以多頻道腦電波貢獻值輔助定位棘波腦部區域之研究

Abstract

癲癇症於世界上並非罕見疾病，而目前醫學上最有效的診治方式為：使用腦波儀量測患者之長期腦電波，醫療人員再從不同樣態的腦電波，來分析患者的發病狀況加以醫治；但是長期腦電波的資料量相當可觀，如由醫師人工判讀將會曠日費時；此外，癲癇患者在發病當下出現的抽搐與不可控行為是患者主要受傷原因，掌握發病前兆（棘波）施以適當處置遂成為重要研究目標；因此，為了能夠減輕醫療人員負擔、找出患者即將發生癲癇之腦部區域，幫助醫療人員更精確的診治患者腦部，本研究提出一演算機制，從多頻道腦電波中定位出棘波的空間位置、時間點，從而輔助醫療人員做出相應措施來減輕傷害；在實驗設計上，會從癲癇患者身上收集16個頻道腦波紀錄，將16個單極腦電波頻道轉至雙極腦電波，並進行預處理與切割分段，再經由總體經驗模態分解法，將腦電訊號分解成八個本質模態函數，從每個本質模態函數提取特徵，並用線性判別分析訓練分類器來辨識棘波與非棘波二分類資料；在建模分類上每個特徵皆有對應的權重值，本研究使用此權重值為基礎，計算出每個特徵在分類上的貢獻值，並藉由特徵貢獻值回溯至頻道貢獻值，最後再依照頻道貢獻值於時間序列上的變化來判斷發生棘波之頻道與時間段；實驗成果準確率達94.9%，召回率為93.8%，不止補足[1]未能指出發生異狀頻道之不足，並在準確率與召回率表現皆優於最新文獻[2]。實驗結果，使用頻道貢獻值能夠有效推斷出該頻道是否正在發生棘波，以此方法也能幫助醫師明確指出每個頻道在哪些時點出現異狀。

Epilepsy is not a rare disease in the world. At present, the most effective way of diagnosis and treatment in medicine is to measure the patient’s long-term electroencephalography. However, the long-term electroencephalography data is considerable, such as manual interpretation by a doctor will be time-consuming. In addition, the main cause of the patient’s injury is withdrawal and uncontrollable behavior of patients at epilepsy. Appropriate treatment has become an important research goal. Therefore, in order to reduce the burden on medical staff, find out the brain area of patients who are about to develop epilepsy, and help medical staff to diagnose and treat the patient’s brain more accurately, this study proposes a calculation mechanism. Locate the spatial location and time point of the spike wave in the multi-channel electroencephalography, so as to assist medical personnel to take corresponding measures to reduce the injury. In terms of experimental design, 16-channel electroencephalography records will be collected from patients with epilepsy, 16 unipolar electroencephalography channels will be transferred to bipolar electroencephalography, pre-processing and segmentation will be performed, and then the total empirical mode decomposition method will be used, Decompose the electroencephalography signal into eight essential modal functions, extract features from each essential modal function, and use linear discriminant analysis to train the classifier to identify the two-classification data of spikes and non-spurs. Each in the modeling category each feature has a corresponding weight value. This study uses this weight value as the basis to calculate the contribution value of each feature in the classification, and traces back to the channel contribution value by the feature contribution value, and finally according to the channel contribution value in the time series The above changes are used to determine the channel and time period where the spike occurs; the accuracy rate of the experimental results is 94.9%, and the recall rate is 93.8%. The performance is better than the latest literature. As a result of the experiment, using the channel contribution value can effectively infer whether a spike is occurring on the channel, and this method can also help the physician to clearly point out when each channel is abnormal.