基於連續支持向量機的心源性呼吸率偵測

Abstract

呼吸率、血壓、脈搏以及體溫並稱為人體生命徵象，即維持生命的基本要素，是顯示人健康狀態不可少的指標，然而呼吸率卻是這四種徵象中最少被觀測的，儘管不正常的呼吸率已被證明為某些嚴重疾病的預測因子。一般而言，呼吸訊號必須藉由笨重的儀器來量測，不但無法在日常生活中進行長期觀測呼吸訊號，亦有可能因為配戴相關儀器而影響自然的呼吸狀態，導致某些與呼吸訊號相關的應用結果失真，例如：壓力測試、睡眠品質分析等。許多心電訊號及呼吸的共同研究指出心電訊號與呼吸之間有相互影響的關係，因此間接由心電訊號取得呼吸資訊的技術變發展而出，最熟為人知的技術即為ECG-Derived Respiration (EDR)。 現有的EDR方法多以DSP來實現，然而，生理訊號因人而異，且容易受到人體運動或健康狀態所影響，故DSP演算法時常無法有效的考慮多種因素以達更佳的效果；此外，有些方法受限於特定呼吸頻率，並無法廣泛地應用。有鑑於此，根據心電訊號與呼吸訊號之間的關係，我們使用不同演算法實現的EDR及小波轉換結果做為特徵擷取的對象，並且導入了機器學習(Machine Learning, ML) 演算法：Least Absolute Shrinkage and Selection Operator (LASSO) Regression、 Support Vector Machines (SVM)分別進行特徵篩選以及資料分析。利用LASSO Regression分析多種特徵與呼吸訊號是否相關及其相關強度，將相關強度高的特徵整理成特徵向量，並由SVM學習大量特徵向量並建立多個分類不同呼吸率的模型。 本論文所提出之序向支持向量機分類是由多個分類器和呼吸率範圍比分組成，由於真實呼吸率每分鐘介於14至20次呼吸的資料數量較多，相對的二次分類模型較不容易有效率的分類，故每個分類器的結果會有相對應的權重，藉由權重與分類結果對不同呼吸率範圍計算比分，依照比分高的呼吸率範圍來決定呼吸率。 特徵選取及分類器建立所使用的資料來源為Physionet所提供的Fantasia database；此外，測試資料除了上述來源外，另有使用Si2實驗室呼吸感測器研究所蒐集的呼吸訊號和博晶醫電所出產的heartwave ECG感測器，兩個資料來源都包含了連續的心電訊號和呼吸訊號。藉由本論文所提出的基於連續向量支持基的心源性呼吸偵測可達到91.78%的平均正確率，其中正確率大於90%的資料就佔所有資料的77.45%；針對實際呼吸率過低或過高的資料，相較傳統的DSP演算法實現的EDR能有效偵測且正確率為100%；未來若是能與穿戴式感測器結合，便可以長期觀測而不受地點與儀器的限制，達到更深入的健康行動照護。

Respiratory rate, blood pressure, pulse and body temperature are the vital signs of human being. However, some studies have indicated that the record of these signs is poor even in the hospital. Of all four signs, respiratory rate is often not recorded, though the abnormal respiratory rate is reported as an important predictor of serious illnesses. Generally, techniques that record respiratory signal require cumbersome importable devices that may cause uncomfortable feelings and interfere with natural breathing. Some application based on the respiratory analysis may even fail such as sleep quality analysis and stress testing. Fortunately, thanks to the joint study of respiratory and electrocardiography, some studies have reported the possibility of indirect methods to extract the respiration information which is well-known as ECG-Derived Respiration (EDR). The existing EDR approaches are achieved by the DSP method. However, the ECG signal is easily influenced by the body motion and the individual health status. The DSP-based EDR is not capable to generate a general model for all cases and is often limited by the specific respiratory frequency. Therefore, we adopted several EDR algorithms and the wavelet transform for the feature extraction. The output features are introduced to machine learning (ML) algorithms: least absolute shrinkage and selection operator (LASSO) regression, support vector machines (SVM) for advance feature select and data analysis. The proposed sequential SVM comprises multiple classifiers and a region score. Since the models for classifying the respiratory rate from 14 to 20 breathes per minute (bpm) are hard to learn, we assign a weight to each model. By calculating the region score with the classification result and the corresponding weight, the respiratory rate is detected. The data source used for the feature selection and the model learning is obtained from Physionet Fantasia database. Besides, we also use the data collected by Si2-lab respiratory sensor studies and the heartwave ECG sensor of bOMDIC Inc. Both data source contains the continuous respiratory signal and the ECG signal. By using the presented EDR-oriented respiratory rate detection based on the sequential SVM, the average accuracy is 91.78%, and 77.45% of the data achieve the accuracy more than 90%. For the cases of high respiratory rate and low respiratory rate, our work performs better than DSP-based EDR, and achieves 100% accuracy. With miniaturized-sensors-integrated mobile devices, it enables the opportunities for ordinary daily healthcare applications, e.g. continuous cardiac signal monitoring of ECG. Accordingly, if the respiratory rate is correctly extracted from cardiac signal, the healthcare system can provide more information for the in-depth monitoring.