| 摘要: | 肌張力不全障礙(Dystonia)是一種自發性的運動障礙疾病，它泛指一群肌肉不自主的過度收縮，由於拮抗肌持續或斷斷續續地不正常收縮導致拮抗肌與收縮肌同時進行收縮的現象。其基本特徵是運動所需的肌肉過度活動。在臨床方面，由於後天性肌張力不全障礙患者的主要致病機轉尚未被完全了解，因此在診斷上只能依賴醫師的經驗來判斷。因此本研究的目的希望能夠藉由機器學習方法，建置一個客觀的寫字型手部痙攣症電腦輔助診斷系統。
本研究收集的對象為寫字型手部痙攣症(Writer’s Cramp)患者12名與正常受試者14名，進行15分鐘內自行數8秒手腕伸展(Wrist extension)的運動，並收集此運動時的腦波訊號(Electroencephalography, EEG)。利用動態因果模型(Dynamic Causal Modeling,DCM)分析患者與正常人運動神經網路連結，並以監督機器學習的方法，將病人與正常人的運動神經網路連結進行分析並找出差異，用以建立電腦輔助診斷系統。
研究結果顯示，在運動開始之前，病人在五個腦區(LM1、LPM、SMA、RM1、RPM)上的Theta頻帶以及三個腦區(LPM、RM1、RPM)上的Alpha頻帶的強度皆明顯小於正常人。而在運動開始之後，患者在Theta頻帶上四個腦區(LM1、LPM、RM1、RPM)) 強度顯大於正常人。另外在Beta頻帶上兩個腦區(RM1、RPM)強度明顯小於正常人。在大腦功能連結的部分，我們的研究結果發現寫病人與正常人之間共有3條連結在統計上有顯著差異，其中輔助運動皮質區有2條異常的連結為較重要的腦區。在輔助診斷系統的部分，我們討論了Theta頻帶改變對診斷系統的影響，我們將Theta頻帶從4-8Hz改為Low Frequency(4-12Hz)，可以看到在4-8Hz的情況下，分類方法為SVM及NaiveBayes演算法便可得到較高的準確率，分別為89.4%及89.8%。若將頻帶改為Low Frequency(4-12Hz)，準確率較高的演算法為SVM、Logistic regression、NaiveBayesn三種，準確率分別為90.99%、90.42%、90.55%。在頻帶組合方面，可以看到在準確率較高的演算法中，Theta頻帶皆包含在最佳組合中，表示在輔助診斷系統中，Theta頻帶被認為是較為重要頻帶。此外，根據我們輔助診斷系統所分析的重要特徵連結也發現，與患側前運動皮質區及輔助運動皮質區被認為是重要腦區，與我們統計出來的結果大致相符。未來希望我們所收的受測者人數可以更多，以避免產生Over Fitting的現象，進而影響輔助診斷的準確率。未來也可以增加更多不同的演算法來找出哪個演算法更適合我們的研究。
;Dystonia is a spontaneous neurological movement disorder, which refers to a group of muscles with involuntary excessive contraction. Its basic feature is excessive muscle activity required for exercise. In clinical terms, the main pathogenesis of patients with acquired dystonia is not fully understood, so the diagnosis can only rely on the experience of the physician to diagnose. Therefore, the purpose of this research is to build an objective computer diagnosis system for task-specific focal hand dystonia by machine learning.
In this study, 12 patients with Writer′s Cramp and 14 normal subjects were collected. They performed a wrist extension of 8 seconds for 15 minutes, and collect electroencephalography (EEG) signals at this time. Use dynamic causal modeling (DCM) to analyze the connection between patients and normal people′s motor neural network, and use the method of supervised machine learning to analyze the connection between patients and normal people′s motor neural network and find out the differences, establish a computer diagnosis system.
The results of the study showed that before the start of exercise, the power of the Theta band on the five brain regions (LM1, LPM, SMA, RM1, RPM) and the Alpha band on the three brain regions (LPM, RM1, RPM) were significantly less than normal peoele. After the start of exercise, the power of the Theta band on the four brain regions (LM1, LPM, RM1, RPM) and the Beta band on the two brain regions (RM1, RPM) were significantly greater than normal people. In the part of brain functional connections, our research results found that there are 3 abnormal connections in patients, of which Two of them are connection with reduced excitement , and one is a connection with a changed connection mechanism. The strength of these three connections is lower than normal people, which means that the increase of connection inhibition is very important for patients to perform specific tasks. Among them, there are two abnormal connections in the supplementary motor area as more important brain areas. In the part of the supporting diagnosis system, we discussed the impact of the change of Theta frequency band on the diagnosis system. We changed the Theta frequency band from 4-8Hz to Low Frequency (4-12Hz). It can be seen that in the case of 4-8Hz, the classification method is SVM and naivebayes algorithm can get higher accuracy, which are 89.4% and 89.8%, respectively. If the frequency band is changed to Low Frequency (4-12Hz), the algorithms with higher accuracy are SVM, Logistic regression, and naivebayes. The accuracy is 90.99%, 90.42%, and 90.55%. In terms of frequency band combination,we can see that in algorithms with higher accuracy, Theta band is included in the best combination, indicating that in our supporting diagnosis system, the Theta band is more able to distinguish between patients and normal people. The important frequency band of the difference. In addition, according to the important characteristic connections analyzed by our supporting diagnostic system, we also found that the connections related to the premotor cortex area and supplementary motor cortex area on the affected side are indeed considered to be used to determine whether it is an important brain area of the patient or normal, Roughly consistent with the results we have calculated. In the future, we hope that the number of subjects we receive can be greater to avoid the phenomenon of over fitting, which will affect the accuracy of auxiliary diagnosis. In the future, more different algorithms can be added to test which algorithm is more suitable for our research. |
