本論文提出基於深度學習的癲癇偵測演算法應用於長時EEG分析，並基於機器學習提出穿戴式癲癇偵測演算法應用於非醫院環境之癲癇偵測。本論文使用CHB-MIT資料集與波恩資料集對演算法效能進行評估，實驗結果顯示本論文深度學習演算法具備最佳的分類效能。本論文深度學習模型在CHB-MIT資料集受試者相依測試平均f1-score為69.34%，跨受試者測試平均f1-score為37.31%。在波恩資料集本論文深度學習模型平均準確度為98.91%，穿戴式演算法平均準確度為97.13%。本論文將穿戴式演算法實現於超低功耗嵌入式系統計算演算法執行時間，實驗結果顯示本論文所提出熵之估算法能夠減少81.58%的計算時間，與近幾年演算法相比本論文穿戴式演算法提供相當的分類效能並具備最少計算時間。;This thesis proposed a deep learning-based seizure detection algorithm for long-term EEG analysis and a machine learning-based wearable seizure detection algorithm for non-hospital seizure detection. We conducted an experiment on the CHB-MIT and Bonn datasets to evaluate the algorithm′s performance. The experimental results show that our deep learning algorithm has the best classification performance. For the CHB-MIT dataset, our deep learning model achieved an average f1-score of 69.34% in the subject dependence experiment and an average f1-score of 37.31% in the cross-subject experiment. In the Bonn dataset, the average accuracy of the deep learning model and the wearable algorithm is 98.91% and 97.13%, respectively. Our wearable algorithm is implemented on the ultra-low power embedded system and analysis the calculation time of the algorithm. The experimental results show that the proposed entropy estimation method can reduce the calculation time by 81.58%. Compared with the previous algorithms, our wearable algorithm provides a comparable classification performance and has the fastest inference speed.
