré 症候群是一個罕見的疾病,發病前期會出現上呼 吸道感染或腸胃炎的症狀,發病後會出現複視、四肢麻木無力。確診 Guillain-Barré 症候群,進行血漿換置術或靜脈注射免疫球蛋白,將近 80%病患於就診治療後完全恢復,10%-20%病患會留下永久性神經後 遺症,5%-10%病患會惡化衰竭死亡。

方法:研究應用邏輯斯回歸與類神經網路,建構 Guillain-Barré 症候 群預後之預測模型,並使用卡方檢定與 T 檢定分析預測因素與預後恢 復之相關性。

結果:研究樣本中,72 位 Guillain-Barré 症候群病患,共有 60 位預後 表現良好,預後恢復率達八成。預測因素分析結果,使用呼吸器輔助 與下肢 MRC 肌力評分為影響預後恢復的顯著性因素。應用邏輯斯回 歸方法建立預測模型,所有可能回歸法建立預測模型,預測準確率達 0.944,逐步選取回歸法建立預測模型,預測準確率達 0.93。研究樣 本經過調整,應用類神經網路方法建立預測模型,挑選測試組預測表現 100%的所有特徵值模組,預測準確率達 1,挑選測試組預測表現 96%的顯著特徵值模組,預測準確率達 0.858。研究後期將呼吸器輔 助之干擾變項移除,重新配適預測模型,其預測表現明顯下降。

結論:綜合比較四組模型,在選入預測因子之建模方法上,均呈現選 入所有預測因子之模型預測表現,優於選入顯著預測因子之模型預測 表現,在配適訓練預測模型之分析方法上,又以類神經網路分析方法, 建構之所有特徵值模組之預測表現最好。

關鍵字

Guillain-Barré症候群 ； 邏輯斯回歸 ； 類神經網路

並列摘要

Background: Guillain-Barré syndrome is a rare disorder. The disease symptoms include upper respiratory tract infection and gastroenteritis. The physical symptoms occur double vision and limb numbness. Diagnosed with Guillain-Barré syndrome, patients conduct treatment of plasma exchange or intravenous immunoglobulin. Nearly 80% of patients recover fully, 10% -20% of patients have permanent neurological sequelae, 5% -10% patients will failure to die.

Method: The study applies logistic regression and artificial neural network to construct predicted models for the prognosis of Guillain-Barré syndrome. The study uses Chi-square test and T test to analysis the relation between factors and the prognosis of Guillain-Barré syndrome.

Result: The study totally includes 72 patients with Guillain-Barré syndrome, and 60 patients recover well. The rate of prognosis for recovery is 80%. Analysis for the predicted factors, patients used mechanical ventilation and patients’ MRC scores of lower limbs are significant factors for the prognosis of Guillain-Barré syndrome. Applying logistic regression to build predicted models, the model of all possible regression procedure that the predicted accuracy is 0.944 and the model of stepwise regression procedure is 0.93. Then, adjusted samples apply artificial neural network to build predicted models. Selecting all eigenvalues model with 100% test set performance, the predicted accuracy is 1. Selecting significant eigenvalues model with 96% test set performance, the predicted accuracy is 0.858. The late of research removes the interferon variable of patients used mechanical ventilation. The predicted performance of rebuild models is obviously decreased.

Conclusion: Comprehensive comparison of four predicted models, the study shows results. The structured method to choose all predicted factors that model predicted performance is better than the structured method to choose significant predicted factors. Applying the analysis method of artificial neural network, the model with all eigenvalues that predicted accuracy and model performance are the best in the study.

並列關鍵字

Guillain-Barrésyndrome ； Logistic regression ； Artificial neural network

參考文獻

[22] B.-C. Cheng, W.-N. Chang, C.-S. Chang, C.-Y. Chee, C.-R. Huang, J.-B. Chen,C.-J. Chang, P.-L. Hung, K.-W. Wang, H.-W. Changeand C.-H. Lu. Guillain–Barre´syndrome in southern Taiwan: clinical features,prognostic factors and therapeutic outcomes. European Journal of Neurology 2003; 10: 655–662.

[1] Yuki, Nobuhiro. Hartung, Hans-Peter. Guillain–Barré Syndrome. New England Journal of Medicine 2012;366; 2294–2304.

[3] van Doorn PA. Kuitwaard K. Walgaard C et al. IVIG treatment and prognosis in Guillain-Barré syndrome. Journal of Clinical Immunology 2010; 30 ;Suppl 1; S74–78.

[4] Wan-Ting Huang, Jen-Hsiang Chuang, Steve Hsu-Sung Kuo. Monitoring the safety of pandemic H1N1 vaccine. The Lancet 2010; 375; 1164.

[5] James J. Sejvar Andrew L. Baughman Matthew Wise Oliver W. Morgan. Population Incidence of Guillain-Barré Syndrome: A Systematic Review and Meta-Analysis. Neuroepidemiology 2011;36:123–133.