睡眠障礙為現代人很嚴重的文明病，如何達到完善良好的睡眠並給予適當的治 療是很重要的議題，現今很多醫生與科學家都利用大數據以及人工智慧來解決此問 題，為了達到此目的本論文使用近幾年熱門的技術深度學習(Deep learning)，將數 名受測者的 PSG 中 EEG 與 EOG 進行分析，利用小波轉換(Wavelet Transform , WT) 將訊號進行時頻分析，將其中 α、β、σ、θ 四個重要特徵取出，以上得到特徵訊號 放入深度學習網路來做分析，使用的網路架構分為三個不同層的架構，第一層為多 層 感 知 層(Multilayer Perceptron ,MLP) 、第 二 層 為 長短期記 憶 (Long Short-Term Memory,LSTM)與最後一層的歸一化指數函數(Softmax)作為睡眠週期的分類器，經 由其網路訓練準確度高達到 6 成 5，將此模型建立好之後，接下來的目標為運用在 硬體上，搭配雲端醫療系統，更能減少醫療資源，也能在自家做檢測，能減少認床 等可排除的問題，進能達到完善的治療以及資料的建檔。;Sleep disorder is a popular modern civil disease. It has become an important issue of how to provide adequate treatment for sustaining well sleep. In recent years, many doctors and scientists try to collect big data and apply artificial intelligence to solve this problem. In this paper, we use the deep learning neural network to analyze subjects’ PSG EEG and EOG. We used wavelet transform (WT) to decompose the measured EEG signals into α, β, σ, and θ bands, so that the temporal-frequency parameters were obtained as input data for deep learning neural network. The proposed network architecture contianed three different layers. The first layer is Multilayer Perceptron (MLP). The second layer is the Long Short-Term Memory (LSTM) and the last layer of the normalized exponential function (Softmax) as classifiers for the sleep cycle. The detection accuracy of the our study results was 65 percent. The continuing work of this paper will use hardware on a cloud-based medical system to process the sleep data. With cloud computation service, users will be able to perform self-diagnosis and homecare service in their home.
