TWI772086B

TWI772086B - A method of monitoring apnea and hypopnea events by using fully convolutional networks

Info

Publication number: TWI772086B
Application number: TW110124440A
Authority: TW
Inventors: 林俊成
Original assignee: 國立勤益科技大學
Priority date: 2021-07-02
Filing date: 2021-07-02
Publication date: 2022-07-21
Also published as: TW202302041A

Abstract

A method of monitoring apnea and hypopnea events by using fully convolutional networks includes a measurement step and a monitoring and identifying step. The measurement step includes getting electrocardiogram signals of a subject by measuring. The monitoring and identifying step includes using training sets and testing sets of electrocardiogram signals of different subjects as a preferable monitoring and identifying model and using a deep learning model of fully convolutional networks for monitoring and identifying the electrocardiogram signals. During the convolution, the probability calculation based on a shared overlap area mode is adopted to extract and classify eigenvalues of the electrocardiogram signals of the subject and finally output an identified result. The method not only requires much less computation but also increases the accuracy of monitoring the degree of severity of apnea events of the subject efficiently.

Description

使用全卷積神經網路之呼吸暫停與不足事件偵測方法Apnea and hypopnea event detection method using fully convolutional neural network

本發明是有關於一種睡眠呼吸功能障礙的偵測，特別是指一種使用全卷積神經網路之呼吸暫停與不足事件偵測方法。The present invention relates to a detection of sleep-disordered breathing, in particular to a method for detecting apnea and insufficiency events using a fully convolutional neural network.

查，阻塞睡眠呼吸暫停(Obstructive Sleep Apnea；以下簡稱OSA)是一種常見且嚴重的睡眠呼吸功能阻礙，這是一種會在睡眠期間因咽部塌陷影響，造成完全或部分上呼吸道阻塞，進而導致呼吸暫停或減弱，同時根據先前的研究顯示，會發生阻塞睡眠呼吸暫停與高血壓、冠心病、心律失常、心臟衰竭和中風的發病率有關，依據目前評估OSA嚴重程度標準方法是透過睡眠多項生理檢查(Polysomnography；以下簡稱PSG)，即受試者必須到睡眠實驗室或睡眠中心睡一個晚上，且在護理人員的監督下，分別在頸部、眼角、下巴、心臟以及腿部貼上電極貼片，並且於胸部及腹部套上感應帶，在手指套上血氧測量器，在口鼻套上呼吸感應器，以及在手臂上套上血壓計，有效透過前述該等感應與測量器來記錄整個晚上的睡眠生理數據，包括腦電圖、眼電圖、心電圖、下巴肌電圖、胸部呼吸訊號、腹部呼吸訊號、口鼻氣流、血氧濃度、血壓變化、心率，以及睡眠體位等，而PSG是結合呼吸氣流、胸部呼吸訊號、腹部呼吸訊號、以及血氧濃度來判斷並計算受試者每小時平均出現的呼吸暫停(Apnea)與呼吸不足(Hypopnea)事件的次數(即呼吸暫停與呼吸不足指標；Apnea and Hypopnea Index (AHI))，藉以評估受試者OSA的嚴重程度，包括呼吸正常(Normal；AHI ＜5)、輕度OSA(Mild；AHI介於5到14)、中度OSA(Moderate；AHI介於15到30) 、以及嚴重OSA (Severe； AHI＞ 30)。Obstructive Sleep Apnea (hereinafter referred to as OSA) is a common and serious sleep-disordered breathing disorder, which can cause complete or partial upper airway obstruction due to the collapse of the pharynx during sleep, which in turn leads to breathing Suspended or weakened, while obstructive sleep apnea can occur according to previous studies, and the incidence of hypertension, coronary heart disease, cardiac arrhythmias, heart failure and stroke, according to the current standard method for assessing the severity of OSA is through multiple physiological examinations of sleep (Polysomnography; hereinafter referred to as PSG), that is, subjects must go to a sleep laboratory or sleep center to sleep for one night, and under the supervision of a nursing staff, electrode patches are attached to the neck, the corners of the eyes, the chin, the heart and the legs, respectively. , and put the induction belt on the chest and abdomen, put the blood oximeter on the finger, put the breathing sensor on the nose and mouth, and put the blood pressure monitor on the arm, and effectively record the whole through the aforementioned sensors and measuring devices. Sleep physiological data at night, including electroencephalogram, electrooculogram, electrocardiogram, chin EMG, chest breathing signal, abdominal breathing signal, nasal airflow, blood oxygen concentration, blood pressure changes, heart rate, and sleep position, etc., while PSG It combines respiratory airflow, chest breathing signal, abdominal breathing signal, and blood oxygen concentration to determine and calculate the average number of apnea (Apnea) and hypopnea (Hypopnea) events per hour (ie, apnea and hypopnea). Index; Apnea and Hypopnea Index (AHI)), which assesses the severity of OSA in subjects, including normal breathing (Normal; AHI <5), mild OSA (Mild; AHI between 5 and 14), moderate OSA ( Moderate; AHI between 15 and 30), and severe OSA (Severe; AHI > 30).

接續前述，有鑒於PSG檢查的費用昂貴且不便，所以近年來便有人致力於研究用量測較少的訊號來開發方便且花費少的呼吸暫停與不足事件偵測系統，其主要被使用的訊號有血氧濃度、呼吸氣流、胸部呼吸、心電圖、聲音訊號，以及結合不同的訊號；然而，因為PSG主要是結合呼吸訊號(呼吸氣流、胸部呼吸及腹部呼吸)與血氧濃度來檢測呼吸暫停與呼吸不足事件，如果單獨使用呼吸氣流、胸部呼吸、腹部呼吸或血氣濃度時，將無法檢測所有的呼吸暫停與呼吸不足事件，同時基於聲音訊號檢測方式則受限於聲音很容易受到心臟聲音與環境噪音的干擾，相較於單獨使用呼吸氣流、胸部呼吸訊號、血氣濃度與聲音訊號，單導程心電圖則是能夠較好的反應出完整呼吸事件訊號，且在先前的研究結果中，基於單導程心電圖的檢測方法在辨識呼吸暫停與呼吸不足時具有較高的準確度；同時，再根據心電圖形訊號來偵測呼吸暫停與呼吸不足事件的原理，是在發生呼吸暫停或呼吸不足事件時，心率會緩慢下降，而在恢復正常呼吸時，心率會加快之後再恢復正常，請參閱圖1，在圖1中顯示的是PGS量測到的呼吸氣流、胸部呼吸訊號、腹部呼吸訊號、心電圖形訊號以及PSG所提供的呼吸註記(準位0表示呼吸正常，準位2表示呼吸暫停間)，心跳間隔時間訊號(RR間隔訊號)則是心電圖形訊號中相鄰R波的間隔時間所組成的訊號，因此從圖1中可以觀察到呼吸暫停期間，心跳間隔時間訊號的變化緩慢，但是呼吸暫停結束之後，心跳間隔時間訊號明顯的減少且持續一段時間之後再恢復正常，是以，如果在原本正常平穩的心跳間隔時間訊號之後，持續出現一段心跳時間訊號的減少再恢復正常平穩的心跳間隔時間訊號，則代表出現一次呼吸暫停或呼吸不足事件，也稱為呼吸暫停與呼吸不足事件的心跳間隔時間變化模式；是以，對目前基於單導程心電圖與機器學習的呼吸暫停和不足的檢測方法，雖已有使用具真實效能且可以測試最佳化後的機器學習模型來進行，但其必需通過數百個百萬浮點數運算來對受測者整晚的心電圖形訊號進行偵測，如此一來對偵測裝置的運算量將會造成很大的負擔，藉此，為能有效進行降低偵測運算量的負擔與正確性等問題加以改進研發，俾提出本案專利申請。Continuing from the above, in view of the high cost and inconvenience of PSG examination, in recent years, some people have devoted themselves to developing a convenient and low-cost apnea and insufficiency event detection system with less measured signals. There are blood oxygen concentration, respiratory airflow, chest breathing, electrocardiogram, sound signal, and combining different signals; however, because PSG mainly combines respiratory signals (respiratory airflow, chest breathing and abdominal breathing) and blood oxygen concentration to detect apnea and Hypopnea events, if the respiratory airflow, chest breathing, abdominal breathing or blood gas concentration are used alone, it will not be able to detect all apnea and hypopnea events. At the same time, the detection method based on sound signals is limited because the sound is easily affected by the heart sound and the environment. Compared with the interference of noise, single-lead ECG can better reflect the complete respiratory event signal compared with the use of respiratory airflow, chest breathing signal, blood gas concentration and sound signal alone. The detection method of Cheng ECG has high accuracy in identifying apnea and hypopnea; at the same time, the principle of detecting apnea and hypopnea events based on ECG signals is that when apnea or hypopnea events occur, The heart rate will decrease slowly, and when returning to normal breathing, the heart rate will increase and then return to normal. Please refer to Figure 1. In Figure 1, it shows the respiratory airflow, chest breathing signal, abdominal breathing signal, and electrocardiogram measured by PGS. The signal and the respiration note provided by PSG (level 0 means normal breathing, level 2 means apnea), and the heartbeat interval time signal (RR interval signal) is composed of the interval time between adjacent R waves in the ECG signal. Therefore, it can be observed from Figure 1 that during the apnea, the heartbeat interval time signal changes slowly, but after the apnea ends, the heartbeat interval time signal decreases significantly and returns to normal after a period of time. Therefore, if the original After a normal and stable heartbeat interval time signal, there is a continuous decrease of the heartbeat time signal for a period of time, and then the normal and stable heartbeat interval time signal is restored, which represents an apnea or hypopnea event, also known as apnea and hypopnea event. Temporal variation pattern; therefore, the current detection methods for apnea and insufficiency based on single-lead ECG and machine learning have been carried out using real-world and optimized machine learning models, but they must be Through hundreds of millions of floating point number operations to detect the ECG signal of the subject throughout the night, it will cause a great burden on the calculation amount of the detection device. To reduce the burden of detection calculation and correctness and other issues, improve research and development, so as to file the patent application in this case.

因此，本發明之目的，是在提供一種使用全卷積神經網路之呼吸暫停與不足事件偵測方法，其得以通過全卷積神經網路技術之共享重疊區域的計算機率方式，針對受測者之心電圖形訊號進行偵測辨識，即能有效快速偵測出受測者具有呼吸暫停與呼吸不足事件的嚴重程度。Therefore, the purpose of the present invention is to provide a method for detecting apnea and insufficiency events using a fully convolutional neural network, which can be used for a subject to be tested in a computationally efficient way of sharing overlapping regions of the fully convolutional neural network technology. By detecting and identifying the ECG signal of the subject, it can effectively and quickly detect the severity of the apnea and hypopnea events of the subject.

於是，本發明使用全卷積神經網路之呼吸暫停與不足事件偵測方法，包含有量測步驟及偵測辨識等步驟；其中，先由該量測步驟中所備具之心電圖機量測模組，得以對受測者分別進行心臟跳動、其周圍肌肉節律性收縮與呼吸反應等，記錄而形成一心電圖形訊號，並經該偵測辨識步驟之運算處理模組先選用不同受測者之心電圖形訊號的訓練資料集及測試資料集來作為較佳的偵測辨識模型，再以一全卷積神經網路技術的深度學習模型，以對帶入之該心電圖形之訊號進行呼吸暫停與呼吸不足事件的偵測與辨識，利用在卷積過程中以共享重疊的計算，使該等心電圖形訊號受到正規化處理、被執行特徵提取而獲得較佳的多個心電圖形訊號特徵圖、並對該等特徵圖轉換為特徵向量及進行計算機率的偵測，最終輸出一辨識結果，藉此可使整體所需運算量大幅減少的情況下，得以有效偵測辨識分類出受測者之呼吸暫停事件嚴重的準確性。Therefore, the method for detecting apnea and insufficiency events using a fully convolutional neural network in the present invention includes the steps of measurement, detection and identification, and the like; wherein, the measurement is performed by the electrocardiograph provided in the measurement step. The module can record the heart beat, the rhythmic contraction of the surrounding muscles and the respiration response of the subject respectively, and record and form an electrocardiogram signal, and the operation processing module of the detection and identification step first selects different subjects The training data set and test data set of the ECG signal are used as a better detection and identification model, and then a deep learning model of fully convolutional neural network technology is used to carry out apnea on the ECG signal brought in. and detection and identification of hypopnea events, the ECG signals are subjected to normalization processing and feature extraction is performed to obtain better multiple ECG signal feature maps, using shared overlapping calculations in the convolution process. And convert these feature maps into feature vectors and carry out computerized detection, and finally output an identification result, which can effectively detect, identify and classify the subject under the condition that the overall required calculation amount can be greatly reduced. Severe accuracy for apnea events.

有關本發明之前述及其他技術內容、特點與功效，在以下配合參考圖式之較佳實施例的詳細說明中，將可清楚的明白。The foregoing and other technical contents, features and effects of the present invention will be clearly understood in the following detailed description of the preferred embodiments with reference to the drawings.

參閱圖2，本發明一較佳實施例，一種使用全卷積神經網路之呼吸暫停與不足事件偵測方法，包含有一量測步驟以及一偵測辨識步驟等；其中，在該量測步驟中備具有一心電圖機，而該心電圖機可針對心臟律動及呼吸頻率感應進行量測，以針對受試者的胸部之心臟自發性跳動與周圍肌肉節律性收縮，且依心臟組織電壓變化記錄而形成一心電圖形訊號。Referring to FIG. 2, a preferred embodiment of the present invention, a method for detecting apnea and insufficiency events using a fully convolutional neural network, includes a measurement step and a detection and identification step, etc.; wherein, in the measurement step The equipment is equipped with an electrocardiograph, which can measure the heart rhythm and respiratory rate induction to measure the spontaneous beating of the heart and the rhythmic contraction of the surrounding muscles in the chest of the subject, and record the changes in the voltage of the heart tissue. An electrocardiogram signal is formed.

接續前述，該偵測辨識步驟備具有一運算處理模組，使該運算處理模組以一全卷積神經網路技術的深度學習模型進行訓練與測試，同時該運算處理模組更可適當搭配一滑動視窗法來與全卷積神經網路技術的深度學習模型配合排列演算，而該深度學習模型為使用記錄有各自獨立且選自不同的受試者之心電圖形訊號的訓練資料集及測試資料集為進行偵測/辨識的資料，以用深度學習計算進行訓練演算以產生偵測辨識結果的基準，同時前述所使用之該訓練資料集與測試資料集的資料是採用睡眠心臟健康研究(Sleep Heart Health Study；簡稱SHHS)所提供的睡眠多項生理檢查(Polysomnography；簡稱PSG)資料庫來建立，而前述該訓練資料集與測試資料集分別包括呼吸正常，以及呼吸暫停與呼吸不足組的30秒心電圖形訊號與心跳間隔時間訊號，以利用該訓練資料集的心電圖形訊號與心跳間隔時間訊號用於訓練出最佳化的深度學習模型，以辨識輸入的心電圖形訊號對應呼吸正常或是呼吸暫停與呼吸不足事件，而該測試資料集的心電圖形訊號與心跳間隔時間訊號是用於測試最佳化後的機器學習模型對於訓練資料集以外的心電圖形訊號與心跳間隔時間訊號的辨識正確性，可以測試最佳化後的機器學習模型的真實效能。Continuing from the above, the detection and identification step is equipped with an arithmetic processing module, so that the arithmetic processing module is trained and tested with a deep learning model of full convolutional neural network technology, and the arithmetic processing module can be appropriately matched A sliding window method is used to arrange and calculate with the deep learning model of the fully convolutional neural network technology. The data set is the data for detection/identification, and is used to perform training algorithms with deep learning calculations to generate the benchmark for detection and identification results. At the same time, the data of the training data set and the test data set used above are based on the sleep heart health research ( The Sleep Heart Health Study (SHHS) provides a sleep polysomnography (PSG) database, and the aforementioned training data set and test data set respectively include normal breathing, and 30 patients in the apnea and hypopnea groups. Second ECG signal and heartbeat interval time signal, to use the ECG signal and heartbeat interval time signal of the training data set to train an optimized deep learning model to identify whether the input ECG signal corresponds to normal breathing or breathing Pause and hypopnea events, and the ECG signal and heartbeat interval time signal of the test data set are used to test the accuracy of the optimized machine learning model for the identification of the ECG signal and heartbeat interval time signal outside the training data set , which can test the real performance of the optimized machine learning model.

再者，請參閱圖3，在本實施例中該運算處理模組以一全卷積神經網路技術的深度學習模型，其特色是針對特徵值提取與分類都採卷積層來實現，因此，該全卷積神經網路技術的深度學習模型的使用包括由具有至少八個結構相同的特徵提取層，一與該八個特徵提取層連接之第一卷積層，一與該第一卷積層連接之第二個卷積層，以及一與該第二卷積層連接第三個卷積層等計算模式所組成，而前述每一特徵提取層包括有一個可取得至少45個1D特徵圖的卷積層、一個批次標準化層、一個激活層、一個池化大小為2的最大池化層及一個具有50%捨棄率的捨棄層，同時該等特徵提取層可針對待測該心電圖形訊號進行正規化處理，以及對該心電圖形訊號執行特徵提取與獲得較佳的多個心電圖形訊號特徵圖，而該第一個卷積層則以採用1×11的運算模式針對該等心電圖形訊號特徵圖中的特徵向量進行計算機率，以至少取得2000個特徵圖輸出，而該第二個卷積層則以採用1×1的運算模式，以針對該第一卷積層取得的該等特徵圖中的特徵向量進行計算機率，並至少取得1000個特徵圖輸出，至於該第三個卷積層則得以採用1×1的運算模式針對該第二卷積層所取得的該等特徵圖的特徵向量再進行計算機率，再至少取得2個特徵圖輸出等運算機率，且在機率高的類別即為辨識的結果，至於該滑動視窗法則對經該全卷積神經網路技術的深度學習模型完成該心電圖形訊號的模型訓練與測試後配合排列演算，並依據該滑動視窗法之視窗的大小來收集某個動作發生前或後的動作，並配合比重值的計算與演算，即如圖4之範例所示，當以3分鐘長度(18000個取樣點)的待測心電圖形訊號為例，輸入的訊號長度為18000個取樣點，經過該八層特徵提取層之後，可輸出45個1×70個特徵圖，經過該第一卷積層的特徵向量進行計算機率之後輸出2000個1×60個特徵圖，再經該第二卷積層的特徵向量進行計算機率之後輸出1000個1×60特徵圖之後，而後由該第三個卷積層的特徵向量進行計算機率輸出2個1×60個特徵圖，最後便會得到60個呼吸暫停與呼吸不足事件之分類機率的辨識結果；因此，當該運算處理模組以全卷積神經網路技術的深度學習模型方式來對輸入之心電圖形訊號進行運算的偵測與辨識時，其得以在卷積的過程與搭配滑動視窗法進行滑動且以共享重疊區域的計算方式對該心電圖形訊號進行特徵值的提取與分類，藉此可使整體偵測辨識過程中所需的運算量大幅減少，以免過多的運算量會造成偵測的負擔與正確的辨識性，有效快速偵測辨識分類出受測者之呼吸暫停事件嚴重的準確性。Furthermore, please refer to FIG. 3, in the present embodiment, the operation processing module adopts a deep learning model of fully convolutional neural network technology, which is characterized by adopting convolutional layers for feature value extraction and classification. Therefore, The use of the deep learning model of the fully convolutional neural network technology includes at least eight feature extraction layers with the same structure, a first convolutional layer connected with the eight feature extraction layers, and a first convolutional layer connected with the first convolutional layer. It consists of a second convolutional layer, and a third convolutional layer connected to the second convolutional layer. Each feature extraction layer includes a convolutional layer that can obtain at least 45 1D feature maps, a A batch normalization layer, an activation layer, a max pooling layer with a pooling size of 2, and a dropout layer with a 50% dropout rate, and these feature extraction layers can be normalized for the ECG signal to be tested. and perform feature extraction on the ECG signal and obtain a plurality of better ECG signal feature maps, and the first convolution layer uses a 1×11 operation mode for the feature vectors in the ECG signal feature maps Carry out computer calculations to obtain at least 2000 feature map outputs, and the second convolutional layer uses a 1×1 operation mode to perform computer calculations for the feature vectors in the feature maps obtained by the first convolutional layer , and obtain at least 1000 feature map outputs. As for the third convolutional layer, the 1×1 operation mode can be used to calculate the feature vectors of the feature maps obtained by the second convolutional layer, and then at least obtain The two feature maps output equal computing probability, and the category with high probability is the result of identification. As for the sliding window rule, the deep learning model of the fully convolutional neural network technology completes the model training and testing of the ECG signal. Then cooperate with the arrangement calculation, and collect the actions before or after a certain action according to the window size of the sliding window method, and cooperate with the calculation and calculation of the specific gravity value, that is, as shown in the example in Figure 4, when the length of 3 minutes is used. (18,000 sampling points) of the ECG signal to be measured as an example, the length of the input signal is 18,000 sampling points. After passing through the eight feature extraction layers, 45 1×70 feature maps can be output. After the first volume After the feature vector of the product layer is calculated, 2000 1×60 feature maps are output, and then 1000 1×60 feature maps are output after the feature vector of the second convolutional layer is calculated, and then the third convolutional layer is used. The eigenvectors are calculated to output 2 1×60 feature maps, and finally the identification results of the classification probabilities of 60 apnea and hypopnea events will be obtained; therefore, when the operation processing module uses a fully convolutional neural network When detecting and identifying the input ECG signal by the deep learning model method of technology, it can slide in the process of convolution and the sliding window method and use the calculation method of sharing the overlapping area to calculate the ECG signal. The extraction and classification of eigenvalues can greatly reduce the amount of computation required in the overall detection and identification process, so as not to cause the burden of detection and positive detection due to excessive computation. Accurate identification, effective and rapid detection and identification of severe apnea events in the subject.

是以，本發明主要針對受測者是否具有呼吸暫停與呼吸不足事件時，當該等心電圖形訊號在經過卷積層時，其可在卷積的運算過程進行滑動視窗法的滑動並且共享重疊區域的計算，得以有別於習知利用機器學習模型之滑動視窗法的每一個窗口的訊號都要獨立輸入模型進行整個模型計算的缺失，同時本發明的使用可使得整體所需的運算量可以大幅減少，即如圖4所示之範例，即當待偵測的心電圖形訊號為1×18000，其長度是基於全卷積神經網路深度學習模型輸入訊號的6倍，因此等於只需要6次用全卷積神經網路深度學習模型運算量，而若使用以往機器學習模型之滑動視窗法便需要進行60次模型的運算量，由此可見，透過本發明該使用全卷積神經網路之呼吸暫停與呼吸不足偵測方法只需要以往偵測法的10%運算量，等於節省了90%的運算量，故，本發明所提出的使用全卷積神經網路之呼吸暫停與不足事件偵測方法之技術，可有效大幅減少呼吸暫停與呼吸不足偵測時所需要的運算量，即能有效快速偵測出受測者具有呼吸暫停與呼吸不足事件的嚴重程度。Therefore, the present invention is mainly aimed at whether the subject has apnea and hypopnea events. When the ECG signals pass through the convolution layer, they can perform sliding window sliding and share the overlapping area during the convolution operation. It is different from the conventional sliding window method using the machine learning model, in which the signal of each window must be independently input into the model to perform the entire model calculation, and the use of the present invention can make the overall required calculation amount can be greatly reduced. Reduction, that is, the example shown in Figure 4, that is, when the ECG signal to be detected is 1×18000, its length is 6 times that of the input signal based on the fully convolutional neural network deep learning model, so it is equivalent to only need 6 times Using a fully convolutional neural network to deeply learn the model's computational complexity, if the sliding window method of the previous machine learning model is used, it requires 60 model computations. It can be seen that, through the present invention, the full convolutional neural network is used to solve the problem. The apnea and hypopnea detection method only needs 10% of the calculation amount of the previous detection method, which is equivalent to saving 90% of the calculation amount. Therefore, the apnea and hypopnea event detection method using a fully convolutional neural network proposed by the present invention The technology of the measurement method can effectively and greatly reduce the amount of computation required for the detection of apnea and hypopnea, that is, it can effectively and quickly detect the severity of the apnea and hypopnea events in the subject.

歸納前述，本發明使用全卷積神經網路之呼吸暫停與不足事件偵測方法，其主要針對受測者是否具有呼吸暫停與呼吸不足事件，並利用該量測步驟所得之心電圖形訊號，經該偵測辨識步驟之運算處理模組以一全卷積神經網路技術的深度學習模型技術的模式下，並選自不同的受試者之心電圖形訊號的訓練資料集與測試資料集的資料做為偵測辨識之基準，並適當搭配一滑動視窗法來配合排列演算，並在卷積的過程進行滑動並以共享重疊區域的計算機率方式進行，使該待測者的心電圖形訊號受到正規化處理、被執行特徵提取而獲得較佳的多個心電圖形訊號特徵圖、並對該等特徵圖轉換為特徵向量及進行計算機率，以進一步採卷積層來實現對該心電圖形訊號的特徵值提取與分類，藉此得以避免產生過多的運算量會造成偵測的負擔，俾利在正確的辨識性的情況下，最終輸出一辨識結果，藉此得以有效快速地提升偵測出受測者之呼吸暫停事件嚴重的準確性。Summarizing the above, the present invention uses a fully convolutional neural network method for detecting apnea and hypopnea events, which is mainly aimed at whether the subject has apnea and hypopnea events. The operation processing module of the detection and identification step is based on the deep learning model technology of a fully convolutional neural network technology, and is selected from the data of the training data set and the test data set of the electrocardiographic signal of different subjects As the benchmark for detection and identification, it is appropriately matched with a sliding window method to cooperate with the arrangement calculation, and the sliding is performed in the process of convolution and is carried out in a computerized manner of sharing the overlapping area, so that the ECG signal of the subject is normalized. After processing, feature extraction is performed to obtain a plurality of better feature maps of the ECG signal, and these feature maps are converted into feature vectors and computerized to further use convolution layers to realize the feature values of the ECG signal. Extraction and classification, so as to avoid the burden of detection caused by excessive computation, so as to finally output an identification result under the condition of correct identification, so as to effectively and quickly improve the detection of the subject Accuracy of severe apnea events.

惟以上所述者，僅為說明本發明之較佳實施例而已，當不能以此限定本發明實施之範圍，即大凡依本發明申請專利範圍及發明說明書內容所作之簡單的等效變化與修飾，皆應仍屬本發明專利涵蓋之範圍內。However, the above descriptions are only to illustrate the preferred embodiments of the present invention, and should not limit the scope of implementation of the present invention. , shall still fall within the scope covered by the patent of the present invention.

無none

圖1是習知呼吸訊號、呼吸註記、心電圖與心跳間隔時間圖例之示意圖。圖2是本發明一較佳實施例之流程圖。圖3是該較佳實施例之基於卷積神經網路的深度學習模型示意圖。圖4是該較佳實施例之使用全卷積神經網路的呼吸暫停與呼吸不足偵測事件之範例示意圖。 FIG. 1 is a schematic diagram of a conventional example of breathing signals, breathing annotations, electrocardiograms and heartbeat intervals. FIG. 2 is a flow chart of a preferred embodiment of the present invention. FIG. 3 is a schematic diagram of a deep learning model based on a convolutional neural network according to the preferred embodiment. FIG. 4 is an exemplary schematic diagram of apnea and hypopnea detection events using a fully convolutional neural network according to the preferred embodiment.

Claims

一種使用全卷積神經網路之呼吸暫停與不足事件偵測方法，其包含有：一量測步驟，其備具有一心電圖機，該心電圖機可針對心臟律動及呼吸頻率感應進行量測，以針對受測者的胸部之心臟自發性跳動與周圍肌肉節律性收縮，且依心臟組織電壓變化記錄而形成一心電圖形訊號；以及一偵測辨識步驟，其備具有一運算處理模組，使該運算處理模組以一全卷積神經網路技術的深度學習模型進行訓練與測試，同時該全卷積神經網路技術的深度學習模型為使用記錄有各自獨立且選自不同的受測者之心電圖形訊號的訓練資料集及測試資料集為進行偵測辨識的資料，且該全卷積神經網路技術的深度學習模型包括由具有至少八個結構相同的特徵提取層，一與該八個特徵提取層連接之第一個卷積層，一與該第一卷積層連接之第二個卷積層，以及及一與該第二卷積層連接第三個卷積層等計算模式所組成，而前述該等特徵提取層對該量測步驟輸入之該心電圖形訊號進行正規化處理，以及對該訊號執行特徵提取與獲得較佳的多個心電圖形訊號特徵圖，而該第一個卷積層得以1×11的運算模式針對該等心電圖形訊號特徵圖中的特徵向量進行計算機率，以至少取得2000個特徵圖輸出，該第二個卷積層得以1×1的運算模式針為該第一卷積層所取得的該等特徵圖中的特徵向量進行計算機率，並至少取得1000個特徵圖輸出，至於該第三個卷積層得以1×1的運算模式針對該第二卷積層所取得的該等特徵圖的特徵向量再進行計算機率，至少取得2個特徵圖輸出，以最終輸出一偵測辨識結果，藉以偵測出該量測步驟所得的心電圖形訊號是否有呼吸暫停與呼吸不足事件態樣。 A method for detecting apnea and insufficiency events using a fully convolutional neural network, which includes: a measuring step, which is equipped with an electrocardiograph, which can measure the heart rhythm and respiratory rate induction, so as to Aiming at the spontaneous beating of the heart of the subject's chest and the rhythmic contraction of the surrounding muscles, and recording the voltage changes of the heart tissue to form an electrocardiographic signal; and a detection and identification step, which is equipped with an arithmetic processing module, so that the The operation processing module is trained and tested with a deep learning model of a fully convolutional neural network technology, and the deep learning model of the fully convolutional neural network technology is used and recorded independently and selected from different subjects. The training data set and the test data set of the electrocardiogram signal are data for detection and identification, and the deep learning model of the fully convolutional neural network technology includes at least eight feature extraction layers with the same structure, one and the eight The feature extraction layer is composed of a first convolutional layer connected to the first convolutional layer, a second convolutional layer connected to the first convolutional layer, and a third convolutional layer connected to the second convolutional layer. The equal feature extraction layer normalizes the ECG signal input in the measurement step, and performs feature extraction on the signal to obtain a plurality of better ECG signal feature maps, and the first convolution layer can be 1× The operation mode of 11 performs computer calculations on the feature vectors in the feature maps of the ECG signals to obtain at least 2000 feature map outputs. The obtained feature vectors in the feature maps are calculated, and at least 1000 feature map outputs are obtained. As for the third convolutional layer, the feature maps obtained by the second convolutional layer can be operated in a 1×1 operation mode. The eigenvectors are then computerized to obtain at least two feature map outputs, so as to finally output a detection and identification result, so as to detect whether the ECG signal obtained in the measurement step has apnea and hypopnea events.

根據請求項1所述使用全卷積神經網路之呼吸暫停與不足事件偵測方法，其中，該每一個特徵提取層包括有一個卷積層、一個批次標準化層、一個激活層、一個最大池化層及一個捨棄層，而前述該卷積層為一個至少可取得45個1D特徵圖的設置，最大池化層為一個池化大小為2的設置，而該捨棄層為一具有50%捨棄率的設置。The method for detecting apnea and insufficiency events using a fully convolutional neural network according to claim 1, wherein each feature extraction layer includes a convolution layer, a batch normalization layer, an activation layer, and a max pooling layer. The aforementioned convolutional layer is a setting that can obtain at least 45 1D feature maps, the max pooling layer is a setting with a pooling size of 2, and the discarding layer is a setting with a 50% discard rate setting.