以心電訊號預測睡眠呼吸事件之深度學習模型

Taipei Medical University Institutional Repository > 醫學院 > 人工智慧醫療碩士在職專班 > 博碩士論文 > Item 987654321/62461

請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/62461

題名:	以心電訊號預測睡眠呼吸事件之深度學習模型 A Deep Learning Model for Predicting Respiratory Events during Sleep by ECG Signals
作者:	梁仲偉 LIANG, CHUNG-WEI
貢獻者:	醫學院人工智慧醫療碩士在職專班劉文德
關鍵詞:	深度學習;遷移式學習;連續小波轉換;呼吸中止;呼吸不足;心電訊號 Deep Learning;Transfer Learning;Continuous Wavelet Transform;Apnea;Hypopnea;ECG
日期:	2022-07-01
上傳時間:	2023-01-17 14:52:53 (UTC+8)
摘要:	呼吸中止與呼吸不足是睡眠呼吸疾病中很常見的症狀。一般的檢驗方式是患者須至專業醫療院所做睡眠多項生理檢測後，由專業技師標註並由專科醫師判讀診斷。較常見的治療方式是患者配戴連續陽壓呼吸器來改善夜間呼吸中止與呼吸不足現象。由於臨床觀察發現，睡眠時，呼吸道完全塌陷前，人體可能有相對應的呼吸補償動作來試圖阻止呼吸道塌陷所造成的呼吸氣流下降，但此現象於連續陽壓呼吸器並無法測得，因此患者可能在連續陽壓呼吸器加壓前仍會感到不適。為確認臨床觀察到呼吸補償現象，本論文利用心電訊號轉換成小波量值圖，並利用遷移式學習方式進行建模。由於並不確定哪種模型能有較佳的效果，因此使用7種不同模型並利用本論文收集的回溯性資料及另外兩個Dublin和MIT-BIH公開資料集進行模型建立與評估，並從中找出整體最佳的EfficientNetB4模型。利用EfficientNetB4模型驗證睡眠呼吸事件開始後30秒及睡眠呼吸事件開始前0到90秒，呼吸中止或呼吸不足的偵測力可達到0.85以上的Accuracy和Marco F1 Score。以此確認可透過心電訊號發現人體於睡眠呼吸事件開始前即有對應的反應，且透過深度學習模型是能夠有效辨識，此外換個方向想，也間接驗證可透過偵測睡眠呼吸事件前的訊號來預測睡眠呼吸事件即將發生。 Apnea and hypopnea are common symptoms of sleep-disordered breathing. Patient goes to the professional hospital to take Polysomnography (PSG), and annotation by professional technician and diagnosis by specialist. Patient wears a Continuous Positive Airway Pressure (CPAP) to improve the appearance of apnea and hypopnea at night is the most common treatment in these days. According to clinical observation, when entering sleep, and before the airway collapses completely, the human body may have corresponding breathing compensation actions to try to prevent the respiratory airflow from falling caused by the collapse of the airway, but this appearance cannot be measured by CPAP, so that the patient still feel uncomfortable before the CPAP is pressurized. In order to confirm the clinical observation of respiratory compensation phenomenon, this paper uses ECG signal convert to Scalogram, and uses transfer learning method to build model. Since it is not sure which model can have the best effect, therefore this paper uses 7 different models and use the retrospective dataset from this paper along with the other two public datasets, Dublin and MIT-BIH to build and evaluate the model, and find the overall best EfficientNetB4 model. Using the EfficientNetB4 model to verify that 30 seconds after the start of a sleep breathing event and 0 to 90 seconds before the start of a sleep breathing event. The detection power of apnea or hypopnea can reach Accuracy and Marco F1 Score above 0.85. This confirms that the human body has a corresponding response before the sleep breathing event can be found through the ECG signal, and the deep learning model can effectively identify it. In addition, thinking in another direction, it is also indirectly verified that the sleep breathing event can be predicted by detecting the signal before the sleep breathing event.
描述:	碩士指導教授：劉文德委員：劉文德委員：何淑娟委員：彭徐鈞
資料類型:	thesis
顯示於類別:	[人工智慧醫療碩士在職專班] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	294	檢視/開啟

在TMUIR中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....