為了實現快速篩檢睡眠呼吸中止症，本研究並透過壓電感測器擷取頸部振動訊號同時監測打鼾震動訊號及頸動脈搏動，使用HRV的方法找出連續睡眠間期再使用深度學習的方法檢測此期間的打鼾事件並加以分類分級。本研究通過長庚醫療財團法人人體試驗委員會之審核通過。另一方面，本研究所使用的深度學習模型為結合卷積神經網路與門控循環單元之模型。為訓練深度學習模型，挑選20位受試者進行打鼾事件的標註，將其分類為無打鼾、打鼾以及噪音，總共標註2870個片段，其中打鼾訊號有1304個、非打鼾數有1167個、噪音有399個。最終嚴重程度分級的部分，挑選40位受試者進行驗證，以本系統所檢測的嚴重程度與睡眠中心進行比較，發現若受試者有多個連續睡眠間期，且第一睡眠間期是在睡眠開始的前30分鐘，會因睡眠狀態不穩定而導致分析結果出現誤差。若排除此狀況，最終驗證結果之準確性為90.0%，靈敏性為100%，特異性為78.9%，證實本系統在OSA的初步檢測上是可信的。本研究成功結合HRV以及深度學習兩種方法檢測病患在睡眠間期發生呼吸中止症事件次數以及睡眠時間長度來檢測病患的嚴重程度，達成居家式快篩系統。

To achieve rapid obstructive sleep apnea screening, in this research, the vibration signal from the neck is captured by piezoelectric sensor. The sensor acquires the snoring vibration and the carotid pulse signals at the same time. Then, the HRV method is used to determine the continuous sleep period and the deep learning method is used to detect snoring events to evalue the severity. This research was approved by the Chang Gung Medical Foundation Institutional Review Board. In addition, the deep learning model that was used in the study was a model combining convolution neural network and gated recurrent unit. Signals from twenty subjects were selected to label the snoring events, categorized as non-snoring, snoring, and noise for deep learning model training. Two thousand eight hundred and seventy snoring events were marked, of which, there are 1304 snoring signals, 1167 non-snoring signals, and 399 noise signals. In the final part of the severity classification, 40 subjects were selected for verification. The severity level detected by the system was compared with that of the sleep center. In the case of subjects had multiple consecutive sleep periods, and the first sleep period was within 30 minutes from the onset of sleep, there were large discrepancies in the analysis results due to this unstable sleep state. If this condition is excluded, the accuracy, sensitivity, and specificity of the final test results are 90.0%, 100%, and 78.9%, confirming that the system is reliable for OSA's initial testing. This research successfully combined HRV and deep learning methods to identify the number of apnea events during sleep and sleep duration to evaluate the patients' severity of apnea, resulting in an at-home rapid screening system.