TWI783343B - A channel information processing system for identifying neonatal epileptic seizures - Google Patents
A channel information processing system for identifying neonatal epileptic seizures Download PDFInfo
- Publication number
- TWI783343B TWI783343B TW110100444A TW110100444A TWI783343B TW I783343 B TWI783343 B TW I783343B TW 110100444 A TW110100444 A TW 110100444A TW 110100444 A TW110100444 A TW 110100444A TW I783343 B TWI783343 B TW I783343B
- Authority
- TW
- Taiwan
- Prior art keywords
- information processing
- data
- neonatal
- epilepsy
- channel
- Prior art date
Links
Images
Landscapes
- Nitrogen And Oxygen Or Sulfur-Condensed Heterocyclic Ring Systems (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
Description
本發明係有關於資訊系統及方法,更詳而言之,係有關於一種應用於新生兒合併有高危險腦病變的情形下使用人工智慧輔助連續性腦電圖監測的環境的辨識新生兒癲癇發作之通道資訊處理系統,針對多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,根據經訊號處理動作後的該些多筆訓練資料,訓練深度卷積神經網路CNN將產生出辨識新生兒癲癇模型,利用所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。 The present invention relates to an information system and method, and more specifically, relates to an environment identification of neonatal epilepsy applied to a newborn with high-risk brain lesions using artificial intelligence-assisted continuous EEG monitoring The seizure channel information processing system performs signal processing for each physiological signal corresponding to each channel of multiple monitoring infant brain wave signal data to generate a deep convolutional neural network CNN (deep Convolutional neural networks) multiple training data, according to the multiple training data after the signal processing action, training the deep convolutional neural network CNN will generate a newborn epilepsy model, and use the generated newborn epilepsy to identify A model that can discriminate the brainwave data of infants and young children and use it to identify whether newborns have epileptic seizures.
新生兒(newborn)抽搐(seizures)是腦部不正常放電(discharge)所導致,而新生兒比成人易發生抽搐,是因為腦部的抑制神經尚未成熟的緣故。新生兒發生抽搐的主要原因,以新生兒於出生過程生產不順所造成的缺氧性腦傷(Hypoxic Encephalopathy)為最大宗,約占一半以上;其他還有中樞神經感染、電解質失衡與糖分過低及其他代謝性疾患等因素。國外大型研究發現,每1,000名兒童約有1.5~5.5例,發生率不低,且多於出生後1週內發生。新生兒抽搐的臨床症狀,比起成人、甚至較大兒童都不典型,且年紀越小表現越不典型,抽搐多是以局部性(partial seizures)為主,必須倚靠家長的良好觀察及醫師的警覺性,配合錄影腦波(video EEG)提供佐證。 Newborn seizures are caused by abnormal brain discharge, and newborns are more prone to seizures than adults because the inhibitory nerves in the brain are not yet mature. The main cause of convulsions in newborns is Hypoxic Encephalopathy (Hypoxic Encephalopathy), which is caused by unsatisfactory delivery during the birth process, accounting for more than half; Other metabolic diseases and other factors. Large-scale foreign studies have found that there are about 1.5 to 5.5 cases per 1,000 children, and the incidence is not low, and more than one week after birth. The clinical symptoms of convulsions in newborns are not typical compared with adults or even older children, and the younger they are, the more atypical the symptoms are. Most of the convulsions are partial seizures, which must rely on the good observation of parents and the doctor's advice Alertness, with video EEG to provide evidence.
腦電圖(EEG,electroencephalography)圖譜可用以偵測癲癇的發生,偵測腦部疾病,在新生兒尤其重要。目前連續性腦電圖(CEEG,continuous EEG)監測在新生兒合併有高危險腦病變的情形下使用的頻率越來越高,並提供了大腦功能狀態和腦電圖上癲癇發作的評估。在新生兒合併有高危險腦病變的情形下,大多數病童可能有腦電圖上癲癇發作,但並沒有臨床的症狀,所以沒有連續性腦電圖監測並無法發現。除此之外,連續性腦電圖監測上的發現往 往會影響臨床的處理流程,例如發現有腦電圖癲癇發作和腦電圖癲癇持續狀態時,則需使用藥物治療。而在多個國外兒童醫學中心的研究報告發現約10-40%新生兒合併有高危險腦病變的情形下曾經歷連續腦電圖監測上的腦電圖癲癇發作。此外,大約三分之一的腦電圖癲癇發作會進展到腦電圖癲癇持續狀態。根據國外統計,平均要花費7小時的腦波檢查才能找到癲癇的證據。而現在也有越來越多的數據顯示,腦電圖發作和腦電圖癲癇持續狀態會造成幼兒腦部發育遲緩等嚴重後果,與神經學預後有很大的關係。且據國外統計,有20-30%的小病童在兒童期會有癲癇的情形。 Electroencephalogram (EEG, electroencephalography) patterns can be used to detect the occurrence of epilepsy and detect brain diseases, especially important in newborns. Continuous electroencephalogram (CEEG, continuous EEG) monitoring is increasingly used in neonates with high-risk encephalopathy and provides assessment of brain functional status and seizures on EEG. In the case of newborns with high-risk brain lesions, most children may have seizures on the EEG, but they have no clinical symptoms, so they cannot be detected without continuous EEG monitoring. In addition, findings on continuous EEG monitoring are often It often affects the clinical treatment process. For example, when EEG seizures and EEG status epilepticus are found, drug treatment is required. However, research reports from several foreign children's medical centers have found that about 10-40% of newborns with high-risk brain lesions have experienced EEG seizures on continuous EEG monitoring. In addition, approximately one-third of EEG seizures progress to EEG status epilepticus. According to foreign statistics, it takes an average of 7 hours of brain wave examination to find evidence of epilepsy. Now there are more and more data showing that EEG seizures and EEG status epilepticus can cause serious consequences such as brain development delay in young children, and have a great relationship with neurological prognosis. And according to foreign statistics, 20-30% of children with minor illnesses will have epilepsy in childhood.
而目前在台灣,新生兒合併有高危險腦病變的情形下已逐漸常規使用連續腦電圖監測。傳統上一次的觀察是狀況往往長達24至48小時以上(long-term)並且需要EEG連續紀錄,因此產生自動化偵測的需求。 At present, in Taiwan, continuous EEG monitoring has gradually been routinely used in neonates with high-risk encephalopathy. The traditional observation is that the condition is often more than 24 to 48 hours (long-term) and requires continuous EEG recording, so there is a need for automatic detection.
所以如何能解決,目前傳統上一次的觀察是狀況往往長達24至48小時以上(long-term)並且需要EEG連續紀錄,而無法利用自動化來進行偵測;如何能利用人工智慧以多筆監測嬰幼兒腦電波訊號數據,利用深度學習演算法經由訓練深度卷積神經網路CNN(Deep Convolutional Neural Network)方式而產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況;如何能對於各種新生兒癲癇發作的辨識正確率高,且辨識時間快速,可節省人工判讀的時間及人力成本;在此,以上種種所述,均是待解決的問題。 So how can it be solved? The current traditional observation is that the situation is often as long as 24 to 48 hours (long-term) and requires continuous EEG recording, which cannot be detected by automation; how can we use artificial intelligence to monitor with multiple records? Infant brain wave signal data, using deep learning algorithm to generate recognition model of neonatal epilepsy by training deep convolutional neural network CNN (Deep Convolutional Neural Network), used to diagnose neonatal epileptic seizure symptoms, identification for more than 10 seconds The state of continuous epilepsy, and to distinguish between non-epilepsy and epilepsy; how to identify various neonatal epileptic seizures with a high accuracy rate, and the identification time is fast, which can save the time and labor costs of manual interpretation; here, All of the above are problems to be solved.
本發明之主要目的便是在於提供一種辨識新生兒癲癇發作之通道資訊處理系統,係應用於新生兒合併有高危險腦病變的情形下使用人工智慧輔助連續性腦電圖監測的環境中,利用本發明之辨識新生兒癲癇發作之通道資訊處理系統時,首先,進行監測嬰幼兒腦電波訊號接收動作,藉由使用一連續性腦電圖(continuous electroencephalography)偵測並獲取嬰幼兒腦電波訊號數據;接著,進行訊號處理動作,將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼 兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;進而,進行辨識新生兒癲癇模型產生動作,根據經訊號處理動作後的該些多筆訓練資料,訓練深度卷積神經網路CNN將產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。利用本發明之辨識新生兒癲癇發作之通道資訊處理系統資訊處理系統所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。 The main purpose of the present invention is to provide a channel information processing system for identifying neonatal epileptic seizures, which is applied in the environment where artificial intelligence is used to assist continuous EEG monitoring in the case of newborns with high-risk brain lesions. In the channel information processing system for identifying neonatal epileptic seizures in the present invention, firstly, monitor the brain wave signal receiving action of infants, and use a continuous electroencephalography (continuous electroencephalography) to detect and obtain brain wave signal data of infants ; Then, the signal processing action will be performed for each physiological signal corresponding to each channel of the multiple monitoring infant brain wave signal data, so as to generate a deep convolutional neural network for training artificial intelligence Multiple training data of CNN (deep convolutional neural networks), wherein, each data of each training data in these multiple training data is corresponding to these multiple monitoring infants and young children Each piece of infant brain wave signal data monitors each physiological signal of each channel in the infant brain wave signal data. Here, all channels are included in the training data, and the training data is marked manually by professional doctors; Identify the neonatal epilepsy model to generate actions. According to the multiple training data after the signal processing actions, the deep convolutional neural network CNN will be trained to generate a neonatal epilepsy model for diagnosing neonatal epilepsy symptoms. Identify 10 The state of continuous epileptic seizures for more than 2 seconds, and distinguish between non-epileptic occurrence and epileptic occurrence. Using the neonatal epilepsy identification model generated by the information processing system of the channel information processing system for identifying neonatal epileptic seizures of the present invention, it is possible to distinguish the brain wave data of infants and thereby identify whether the newborn has epileptic seizures.
本發明之再一目的便是在於提供一種辨識新生兒癲癇發作之通道資訊處理系統,係應用於新生兒合併有高危險腦病變的情形下使用人工智慧輔助連續性腦電圖監測的環境中,能解決目前傳統上一次的觀察是狀況往往長達24至48小時以上(long-term)並且需要EEG連續紀錄,而無法利用自動化來進行偵測;能利用人工智慧以多筆監測嬰幼兒腦電波訊號數據,利用深度學習演算法經由訓練深度卷積神經網路CNN(Deep Convolutional Neural Network)方式而產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況;能對於各種新生兒癲癇發作的辨識正確率高,且辨識時間快速,可節省人工判讀的時間及人力成本。 Another object of the present invention is to provide a channel information processing system for identifying epileptic seizures in newborns, which is applied in the environment where artificial intelligence is used to assist continuous EEG monitoring in the case of newborns with high-risk brain lesions. It can solve the problem that the traditional one-time observation is often as long as 24 to 48 hours (long-term) and requires continuous EEG recording, and cannot be detected by automation; it can use artificial intelligence to monitor the brain waves of infants and young children with multiple records Signal data, using deep learning algorithms to generate a model for identifying neonatal epilepsy by training a deep convolutional neural network CNN (Deep Convolutional Neural Network), which is used to diagnose the symptoms of neonatal epilepsy and identify epileptic seizures that occur continuously for more than 10 seconds It can distinguish non-epileptic seizures from epileptic seizures; it can identify various neonatal epileptic seizures with a high accuracy rate, and the identification time is fast, which can save the time and labor costs of manual interpretation.
根據以上所述之目的,本發明提供一種辨識新生兒癲癇發作之通道資訊處理系統,該通道資訊處理系統包含資訊處理模組、人工智慧深度卷積神經網路CNN模組、以及資料庫。 According to the above-mentioned purpose, the present invention provides a channel information processing system for identifying neonatal epileptic seizures. The channel information processing system includes an information processing module, an artificial intelligence deep convolutional neural network (CNN) module, and a database.
資訊處理模組,該資訊處理模組將接收多筆與訓練人工智慧卷積神經網路CNN相關的監測嬰幼兒腦電波訊號,藉由使用一連續性腦電圖偵測並獲取嬰幼兒腦電波訊號數據;該資訊處理模組將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;以及,該資訊處理模組將經訊號處理動作後的該些多筆訓練資料傳送至人工智慧深度卷積神經網路CNN模組。 Information processing module, the information processing module will receive multiple monitoring infant brain wave signals related to training artificial intelligence convolution neural network CNN, by using a continuous EEG to detect and obtain infant brain wave Signal data; the information processing module will perform signal processing on each physiological signal corresponding to each channel of these multiple monitoring infant brain wave signal data to generate a deep convolutional neural network for training artificial intelligence Multiple training data of CNN (deep convolutional neural networks), wherein each of the multiple training data corresponds to each of the multiple monitoring infant brain wave signal data Monitor each physiological signal of each channel in the brainwave signal data of infants and young children. Here, all channels are included in the training data, and the training data is marked manually by professional doctors; and the information processing module will be processed by the signal The multiple training data after the action are sent to the artificial intelligence deep convolutional neural network (CNN) module.
在此,嬰幼兒腦電波訊號數據的資料型態: Here, the data type of infant brain wave signal data:
1)類比電位訊號經接收之後,放大(amplifier),轉換成類比訊號。 1) After the analog potential signal is received, it is amplified (amplifier) and converted into an analog signal.
2)取樣頻率125Hz。 2) The sampling frequency is 125Hz.
3)對於每個通道,Y軸單位為micro-volt,X軸為時間。 3) For each channel, the Y-axis unit is micro-volt, and the X-axis is time.
4)每個通道的訊號值會減去不同的參考通道(reference channel),以雙蕉範式(double banana montage)為參考圖譜。 4) Different reference channels are subtracted from the signal value of each channel, and the double banana montage is used as the reference spectrum.
5)每個通道對應於不同腦區,可作為臨床腦區異常判斷依據。 5) Each channel corresponds to a different brain region, which can be used as a basis for judging abnormalities in clinical brain regions.
6)列出所有channel編號數目名稱。以11位置(location)(Fp1,C3,O1,T3,Fp2,C4,O2,T4,Fz,Cz,Pz);另,以安裝(Montage)(Fp1-C3,C3-O2,Fp2-C4,C4-O2,Fp1-T3,T3-O1,Fp2-T4,T4-O2,T3-C3,C3-Cz,Cz-C4,C4-T4,Fz-Cz,Cz-Pz)。 6) List all channel numbers and names. With 11 locations (Fp1, C3, O1, T3, Fp2, C4, O2, T4, Fz, Cz, Pz); in addition, with Montage (Fp1-C3, C3-O2, Fp2-C4, C4-O2, Fp1-T3, T3-O1, Fp2-T4, T4-O2, T3-C3, C3-Cz, Cz-C4, C4-T4, Fz-Cz, Cz-Pz).
7)嬰幼兒的癲癇波多界於0.5-15Hz之間,週期(periodic)或非週期性(non-periodic),波形特徵隨時間緩慢變化,有些呈現高頻棘波(ictal or preictal spikes)。 7) The epilepsy waves of infants and young children are mostly between 0.5-15Hz, periodic (periodic) or non-periodic (non-periodic), and the waveform characteristics change slowly with time, and some present high-frequency spikes (ictal or preictal spikes).
在此,專業醫師人工標註訓練資料時,進行資料處理: Here, when professional doctors manually label training data, data processing is performed:
1)人工標註嬰幼兒(早產兒(24周至37周)和足月新生兒(37周至48周)EEG上癲癇發生的時間區域,有癲癇發生起始點和終點。 1) Manually mark the time zone of epilepsy on the EEG of infants (premature infants (24 weeks to 37 weeks) and full-term neonates (37 weeks to 48 weeks), with the start point and end point of epilepsy.
2)以10秒完單位切割片段(segments)。 2) Cut the segments in units of 10 seconds.
3)每個片段有2-8秒長度和前一個片段重疊。 3) Each segment overlaps the previous segment by 2-8 seconds in length.
4)去除片段大於±500 micro-volt的過大值,高機率為雜訊。 4) Remove the excessive value of the fragment greater than ±500 micro-volt, the high probability is noise.
5)資料”不”正規化(normalization,取所有通道的最大值當作1最小值當作0,縮放資料於0-1之間)。 5) The data is "not" normalized (normalization, take the maximum value of all channels as 1 and the minimum value as 0, and scale the data between 0-1).
6)每個片段都有獨立標記為癲癇或非癲癇,片段未充滿10秒癲癇狀態皆列為正常,反之充滿10秒則列為癲癇狀態。 6) Each segment is independently marked as epileptic or non-epileptic. If the segment is not full for 10 seconds, the epileptic state is classified as normal, and if it is full for 10 seconds, it is classified as epileptic state.
7)全部通道皆納入模型訓練,訓練時隨機剔除1-2個通道,以零值取代增加名行的耐受度(tolerance)。 7) All channels are included in the model training, and 1-2 channels are randomly eliminated during training, and zero values are used instead to increase the tolerance (tolerance).
8)資料會隨機沿時間軸平移,對資料做增集(Augmentation)。 8) The data will be randomly shifted along the time axis, and the data will be augmented (Augmentation).
9)隨機亂序排列通道對資料做增集,通道間時間仍保持同步(Synchronization)關係。 9) Randomly arrange the channels in random order to augment the data, and the time between the channels still maintains the synchronization (Synchronization) relationship.
10)資料會上傳至資料中心。 10) The data will be uploaded to the data center.
11)運算處理於運算中心。 11) Computing is processed in the computing center.
人工智慧深度卷積神經網路CNN模組,該人工智慧深度卷積神經網路CNN模組使用深度學習演算法,將根據經資訊處理模組的訊號處理動作後的該些多筆訓練資料,進行辨識新生兒癲癇模型產生動作,訓練深度卷積神經網路CNN產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。利用本發明之辨識新生兒癲癇發作之通道資訊處理系統資訊處理系統所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。 Artificial intelligence deep convolutional neural network CNN module, the artificial intelligence deep convolutional neural network CNN module uses a deep learning algorithm to process the multiple training data according to the signal processing action of the information processing module, Identify the neonatal epilepsy model to generate actions, and train the deep convolutional neural network CNN to generate a neonatal epilepsy model for diagnosing the symptoms of neonatal epilepsy, identifying the state of continuous epilepsy for more than 10 seconds, and distinguishing non-epileptic occurrence with epilepsy. Using the neonatal epilepsy identification model generated by the information processing system of the channel information processing system for identifying neonatal epileptic seizures of the present invention, it is possible to distinguish the brain wave data of infants and thereby identify whether the newborn has epileptic seizures.
在此,採用深度學習演算法的人工智慧深度卷積神經網路CNN模組,可將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;對於此深度學習模型,可使用三種方式選取通道:第一、選擇模型內參數較高的通道作為重要的通道;第二、訓練單一個通道作辨識任務,選取最辨識準確性者;以及,第三、只剔除一個通道(leave-one out)訓練模型,若通道有重要貢獻則辨識效率將降低。 Here, the artificial intelligence deep convolutional neural network (CNN) module using deep learning algorithms can incorporate all channels into the training data, and the training data is marked manually by professional doctors; for this deep learning model, three methods can be used Channel selection: first, select the channel with higher parameters in the model as an important channel; second, train a single channel for identification tasks, and select the one with the highest identification accuracy; and, third, only remove one channel (leave-one out) training model, if the channel has an important contribution, the identification efficiency will decrease.
以人工智慧深度卷積神經網路CNN模組使用深度學習演算法而產生出辨識新生兒癲癇模型的模型方法而言: In terms of the model method for identifying neonatal epilepsy models using artificial intelligence deep convolutional neural network CNN modules using deep learning algorithms:
1)CNN:使用多層卷積層(convolution layers)組成一個密集層(Dense Block),許多密集層可以藉由轉移層(Transition Block)連接,最後經過線性層(Linear Block)輸出,Softmax運算輸出。 1) CNN: Multi-layer convolution layers are used to form a dense layer (Dense Block). Many dense layers can be connected by a transition layer (Transition Block), and finally output through a linear layer (Linear Block), and Softmax operation output.
2)資料向前傳遞經過各層可逐漸萃取重要特徵,於密集層時特徵會萃取重要特徵,這些特徵會於轉移層疊加(concatenate),此疊加效果較一般傳統的CNN會保留上游特徵。 2) The forward transfer of data through each layer can gradually extract important features. In the dense layer, the features will extract important features, and these features will be concatenated in the transfer layer. This superposition effect is better than that of traditional CNNs, which will retain upstream features.
3)每次訓練結果會藉由向後傳遞(Back-propagation)更新參數,藉此修正錯誤辨識的參數。 3) Each training result will update the parameters through Back-propagation, thereby correcting the wrongly identified parameters.
4)CNN層為單通道特徵萃取,訓練或辨識時會對所有一個以上的通道做特徵萃取。模型最後輸輸出會經過注意力層(Attention)將權重從新分配,增強通道之間與時間序列前後的關聯性。 4) The CNN layer is a single-channel feature extraction, and feature extraction will be performed on more than one channel during training or identification. The final input and output of the model will redistribute the weights through the attention layer (Attention) to enhance the correlation between channels and time series.
5)以正確率(Accuracy),操作特征曲線(ROC),曲線下面積(AUC),F1 Scores,敏感性Sensitivity,特異性(Specificity)為模型衡量標準。 5) Accuracy, Operating Characteristic Curve (ROC), Area Under the Curve (AUC), F1 Scores, Sensitivity, and Specificity are used as model metrics.
以辨識新生兒癲癇模型的估測模型之運作而言: In terms of the operation of the estimation model to identify neonatal epilepsy models:
模型作用: Model role:
1)可判別經典數種不同的嬰幼兒癲癇型態,是一種分類模型(Classification)。 1) It is a classification model (Classification) that can distinguish several different types of epilepsy in infants and young children.
2)預測(forecasting):提前10~30秒預測癲癇即將發生。 2) Forecasting: 10-30 seconds in advance to predict the impending epilepsy.
3)人機交互作用界面,手動標註之癲癇區皆可被納入新的癲癇訓練集,即時回饋修改模型參數。 3) Human-computer interaction interface, manually marked epilepsy areas can be included in the new epilepsy training set, and the model parameters can be modified in real time.
4)可達成及時偵測,紀錄後偵測。 4) It can achieve real-time detection and detection after recording.
5)配合警報系統可設置長期無人自動提醒。 5) With the alarm system, automatic reminders can be set for long-term unmanned.
6)模型可調整敏感性sensitivity升高或降低需求。 6) The model can adjust sensitivity to increase or decrease demand.
7)模型參數少佈署容易。 7) The model parameters are few and easy to deploy.
資料庫,該資料庫配合資訊處理模組、人工智慧深度卷積神經網路CNN模組共同運作,可供資訊處理模組、人工智慧深度卷積神經網路CNN模組存取所需的資料/數據。 The database, which works together with the information processing module and the artificial intelligence deep convolutional neural network CNN module, can be used by the information processing module and the artificial intelligence deep convolutional neural network CNN module to access the required data /data.
利用本發明之辨識新生兒癲癇發作之通道資訊處理系統時,首先,進行監測嬰幼兒腦電波訊號接收動作,藉由使用一連續性腦電圖(continuous electroencephalography)偵測並獲取嬰幼兒腦電波訊號數據。 When using the channel information processing system for identifying neonatal epileptic seizures of the present invention, first, monitor the brainwave signal reception of infants and young children, and use a continuous electroencephalography (continuous electroencephalography) to detect and obtain brainwave signals of infants and young children data.
接著,進行訊號處理動作,將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註。 Then, the signal processing operation will be performed on each physiological signal corresponding to each channel of these multiple monitoring infant brain wave signal data to generate a deep convolutional neural network for training artificial intelligence Multiple training data of CNN (deep convolutional neural networks), wherein each of the multiple training data corresponds to each monitoring of the multiple monitoring infant brain wave signal data For each physiological signal of each channel in the infant brain wave signal data, all channels are included in the training data, and the training data is marked manually by professional doctors.
進而,進行辨識新生兒癲癇模型產生動作,根據經訊號處理動作後的該些多筆訓練資料,訓練深度卷積神經網路CNN將產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。 Furthermore, the neonatal epilepsy model is identified to generate actions. According to the multiple training data after the signal processing actions, the deep convolutional neural network CNN will be trained to generate a neonatal epilepsy model for diagnosing neonatal epilepsy symptoms. , to identify the state of continuous epileptic seizures for more than 10 seconds, and to distinguish between non-epileptogenesis and epilepsy.
利用本發明之辨識新生兒癲癇發作之通道資訊處理系統資訊處理系統所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。 Using the neonatal epilepsy identification model generated by the information processing system of the channel information processing system for identifying neonatal epileptic seizures of the present invention, it is possible to distinguish the brain wave data of infants and thereby identify whether the newborn has epileptic seizures.
為使熟悉該項技藝人士瞭解本發明之目的、特徵及功效,茲藉由下述具體實施例,並配合所附之圖式,對本發明詳加說明如後: In order to make those familiar with the art understand the purpose, characteristics and effects of the present invention, the present invention is described in detail as follows by the following specific embodiments and in conjunction with the accompanying drawings:
1:辨識新生兒癲癇發作之通道資訊處理系統 1: Channel information processing system for identifying neonatal epileptic seizures
2:資訊處理模組 2: Information processing module
3:人工智慧深度卷積神經網路CNN模組 3: Artificial intelligence deep convolutional neural network CNN module
4:資料庫 4: Database
5:電子裝置 5: Electronic device
11:密集模組 11: Dense module
12:轉移模組 12: Transfer Module
13:轉移模組 13: Transfer Module
14:線性層 14: Linear layer
15:Softmax函數 15: Softmax function
101 102 103:步驟 101 102 103: Steps
201 202 203:步驟 201 202 203: Steps
1001 1002 1003:步驟 1001 1002 1003: steps
第1圖為一系統示意圖,用以顯示說明本發明之通道資訊處理系統之系統架構、以及運作情形;第2圖為一流程圖,用以顯示說明利用如第1圖中之本發明之辨識新生兒癲癇發作之通道資訊處理系統以進行辨識新生兒癲癇發作之通道資訊處理方法的流程步驟;第3圖為一示意圖,用以顯示說明本發明之辨識新生兒癲癇發作之通道資訊處理系統的一實施例、以及運作情形;第4圖為一示意圖,用以顯示說明於第3圖中之11位置(location)、以及安裝(Montage)的情形;第5圖為一示意圖,用以顯示說明於第3圖中的實施例的人工智慧深度卷積神經網路CNN模組之CNN模型訓練方式及組成;第6圖為一流程圖,用以顯示說明利用於第3圖中的實施例的辨識新生兒癲癇模型,偵測新生兒癲癇發作的流程;第7圖為一示意圖,用以顯示說明於第6圖中的進行記錄動作時,顯示正常腦波、較不規律的情況;第8圖為一示意圖,用以顯示說明於第6圖中的進行記錄動作時,顯示癲癇腦波、規律性波形的情況;第9圖為一示意圖,用以顯示說明於第3圖中的一CNN模型確效性及訓練資料集之表現;第10圖為一示意圖,用以顯示說明於第3圖中的另一CNN模型確效性及訓練資料集之表現; 第11圖為一示意圖,用以顯示說明利用第3圖中的實施例的辨識新生兒癲癇發作之通道資訊處理系統所產生出的辨識新生兒癲癇模型,判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作的情況;以及第12圖為一流程圖,用以顯示說明利用如第3圖中之本發明之辨識新生兒癲癇發作之通道資訊處理系統以進行辨識新生兒癲癇發作之通道資訊處理方法的一流程步驟。 Figure 1 is a schematic diagram of a system for illustrating the system architecture and operation of the channel information processing system of the present invention; Figure 2 is a flow chart for illustrating the identification of the present invention using the method in Figure 1 The channel information processing system for neonatal epileptic seizures is used to perform the flow steps of the channel information processing method for identifying neonatal epileptic seizures; Figure 3 is a schematic diagram for illustrating the channel information processing system for identifying neonatal epileptic seizures of the present invention One embodiment, and operation situation; The 4th figure is a schematic diagram, in order to show the situation of 11 positions (location) and installation (Montage) described in the 3rd figure; The 5th figure is a schematic diagram, in order to show explanation The CNN model training method and composition of the artificial intelligence deep convolutional neural network CNN module of the embodiment in the 3rd figure; the 6th figure is a flow chart, in order to show and explain and utilize in the embodiment in the 3rd figure The process of identifying neonatal epilepsy models and detecting neonatal epileptic seizures; Figure 7 is a schematic diagram to show the normal and irregular brain waves during the recording action described in Figure 6; Figure 8 The picture is a schematic diagram for displaying epileptic brain waves and regular waveforms during the recording action described in Figure 6; Figure 9 is a schematic diagram for displaying a CNN described in Figure 3 The validity of the model and the performance of the training data set; Figure 10 is a schematic diagram for showing the validity of another CNN model and the performance of the training data set explained in Figure 3; Fig. 11 is a schematic diagram for illustrating the neonatal epilepsy identification model generated by using the channel information processing system for identifying neonatal epileptic seizures in the embodiment in Fig. 3 to identify infant brain wave data and identify Whether the newborn has epileptic seizures; and Fig. 12 is a flow chart for illustrating the process of identifying neonatal epileptic seizures by using the channel information processing system for identifying neonatal epileptic seizures of the present invention as shown in Fig. 3 A process step of the channel information processing method.
第1圖為一系統示意圖,用以顯示說明本發明之通道資訊處理系統之系統架構、以及運作情形。如第1圖中所示之,辨識新生兒癲癇發作之通道資訊處理系統1包含資訊處理模組2、人工智慧深度卷積神經網路CNN模組3、以及資料庫4。
Fig. 1 is a schematic diagram of the system, which is used to illustrate the system architecture and operation of the channel information processing system of the present invention. As shown in FIG. 1 , the channel
資訊處理模組2,該資訊處理模組2將接收多筆與訓練人工智慧卷積神經網路CNN相關的監測嬰幼兒腦電波訊號,藉由使用一連續性腦電圖偵測並獲取嬰幼兒腦電波訊號數據;該資訊處理模組2將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;以及,該資訊處理模組2將經訊號處理動作後的該些多筆訓練資料傳送至人工智慧深度卷積神經網路CNN模組3。
在此,嬰幼兒腦電波訊號數據的資料型態: Here, the data type of infant brain wave signal data:
1)類比電位訊號經接收之後,放大(amplifier),轉換成類比訊號。 1) After the analog potential signal is received, it is amplified (amplifier) and converted into an analog signal.
2)取樣頻率125Hz。 2) The sampling frequency is 125Hz.
3)對於每個通道,Y軸單位為micro-volt,X軸為時間。 3) For each channel, the Y-axis unit is micro-volt, and the X-axis is time.
4)每個通道的訊號值會減去不同的參考通道(reference channel),以雙蕉範式(double banana montage)為參考圖譜。 4) Different reference channels are subtracted from the signal value of each channel, and the double banana montage is used as the reference spectrum.
5)每個通道對應於不同腦區,可作為臨床腦區異常判斷依據。 5) Each channel corresponds to a different brain region, which can be used as a basis for judging abnormalities in clinical brain regions.
6)列出所有channel編號數目名稱。以11位置(location)(Fp1,C3,O1,T3,Fp2,C4,O2,T4,Fz,Cz,Pz);另,以安裝(Montage)(Fp1-C3,C3-O2,Fp2-C4,C4-O2,Fp1-T3,T3-O1,Fp2-T4,T4-O2,T3-C3,C3-Cz,Cz-C4,C4-T4,Fz-Cz,Cz-Pz)。 6) List all channel numbers and names. With 11 locations (Fp1, C3, O1, T3, Fp2, C4, O2, T4, Fz, Cz, Pz); in addition, with Montage (Fp1-C3, C3-O2, Fp2-C4, C4-O2, Fp1-T3, T3-O1, Fp2-T4, T4-O2, T3-C3, C3-Cz, Cz-C4, C4-T4, Fz-Cz, Cz-Pz).
7)嬰幼兒的癲癇波多界於0.5-15Hz之間,週期(periodic)或非週期性(non-periodic),波形特徵隨時間緩慢變化,有些呈現高頻棘波(ictal or preictal spikes)。 7) The epilepsy waves of infants and young children are mostly between 0.5-15Hz, periodic (periodic) or non-periodic (non-periodic), and the waveform characteristics change slowly with time, and some present high-frequency spikes (ictal or preictal spikes).
在此,專業醫師人工標註訓練資料時,進行資料處理: Here, when professional doctors manually label training data, data processing is performed:
1)人工標註嬰幼兒(早產兒(24周至37周)和足月新生兒(37周至48周)EEG上癲癇發生的時間區域,有癲癇發生起始點和終點。 1) Manually mark the time zone of epilepsy on the EEG of infants (premature infants (24 weeks to 37 weeks) and full-term neonates (37 weeks to 48 weeks), with the start point and end point of epilepsy.
2)以10秒完單位切割片段(segments)。 2) Cut the segments in units of 10 seconds.
3)每個片段有2-8秒長度和前一個片段重疊。 3) Each segment overlaps the previous segment by 2-8 seconds in length.
4)去除片段大於±500 micro-volt的過大值,高機率為雜訊。 4) Remove the excessive value of the fragment greater than ±500 micro-volt, the high probability is noise.
5)資料”不”正規化(normalization,取所有通道的最大值當作1最小值當作0,縮放資料於0-1之間)。 5) The data is "not" normalized (normalization, take the maximum value of all channels as 1 and the minimum value as 0, and scale the data between 0-1).
6)每個片段都有獨立標記為癲癇或非癲癇,片段未充滿10秒癲癇狀態皆列為正常,反之充滿10秒則列為癲癇狀態。 6) Each segment is independently marked as epileptic or non-epileptic. If the segment is not full for 10 seconds, the epileptic state is classified as normal, and if it is full for 10 seconds, it is classified as epileptic state.
7)全部通道皆納入模型訓練,訓練時隨機剔除1-2個通道,以零值取代增加名行的耐受度(tolerance)。 7) All channels are included in the model training, and 1-2 channels are randomly eliminated during training, and zero values are used instead to increase the tolerance (tolerance).
8)資料會隨機沿時間軸平移,對資料做增集(Augmentation)。 8) The data will be randomly shifted along the time axis, and the data will be augmented (Augmentation).
9)隨機亂序排列通道對資料做增集,通道間時間仍保持同步(Synchronization)關係。 9) Randomly arrange the channels in random order to augment the data, and the time between the channels still maintains the synchronization (Synchronization) relationship.
10)資料會上傳至資料中心。 10) The data will be uploaded to the data center.
11)運算處理於運算中心。 11) Computing is processed in the computing center.
人工智慧深度卷積神經網路CNN模組3,該人工智慧深度卷積神經網路CNN模組3使用深度學習演算法,將根據經資訊處理模組2的訊號處理動作後的該些多筆訓練資料,進行辨識新生兒癲癇模型產生動作,訓練深度卷積神經網路CNN產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。
Artificial intelligence deep convolutional neural
利用本發明之辨識新生兒癲癇發作之通道資訊處理系統1及其方法所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。
Using the neonatal epilepsy identification model generated by the channel
在此,採用深度學習演算法的人工智慧深度卷積神經網路CNN模組3,可將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;對於此深度學習模型,可使用三種方式選取通道:第一、選擇模型內參數較高的通道作為重要的通道;第二、訓練單一個通道作辨識任務,選取最辨識準確性者;以及,第三、只剔除一個通道(leave-one out)訓練模型,若通道有重要貢獻則辨識效率將降低。
Here, the artificial intelligence deep convolutional neural
以人工智慧深度卷積神經網路CNN模組3使用深度學習演算法而產生出辨識新生兒癲癇模型的模型方法而言:
In terms of the model method for identifying neonatal epilepsy models using artificial intelligence deep convolutional neural
1)CNN:使用多層卷積層(convolution layers)組成一個密集層(Dense Block),許多密集層可以藉由轉移層(Transition Block)連接,最後經過線性層(Linear Block)輸出,Softmax運算輸出。 1) CNN: Multi-layer convolution layers are used to form a dense layer (Dense Block). Many dense layers can be connected by a transition layer (Transition Block), and finally output through a linear layer (Linear Block), and Softmax operation output.
2)資料向前傳遞經過各層可逐漸萃取重要特徵,於密集層時特徵會萃取重要特徵,這些特徵會於轉移層疊加(concatenate),此疊加效果較一般傳統的CNN會保留上游特徵。 2) The forward transfer of data through each layer can gradually extract important features. In the dense layer, the features will extract important features, and these features will be concatenated in the transfer layer. This superposition effect is better than that of traditional CNNs, which will retain upstream features.
3)每次訓練結果會藉由向後傳遞(Back-propagation)更新參數,藉此修正錯誤辨識的參數。 3) Each training result will update the parameters through Back-propagation, thereby correcting the wrongly identified parameters.
4)CNN層為單通道特徵萃取,訓練或辨識時會對所有一個以上的通道做特徵萃取。模型最後輸輸出會經過注意力層(Attention)將權重從新分配,增強通道之間與時間序列前後的關聯性。 4) The CNN layer is a single-channel feature extraction, and feature extraction will be performed on more than one channel during training or identification. The final input and output of the model will redistribute the weights through the attention layer (Attention) to enhance the correlation between channels and time series.
5)以正確率(Accuracy),操作特征曲線(ROC),曲線下面積(AUC),F1 Scores,敏感性Sensitivity,特異性(Specificity)為模型衡量標準。 5) Accuracy, Operating Characteristic Curve (ROC), Area Under the Curve (AUC), F1 Scores, Sensitivity, and Specificity are used as model metrics.
以辨識新生兒癲癇模型的估測模型之運作而言: In terms of the operation of the estimation model to identify neonatal epilepsy models:
模型作用: Model role:
1)可判別經典數種不同的嬰幼兒癲癇型態,是一種分類模型(Classification)。 1) It is a classification model (Classification) that can distinguish several different types of epilepsy in infants and young children.
2)預測(forecasting):提前10~30秒預測癲癇即將發生。 2) Forecasting: 10-30 seconds in advance to predict the impending epilepsy.
3)人機交互作用界面,手動標註之癲癇區皆可被納入新的癲癇訓練集,即時回饋修改模型參數。 3) Human-computer interaction interface, manually marked epilepsy areas can be included in the new epilepsy training set, and the model parameters can be modified in real time.
4)可達成及時偵測,紀錄後偵測。 4) It can achieve real-time detection and detection after recording.
5)配合警報系統可設置長期無人自動提醒。 5) With the alarm system, automatic reminders can be set for long-term unmanned.
6)模型可調整敏感性sensitivity升高或降低需求。 6) The model can adjust sensitivity to increase or decrease demand.
7)模型參數少佈署容易。 7) The model parameters are few and easy to deploy.
資料庫4,該資料庫4配合資訊處理模組2、人工智慧深度卷積神經網路CNN模組3共同運作,可供資訊處理模組2、人工智慧深度卷積神經網路CNN模組3存取所需的資料/數據。
The
視實施狀況,資訊處理模組2及/或人工智慧深度卷積神經網路CNN模組3,係由電子硬體、韌體、以及軟體的至少其中之一所組成,配合辨識新生兒癲癇發作之通道資訊處理系統1所在之系統/裝置的處理器(未圖示之)而進行動作;而資料庫4則位於辨識新生兒癲癇發作之通道資訊處理系統1所在之系統/裝置的儲存模組(未圖示之)。
Depending on the implementation status, the
第2圖為一流程圖,用以顯示說明利用如第1圖中之本發明之辨識新生兒癲癇發作之通道資訊處理系統以進行辨識新生兒癲癇發作之通道資訊處理方法的流程步驟。如第2圖中所示之,首先,於步驟101,進行監測嬰幼兒腦電波訊號接收動作,藉由使用一連續性腦電圖(continuous electroencephalography)偵測並獲取嬰幼兒腦電波訊號數據,並進到步驟102。
FIG. 2 is a flow chart for illustrating the process steps of the channel information processing method for identifying neonatal epileptic seizures by using the channel information processing system for identifying neonatal epileptic seizures of the present invention as shown in FIG. 1 . As shown in Figure 2, firstly, in
在此,資訊處理模組2將接收多筆與訓練人工智慧卷積神經網路CNN相關的監測嬰幼兒腦電波訊號,藉由使用一連續性腦電圖偵測並獲取嬰幼兒腦電波訊號數據。
Here, the
於步驟102,進行訊號處理動作,將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註,並進到步驟103。
In
在此,資訊處理模組2將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;以及,該資訊處理模組2將經訊號處理動作後的該些多筆訓練資料傳送至人工智慧深度卷積神經網路CNN模組3。
Here, the
於步驟103,進行辨識新生兒癲癇模型產生動作,根據經訊號處理動作後的該些多筆訓練資料,訓練深度卷積神經網路CNN將產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。
In
在此,人工智慧深度卷積神經網路CNN模組3使用深度學習演算法,將根據經資訊處理模組2的訊號處理動作後的該些多筆訓練資料,進行辨識新生兒癲癇模型產生動作,訓練深度卷積神經網路CNN產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。
Here, the artificial intelligence deep convolutional neural
利用本發明之辨識新生兒癲癇發作之通道資訊處理系統1及其方法所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。
Using the neonatal epilepsy identification model generated by the channel
在此,採用深度學習演算法的人工智慧深度卷積神經網路CNN模組3,可將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;對於此深度學習模型,可使用三種方式選取通道:第一、選擇模型內參數較高的通道作為重要的通道;第二、訓練單一個通道作辨識任務,選取最辨識準確性者;以及,第三、只剔除一個通道(leave-one out)訓練模型,若通道有重要貢獻則辨識效率將降低。
Here, the artificial intelligence deep convolutional neural
第3圖為一示意圖,用以顯示說明本發明之辨識新生兒癲癇發作之通道資訊處理系統的一實施例、以及運作情形。如第3圖中所示之,辨識新生兒癲癇發作之通道資訊處理系統1包含資訊處理模組2、人工智慧深度卷積神經網路CNN模組3、以及資料庫4,其中,辨識新生兒癲癇發作之通道資訊處理系統1
係位於,例如,醫院運算中心/資料儲存中心的電子裝置5中,電子裝置5可為,例如,攜帶型、桌上型、伺服器型系統,為資訊系統的電子裝置5負責串流腦波資料,進行模型演算,儲存資料,匯出報告;其中,資訊處理模組2及/或人工智慧深度卷積神經網路CNN模組3,係由電子硬體、韌體、以及軟體的至少其中之一所組成,配合辨識新生兒癲癇發作之通道資訊處理系統1所在之電子裝置5的處理器而進行動作,而資料庫4則位於辨識新生兒癲癇發作之通道資訊處理系統1所在之電子裝置5的儲存模組。
FIG. 3 is a schematic diagram for illustrating an embodiment of the channel information processing system for identifying neonatal epileptic seizures and its operation according to the present invention. As shown in Figure 3, the channel
資訊處理模組2,該資訊處理模組2將接收多筆與訓練人工智慧卷積神經網路CNN相關的監測嬰幼兒腦電波訊號,藉由使用一連續性腦電圖偵測並獲取嬰幼兒腦電波訊號數據;該資訊處理模組2將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;以及,該資訊處理模組2將經訊號處理動作後的該些多筆訓練資料傳送至人工智慧深度卷積神經網路CNN模組3。
又,在此,例如,嬰幼兒腦電波訊號數據的資料型態: Also, here, for example, the data type of infant brain wave signal data:
1)類比電位訊號經接收之後,放大(amplifier),轉換成類比訊號。 1) After the analog potential signal is received, it is amplified (amplifier) and converted into an analog signal.
2)取樣頻率125Hz。 2) The sampling frequency is 125Hz.
3)對於每個通道,Y軸單位為micro-volt,X軸為時間。 3) For each channel, the Y-axis unit is micro-volt, and the X-axis is time.
4)每個通道的訊號值會減去不同的參考通道(reference channel),以雙蕉範式(double banana montage)為參考圖譜。 4) Different reference channels are subtracted from the signal value of each channel, and the double banana montage is used as the reference spectrum.
5)每個通道對應於不同腦區,可作為臨床腦區異常判斷依據。 5) Each channel corresponds to a different brain region, which can be used as a basis for judging abnormalities in clinical brain regions.
6)列出所有channel編號數目名稱。在此,第4圖為一示意圖,用以顯示說明於第3圖中之11位置(location)、以及安裝(Montage)的情形。如第4圖中所示之,11位置(location):Fp1,C3,O1,T3,Fp2,C4,O2,T4,Fz,Cz,Pz;而安裝(Montage):Fp1-C3,C3-O2,Fp2-C4,C4-O2,Fp1-T3,T3-O1,Fp2-T4,T4-O2,T3-C3,C3-Cz,Cz-C4,C4-T4,Fz-Cz,Cz-Pz。 6) List all channel numbers and names. Here, FIG. 4 is a schematic diagram for showing the 11 positions (location) and the installation (Montage) described in FIG. 3 . As shown in Figure 4, 11 locations (location): Fp1, C3, O1, T3, Fp2, C4, O2, T4, Fz, Cz, Pz; and installation (Montage): Fp1-C3, C3-O2 , Fp2-C4, C4-O2, Fp1-T3, T3-O1, Fp2-T4, T4-O2, T3-C3, C3-Cz, Cz-C4, C4-T4, Fz-Cz, Cz-Pz.
7)嬰幼兒的癲癇波多界於0.5-15Hz之間,週期(periodic)或非週期性(non-periodic),波形特徵隨時間緩慢變化,有些呈現高頻棘波(ictal or preictal spikes)。 7) The epilepsy waves of infants and young children are mostly between 0.5-15Hz, periodic (periodic) or non-periodic (non-periodic), and the waveform characteristics change slowly with time, and some present high-frequency spikes (ictal or preictal spikes).
再,在此,例如,嬰幼兒腦電波訊號數據的資料型態包括,但不限於:(1)類比電位訊號經接收之後,放大(amplifier),轉換成類比訊號;(2)取樣頻率125Hz;(3)對於每個通道,Y軸單位為微伏特,X軸為時間;(4)每個通道的訊號值會減去不同的參考通道(reference channel),以雙蕉範式(double banana montage)為參考圖譜;(5)每個通道對應於不同腦區,可作為臨床腦區異常判斷依據;(6)嬰幼兒的癲癇波多界於0.5~15Hz之間,週期(periodic)或非週期性(non-periodic),波形特徵隨時間緩慢變化,有些呈現癲癇發作時或癲癇發作前的高頻棘波(ictal or pre-ictal spikes)。 Furthermore, here, for example, the data types of brain wave signal data of infants include, but are not limited to: (1) After receiving the analog potential signal, it is amplified (amplifier) and converted into an analog signal; (2) The sampling frequency is 125Hz; (3) For each channel, the Y-axis unit is microvolts, and the X-axis is time; (4) The signal value of each channel will be subtracted from a different reference channel (reference channel), in double banana montage (5) Each channel corresponds to a different brain region, which can be used as a basis for judging abnormalities in clinical brain regions; (6) The epileptic waves of infants and young children range from 0.5 to 15 Hz, periodic (periodic) or aperiodic ( non-periodic), the waveform characteristics change slowly over time, and some show high-frequency spikes (ictal or pre-ictal spikes) during or before epileptic seizures.
在本實施例中,資料來源為林口長庚醫院小兒加護病房,取樣對象為早產兒(24周至37周)及足月新生兒(37周至48周)。 In this example, the source of data is the Pediatric Intensive Care Unit of Linkou Chang Gung Hospital, and the sampling objects are premature infants (24 weeks to 37 weeks) and full-term newborns (37 weeks to 48 weeks).
本發明於實際施行時,嬰幼兒腦電波訊號數據的資料型態並非限於本實施例中所述之,其他嬰幼兒腦電波訊號數據的資料型態,其理相同、類似於本實施例中所述,是故,在此不再贅述。 When the present invention is actually implemented, the data type of the brain wave signal data of infants is not limited to the data type described in this embodiment, and the data type of other brain wave signal data of infants is the same, similar to that described in this embodiment. Therefore, I will not repeat them here.
在本實施例中,人工智慧深度卷積神經網路CNN模組3的CNN輸入為EEG生理訊號,對資料結構來說,本發明稱每個生理訊號來源叫做通道(channel)。每一個通道為單維度資料,以電壓(voltages)(電壓是機器收取訊號的方式)為記錄單位,隨時間(time)累積資料量。每個通道為單維度資料(1通道×時間)。因為本發明有13通道,CNN輸入為13通道×時間資料(視為13個單維度資料)。本發明的CNN即是用這樣的資料作為訓練。每一筆資料會由訓練過的醫師使用軟體呈現後對其圖形做判讀,定義出病徵時段。本發明的CNN是訓練13個電壓型態的單維度資料而非技師用肉眼看到的波形,雖然在CNN與人腦判斷方式相似,但資料形式不同。
In this embodiment, the CNN input of the artificial intelligence deep convolutional neural
在此,專業醫師人工標註訓練資料時,進行資料處理: Here, when professional doctors manually label training data, data processing is performed:
1)人工標註嬰幼兒(早產兒(24周至37周)和足月新生兒(37周至48周)EEG上癲癇發生的時間區域,有癲癇發生起始點和終點。 1) Manually mark the time zone of epilepsy on the EEG of infants (premature infants (24 weeks to 37 weeks) and full-term neonates (37 weeks to 48 weeks), with the start point and end point of epilepsy.
2)以10秒完單位切割片段(segments)。 2) Cut the segments in units of 10 seconds.
3)每個片段有2-8秒長度和前一個片段重疊。 3) Each segment overlaps the previous segment by 2-8 seconds in length.
4)去除片段大於±500 micro-volt的過大值,高機率為雜訊。 4) Remove the excessive value of the fragment greater than ±500 micro-volt, the high probability is noise.
5)資料”不”正規化(normalization,取所有通道的最大值當作1最小值當作0,縮放資料於0-1之間)。 5) The data is "not" normalized (normalization, take the maximum value of all channels as 1 and the minimum value as 0, and scale the data between 0-1).
6)每個片段都有獨立標記為癲癇或非癲癇,片段未充滿10秒癲癇狀態皆列為正常,反之充滿10秒則列為癲癇狀態。 6) Each segment is independently marked as epileptic or non-epileptic. If the segment is not full for 10 seconds, the epileptic state is classified as normal, and if it is full for 10 seconds, it is classified as epileptic state.
7)全部通道皆納入模型訓練,訓練時隨機剔除1-2個通道,以零值取代增加名行的耐受度(tolerance)。 7) All channels are included in the model training, and 1-2 channels are randomly eliminated during training, and zero values are used instead to increase the tolerance (tolerance).
8)資料會隨機沿時間軸平移,對資料做增集(Augmentation)。 8) The data will be randomly shifted along the time axis, and the data will be augmented (Augmentation).
9)隨機亂序排列通道對資料做增集,通道間時間仍保持同步(Synchronization)關係。 9) Randomly arrange the channels in random order to augment the data, and the time between the channels still maintains the synchronization (Synchronization) relationship.
10)資料會上傳至資料中心。 10) The data will be uploaded to the data center.
11)運算處理於運算中心。 11) Computing is processed in the computing center.
人工智慧深度卷積神經網路CNN模組3,該人工智慧深度卷積神經網路CNN模組3使用深度學習演算法,將根據經資訊處理模組2的訊號處理動作後的該些多筆訓練資料,進行辨識新生兒癲癇模型產生動作,訓練深度卷積神經網路CNN產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。
Artificial intelligence deep convolutional neural
舉例而言,包括卷積神經網絡的人工智能模型之辨識新生兒癲癇模型包括,但不限於:(1)可判別經典數種不同的嬰幼兒癲癇型態,是一種分類模型(Classification model);(2)人機交互作用界面,手動標註之癲癇區皆可被納入新的癲癇訓練集,即時回饋修改模型之特徵參數;(3)可達成及時偵測,紀錄後偵測;(4)配合自動警報系統可設置長期無人自動提醒;(4)模型可藉由警報系統調整敏感性升高或降低需求,並產生一臨床推薦報告;(5)模型參數少佈署容易。 For example, the neonatal epilepsy identification model of the artificial intelligence model including the convolutional neural network includes, but is not limited to: (1) It can distinguish several classic types of infantile epilepsy, which is a classification model (Classification model); (2) Human-computer interaction interface, manually marked epilepsy areas can be included in the new epilepsy training set, and the characteristic parameters of the model can be modified in real time; (3) Real-time detection can be achieved, and detection after recording; (4) Cooperate The automatic alarm system can be set to automatically remind long-term unattended users; (4) The model can be adjusted to increase sensitivity or reduce demand through the alarm system, and generate a clinical recommendation report; (5) The model has few parameters and is easy to deploy.
舉例而言,資料於運算中心進行運算處理時除了進行模型辨識之外,亦可產生個人化模型,進而產生泛化模型,泛化模型與運算中心會隨時進行訓練更新。 For example, when the data is processed in the computing center, in addition to model identification, a personalized model can also be generated, and then a generalized model can be generated. The generalized model and the computing center will be trained and updated at any time.
在本實施例中,每次訓練結果會藉由一向後傳遞(back-propagation)更新特徵參數,藉此修正錯誤判別的特徵參數。 In this embodiment, the feature parameters are updated through a back-propagation for each training result, so as to correct incorrectly identified feature parameters.
第4圖為一示意圖,用以顯示說明於第3圖中之11位置(location)、以及安裝(Montage)的情形。如第4圖中所示之,11位置(location):Fp1,C3,O1,T3,Fp2,C4,O2,T4,Fz,Cz,Pz;而安裝(Montage):Fp1-C3,C3-O2,Fp2-C4,C4-O2,Fp1-T3,T3-O1,Fp2-T4,T4-O2,T3-C3,C3-Cz,Cz-C4,C4-T4,Fz-Cz,Cz-Pz。 Figure 4 is a schematic diagram for illustrating the 11 locations (location) and installation (Montage) in Figure 3. As shown in Figure 4, 11 locations (location): Fp1, C3, O1, T3, Fp2, C4, O2, T4, Fz, Cz, Pz; and installation (Montage): Fp1-C3, C3-O2 , Fp2-C4, C4-O2, Fp1-T3, T3-O1, Fp2-T4, T4-O2, T3-C3, C3-Cz, Cz-C4, C4-T4, Fz-Cz, Cz-Pz.
在此,採用深度學習演算法的人工智慧深度卷積神經網路CNN模組3,可將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;對於此深度學習模型,可使用三種方式選取通道:第一、選擇模型內參數較高的通道作為重要的通道;第二、訓練單一個通道作辨識任務,選取最辨識準確性者;以及,第三、只剔除一個通道(leave-one out)訓練模型,若通道有重要貢獻則辨識效率將降低。
Here, the artificial intelligence deep convolutional neural
以人工智慧深度卷積神經網路CNN模組3使用深度學習演算法而產生出辨識新生兒癲癇模型的模型方法而言:
In terms of the model method for identifying neonatal epilepsy models using artificial intelligence deep convolutional neural
1)CNN:使用多層卷積層(convolution layers)組成一個密集層(Dense Block),許多密集層可以藉由轉移層(Transition Block)連接,最後經過線性層(Linear Block)輸出,Softmax運算輸出。 1) CNN: Multi-layer convolution layers are used to form a dense layer (Dense Block). Many dense layers can be connected by a transition layer (Transition Block), and finally output through a linear layer (Linear Block), and Softmax operation output.
2)資料向前傳遞經過各層可逐漸萃取重要特徵,於密集層時特徵會萃取重要特徵,這些特徵會於轉移層疊加(concatenate),此疊加效果較一般傳統的CNN會保留上游特徵。 2) The forward transfer of data through each layer can gradually extract important features. In the dense layer, the features will extract important features, and these features will be concatenated in the transfer layer. This superposition effect is better than that of traditional CNNs, which will retain upstream features.
3)每次訓練結果會藉由向後傳遞(Back-propagation)更新參數,藉此修正錯誤辨識的參數。 3) Each training result will update the parameters through Back-propagation, thereby correcting the wrongly identified parameters.
4)CNN層為單通道特徵萃取,訓練或辨識時會對所有一個以上的通道做特徵萃取。模型最後輸輸出會經過注意力層(Attention)將權重從新分配,增強通道之間與時間序列前後的關聯性。 4) The CNN layer is a single-channel feature extraction, and feature extraction will be performed on more than one channel during training or identification. The final input and output of the model will redistribute the weights through the attention layer (Attention) to enhance the correlation between channels and time series.
5)以正確率(Accuracy),操作特征曲線(ROC),曲線下面積(AUC),F1 Scores,敏感性Sensitivity,特異性(Specificity)為模型衡量標準。 5) Accuracy, Operating Characteristic Curve (ROC), Area Under the Curve (AUC), F1 Scores, Sensitivity, and Specificity are used as model metrics.
以辨識新生兒癲癇模型的估測模型之運作而言: In terms of the operation of the estimation model to identify neonatal epilepsy models:
模型作用: Model role:
1)可判別經典數種不同的嬰幼兒癲癇型態,是一種分類模型(Classification)。 1) It is a classification model (Classification) that can distinguish several different types of epilepsy in infants and young children.
2)預測(forecasting):提前10~30秒預測癲癇即將發生。 2) Forecasting: 10-30 seconds in advance to predict the impending epilepsy.
3)人機交互作用界面,手動標註之癲癇區皆可被納入新的癲癇訓練集,即時回饋修改模型參數。 3) Human-computer interaction interface, manually marked epilepsy areas can be included in the new epilepsy training set, and the model parameters can be modified in real time.
4)可達成及時偵測,紀錄後偵測。 4) It can achieve real-time detection and detection after recording.
5)配合警報系統可設置長期無人自動提醒。 5) With the alarm system, automatic reminders can be set for long-term unmanned.
6)模型可調整敏感性sensitivity升高或降低需求。 6) The model can adjust sensitivity to increase or decrease demand.
7)模型參數少佈署容易。 7) The model parameters are few and easy to deploy.
資料庫4,該資料庫4配合資訊處理模組2、人工智慧深度卷積神經網路CNN模組3共同運作,可供資訊處理模組2、人工智慧深度卷積神經網路CNN模組3存取所需的資料/數據。
The
在本實施例中,人工智慧深度卷積神經網路CNN模組3的CNN輸入為EEG生理訊號,對資料結構來說,本發明稱每個生理訊號來源叫做通道(channel)。每一個通道為單維度資料,以電壓(voltages)(電壓是機器收取訊號的方式)為記錄單位,隨時間(time)累積資料量。每個通道為單維度資料(1通道×時間)。因為本發明有13通道,CNN輸入為13通道×時間資料(視為13個單維度資料)。本發明的CNN即是用這樣的資料作為訓練。每一筆資料會由訓練過的醫師使用軟體呈現後對其圖形做判讀,定義出病徵時段。本發明的CNN是訓練13個電壓型態的單維度資料而非技師用肉眼看到的波形,雖然在CNN與人腦判斷方式相似,但資料形式不同。
In this embodiment, the CNN input of the artificial intelligence deep convolutional neural
第5圖為一示意圖,用以顯示說明於第3圖中的實施例的人工智慧深度卷積神經網路CNN模組之CNN模型訓練方式及組成。 FIG. 5 is a schematic diagram for illustrating the CNN model training method and composition of the artificial intelligence deep convolutional neural network CNN module of the embodiment illustrated in FIG. 3 .
如第5圖中所示之,CNN模型訓練方式及組成為由密集模組(dense module)11、轉移模組(translation module)12、轉移模組(translation module)13、線性層(linear block)14、以及Softmax函數(Softmax regression)15所組成,而輸出為資料分類(Classification)。 As shown in Figure 5, the training method and composition of the CNN model consists of a dense module (dense module) 11, a transfer module (translation module) 12, a transfer module (translation module) 13, and a linear layer (linear block) 14, and the Softmax function (Softmax regression) 15, and the output is data classification (Classification).
密集模組11包含二次之批量標準化(Batch Normalization)、線性整流函式ReLU(Rectified Linear Unit)、以及卷積層(Convolution);以及,轉移模組12、以及轉移模組13分別包含批量標準化、線性整流函式ReLU、卷積層、以及池化(polling)層。
例如,批量標準化算法使得深層神經網絡訓練更加穩定,加快了收斂的速度;而線性整流函式ReLU函數和它的導數計算簡單,在向前傳遞和向後傳遞時都減少了計算量,由於在時函數的導數值為1,可以在一定程度上解決梯度消失問題,訓練時有更快的收斂速度;卷積層可有n個m*p的卷積核,n,m,p為整數,例如,作用於灰度圖像,每個卷積核作用於前一層輸出圖像的部分通道上,產生多張的輸出圖像;以及,池化層作用於卷積層的輸出圖像,執行q*r的池化,q,r為整數,產生多張的輸出圖像。 For example, the batch normalization algorithm makes the deep neural network training more stable and speeds up the convergence speed; while the linear rectification function ReLU function and its derivatives are easy to calculate, and the amount of calculation is reduced when passing forward and backward. The derivative value of the function is 1, which can solve the gradient disappearance problem to a certain extent, and has faster convergence speed during training; the convolution layer can have n m*p convolution kernels, n, m, and p are integers, for example, Acting on grayscale images, each convolution kernel acts on some channels of the output image of the previous layer to generate multiple output images; and, the pooling layer acts on the output image of the convolution layer, performing q*r The pooling, q, r are integers, generate multiple output images.
以人工智慧深度卷積神經網路CNN模組3使用深度學習演算法而產生出估測模型的CNN模型訓練方法而言,如第5圖中所示之,以13個通道的各別生理訊號當成輸入資料:
In terms of the CNN model training method using artificial intelligence deep convolutional neural
(1)使用多層卷積層(convolution layers)組成一個密集層(Dense Block),許多密集層可以藉由轉移層(Transition Block)連接,最後經過線性層(Linear Block)輸出,Softmax函數(Softmax regression)運算而輸出。 (1) Use multiple convolution layers to form a dense layer (Dense Block), many dense layers can be connected by a transition layer (Transition Block), and finally output through a linear layer (Linear Block), Softmax function (Softmax regression) operation and output.
(2)資料向前傳遞(Forward-propagation)經過各層可逐漸萃取重要特徵,於密集層時特徵會萃取重要特徵,這些特徵會於轉移層疊加(concatenate),此疊加效果較一般傳統的CNN會保留上游特徵。 (2) Forward-propagation of the data can gradually extract important features through each layer. In the dense layer, the features will extract important features. These features will be superimposed on the transfer layer (concatenate). This superposition effect is better than that of traditional CNN. Preserve the upstream features.
(3)每次訓練結果會藉由向後傳遞(Back-propagation)更新參數,藉此修正錯誤辨識的參數。 (3) Each training result will update the parameters by back-propagation, thereby correcting the wrongly identified parameters.
(4)CNN層為單通道特徵萃取,訓練或辨識時會對所有一個以上的通道做特徵萃取。估測模型之模型最後輸出會經過注意力層(Attention)將權重重新分配,增強通道之間與時間序列前後的關聯性。 (4) The CNN layer is a single-channel feature extraction, and feature extraction will be performed on all or more channels during training or identification. The final output of the estimation model will redistribute the weights through the attention layer (Attention) to enhance the correlation between channels and time series.
(5)以正確率(Accuracy),操作特征曲線(ROC),曲線下面積(AUC),F1 Scores,敏感性Sensitivity,特異性(Specificity)為模型衡量標準。 (5) The accuracy rate (Accuracy), operating characteristic curve (ROC), area under the curve (AUC), F1 Scores, Sensitivity, and Specificity are used as model measurement standards.
利用本發明之辨識新生兒癲癇發作之通道資訊處理系統1及其方法所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。
Using the neonatal epilepsy identification model generated by the channel
第6圖為一流程圖,用以顯示說明利用於第3圖中的實施例的辨識新生兒癲癇模型,偵測新生兒癲癇發作的流程。如第6圖中所示之,首先,於步驟1001,進行自動偵測動作;由電子裝置5的資訊系統進行無人長期自動偵測(亦
即不需要醫師待命在旁),然後進行判讀與警示,並送至ICU中辨識癲癇發作新生兒,並進到步驟1002。
FIG. 6 is a flow chart for illustrating the process of detecting neonatal epileptic seizures by utilizing the neonatal epilepsy model of the embodiment in FIG. 3 . As shown in Figure 6, at first, in
於步驟1002,進行記錄動作;進行13通道(包括Fp1-T3、T3-O1、Fp2-T4、T4-O2、Fp1-C3、C3-O1、Fp2-C4、C4-O2、T3-C3、C3-Cz、Cz-C4、C4-T4及EKG,參閱第7圖與第8圖)腦波記錄,然後由醫師手動判讀癲癇發生,並進到步驟1003。
In
於步驟1003,進行資料蒐集或校準動作;在資料蒐集或校準之後,進行CNN模型學習並傳送至電子裝置5的資訊系統。
In
第7圖為一示意圖,用以顯示說明於第6圖中的進行記錄動作時,顯示正常腦波、較不規律的情況。 Fig. 7 is a schematic diagram for illustrating normal and irregular brain waves during the recording action described in Fig. 6 .
第8圖為一示意圖,用以顯示說明於第6圖中的進行記錄動作時,顯示癲癇腦波、規律性波形的情況。 Fig. 8 is a schematic diagram for illustrating the situation of displaying epileptic brain waves and regular waveforms during the recording action in Fig. 6 .
第9圖為一示意圖,用以顯示說明於第3圖中的一CNN模型確效性及訓練資料集之表現。如第9圖中所示之,本發明用於辨識新生兒癲癇發作的辨識新生兒癲癇模型,其顯示一CNN模型確效性及訓練資料集之表現。 FIG. 9 is a schematic diagram for illustrating the validity of a CNN model illustrated in FIG. 3 and the performance of the training data set. As shown in Fig. 9, the neonatal epilepsy model for identifying neonatal epileptic seizures according to the present invention shows the validity of a CNN model and the performance of the training data set.
第10圖為一示意圖,用以顯示說明於第3圖中的另一CNN模型確效性及訓練資料集之表現。如第10圖中所示之,本發明用於辨識新生兒癲癇發作的辨識新生兒癲癇模型,其顯示另一CNN模型確效性及訓練資料集之表現。 Fig. 10 is a schematic diagram to show the validity of another CNN model and the performance of the training data set illustrated in Fig. 3 . As shown in Fig. 10, the present invention is used to identify neonatal epilepsy model for identifying neonatal epilepsy, which shows the validity of another CNN model and the performance of the training data set.
第11圖為一示意圖,用以顯示說明利用第3圖中的實施例的辨識新生兒癲癇發作之通道資訊處理系統所產生出的辨識新生兒癲癇模型,判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作的情況。如第11圖中所示之,顯示癲癇新生兒病患與癲癇腦波出現示意圖。 Fig. 11 is a schematic diagram for illustrating the neonatal epilepsy identification model generated by using the channel information processing system for identifying neonatal epileptic seizures in the embodiment in Fig. 3 to identify infant brain wave data and identify Whether the newborn has epileptic seizures. As shown in Figure 11, it shows a schematic diagram of epileptic neonatal patients and epileptic brainwaves.
第12圖為一流程圖,用以顯示說明利用如第3圖中之本發明之辨識新生兒癲癇發作之通道資訊處理系統以進行辨識新生兒癲癇發作之通道資訊處理方法的一流程步驟。如第12圖中所示之,首先,於步驟201,進行監測嬰幼兒腦電波訊號接收動作,藉由使用一連續性腦電圖(continuous electroencephalography)偵測並獲取嬰幼兒腦電波訊號數據,並進到步驟202。
FIG. 12 is a flow chart for illustrating a process step of a channel information processing method for identifying neonatal epileptic seizures by using the channel information processing system for identifying neonatal epileptic seizures as shown in FIG. 3 of the present invention. As shown in Fig. 12, firstly, in
在此,資訊處理模組2將接收多筆與訓練人工智慧卷積神經網路CNN相關的監測嬰幼兒腦電波訊號,藉由使用一連續性腦電圖偵測並獲取嬰幼兒腦電波訊號數據。
Here, the
於步驟202,進行訊號處理動作,將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註,並進到步驟203。
In
在此,資訊處理模組2將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN(deep convolutional neural networks)的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;以及,該資訊處理模組2將經訊號處理動作後的該些多筆訓練資料傳送至人工智慧深度卷積神經網路CNN模組3。
Here, the
於步驟203,進行辨識新生兒癲癇模型產生動作,根據經訊號處理動作後的該些多筆訓練資料,訓練深度卷積神經網路CNN將產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。
In
在此,人工智慧深度卷積神經網路CNN模組3使用深度學習演算法,將根據經資訊處理模組2的訊號處理動作後的該些多筆訓練資料,進行辨識新生兒癲癇模型產生動作,訓練深度卷積神經網路CNN產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。
Here, the artificial intelligence deep convolutional neural
利用本發明之辨識新生兒癲癇發作之通道資訊處理系統1及其方法所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。
Using the neonatal epilepsy identification model generated by the channel
在此,採用深度學習演算法的人工智慧深度卷積神經網路CNN模組3,可將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;對於此深度學習模型,可使用三種方式選取通道:第一、選擇模型內參數較高的通道
作為重要的通道;第二、訓練單一個通道作辨識任務,選取最辨識準確性者;以及,第三、只剔除一個通道(leave-one out)訓練模型,若通道有重要貢獻則辨識效率將降低。
Here, the artificial intelligence deep convolutional neural
綜合以上之實施例,我們可以得到一種辨識新生兒癲癇發作之通道資訊處理系統,係應用於新生兒合併有高危險腦病變的情形下使用人工智慧輔助連續性腦電圖監測的環境中,利用本發明之辨識新生兒癲癇發作之通道資訊處理系統時,首先,進行監測嬰幼兒腦電波訊號接收動作,藉由使用一連續性腦電圖偵測並獲取嬰幼兒腦電波訊號數據;接著,進行訊號處理動作,將針對該些多筆監測嬰幼兒腦電波訊號數據之對應每一通道的每一生理訊號,進行訊號處理,以產生出用以訓練人工智慧之深度卷積神經網路CNN的多筆訓練資料,其中,該些多筆訓練資料中的每一筆訓練資料的每一項資料係對應於該些多筆監測嬰幼兒腦電波訊號數據之每一筆監測嬰幼兒腦電波訊號數據中之每一通道的每一生理訊號,在此,將所有通道納入訓練資料,而訓練資料之標註為專業醫師人工標註;進而,進行辨識新生兒癲癇模型產生動作,根據經訊號處理動作後的該些多筆訓練資料,訓練深度卷積神經網路CNN將產生出辨識新生兒癲癇模型,用以診斷新生兒癲癇發作症狀,辨識10秒以上連續癲癇出現的狀態,並且區別出非癲癇發生與癲癇發生的情況。利用本發明之辨識新生兒癲癇發作之通道資訊處理系統資訊處理系統所產生出的辨識新生兒癲癇模型,可判別嬰幼兒腦電波數據並藉此辨識新生兒是否為癲癇發作。 Combining the above embodiments, we can obtain a channel information processing system for identifying neonatal epileptic seizures, which is applied in the environment where artificial intelligence is used to assist continuous EEG monitoring in the case of neonates with high-risk brain lesions. In the channel information processing system for identifying neonatal epileptic seizures of the present invention, first, monitor the infant’s brain wave signal receiving action, and use a continuous electroencephalogram to detect and obtain the infant’s brain wave signal data; then, carry out The signal processing action will carry out signal processing on each physiological signal corresponding to each channel of these multiple monitoring infant brain wave signal data, so as to generate multiple data of the deep convolutional neural network CNN used to train artificial intelligence. A piece of training data, wherein, each piece of data in each piece of training data in the multiple pieces of training data corresponds to each piece of data in each piece of monitoring infant brainwave signal data in the multiple pieces of monitoring infant brainwave signal data Each physiological signal of a channel, here, all channels are included in the training data, and the training data is marked manually by professional doctors; and then, identify the actions generated by the neonatal epilepsy model, and according to these multiple actions after signal processing Write training data, train the deep convolutional neural network CNN to generate a model for identifying neonatal epilepsy, which is used to diagnose the symptoms of neonatal epilepsy, identify the state of continuous epilepsy for more than 10 seconds, and distinguish between non-epilepsy and epilepsy Condition. Using the neonatal epilepsy identification model generated by the information processing system of the channel information processing system for identifying neonatal epileptic seizures of the present invention, it is possible to distinguish the brain wave data of infants and thereby identify whether the newborn has epileptic seizures.
以上所述僅為本發明之較佳實施例而已,並非用以限定本發明之範圍;凡其它未脫離本發明所揭示之精神下所完成之等效改變或修飾,均應包含在下述之專利範圍內。 The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention; all other equivalent changes or modifications that do not deviate from the spirit disclosed in the present invention should be included in the following patents within range.
1:辨識新生兒癲癇發作之通道資訊處理系統 1: Channel information processing system for identifying neonatal epileptic seizures
2:資訊處理模組 2: Information processing module
3:人工智慧深度卷積神經網路CNN模組 3: Artificial intelligence deep convolutional neural network CNN module
4:資料庫 4: Database
Claims (5)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW110100444A TWI783343B (en) | 2021-01-06 | 2021-01-06 | A channel information processing system for identifying neonatal epileptic seizures |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW110100444A TWI783343B (en) | 2021-01-06 | 2021-01-06 | A channel information processing system for identifying neonatal epileptic seizures |
Publications (2)
Publication Number | Publication Date |
---|---|
TW202228148A TW202228148A (en) | 2022-07-16 |
TWI783343B true TWI783343B (en) | 2022-11-11 |
Family
ID=83436881
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW110100444A TWI783343B (en) | 2021-01-06 | 2021-01-06 | A channel information processing system for identifying neonatal epileptic seizures |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI783343B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI446297B (en) * | 2007-12-28 | 2014-07-21 | 私立中原大學 | Drowsiness detection system |
CN109671500A (en) * | 2019-02-26 | 2019-04-23 | 上海交通大学 | Schizophrenia auxiliary diagnosis classification method based on electroencephalogram time domain data |
CN111616682A (en) * | 2020-05-31 | 2020-09-04 | 天津大学 | Epileptic seizure early warning system based on portable electroencephalogram acquisition equipment and application |
CN112106074A (en) * | 2018-05-01 | 2020-12-18 | 国际商业机器公司 | Seizure detection and prediction using techniques such as deep learning methods |
-
2021
- 2021-01-06 TW TW110100444A patent/TWI783343B/en active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI446297B (en) * | 2007-12-28 | 2014-07-21 | 私立中原大學 | Drowsiness detection system |
CN112106074A (en) * | 2018-05-01 | 2020-12-18 | 国际商业机器公司 | Seizure detection and prediction using techniques such as deep learning methods |
CN109671500A (en) * | 2019-02-26 | 2019-04-23 | 上海交通大学 | Schizophrenia auxiliary diagnosis classification method based on electroencephalogram time domain data |
CN111616682A (en) * | 2020-05-31 | 2020-09-04 | 天津大学 | Epileptic seizure early warning system based on portable electroencephalogram acquisition equipment and application |
Also Published As
Publication number | Publication date |
---|---|
TW202228148A (en) | 2022-07-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Choi et al. | Real-time apnea-hypopnea event detection during sleep by convolutional neural networks | |
Surrel et al. | Online obstructive sleep apnea detection on medical wearable sensors | |
Venkatesan et al. | Mobile cloud computing for ECG telemonitoring and real-time coronary heart disease risk detection | |
US5775330A (en) | Neurometric assessment of intraoperative anesthetic | |
US20220175324A1 (en) | Computer-based prediction of fetal and maternal outcomes | |
Thompson et al. | Detection of obstructive sleep apnoea using features extracted from segmented time-series ECG signals using a one dimensional convolutional neural network | |
Werth et al. | Unobtrusive assessment of neonatal sleep state based on heart rate variability retrieved from electrocardiography used for regular patient monitoring | |
Altıntop et al. | A novel approach for detection of consciousness level in comatose patients from EEG signals with 1-D convolutional neural network | |
Yadav et al. | Variational mode decomposition-based seizure classification using Bayesian regularized shallow neural network | |
US7231246B2 (en) | Detection of encephalopathy | |
Rajput et al. | Automated detection of hypertension using wavelet transform and nonlinear techniques with ballistocardiogram signals | |
Moghadam et al. | Sleep State Trend (SST), a bedside measure of neonatal sleep state fluctuations based on single EEG channels | |
Zhang et al. | Automatic Sleep Staging using Multi‐dimensional Feature Extraction and Multi‐kernel Fuzzy Support Vector Machine | |
TWI783343B (en) | A channel information processing system for identifying neonatal epileptic seizures | |
Ranta et al. | An open source classifier for bed mattress signal in infant sleep monitoring | |
Chiang et al. | Online incremental learning for sleep quality assessment using associative Petri net | |
Hamlin et al. | Assessing the feasibility of detecting epileptic seizures using non-cerebral sensor data | |
TWI748485B (en) | Information processing system and method | |
Ríos et al. | An automatic apnea screening algorithm for children | |
Satapathy et al. | An automated system for sleep staging using EEG brain signals based on a machine learning approach | |
Green et al. | Sleep-Wake and Body Position Classification with Deep Learning using Pressure Sensor Mat Measurements | |
Locharla et al. | EEG-based deep learning neural net for apnea detection | |
Li et al. | A deep learning model developed for sleep apnea detection: A multi-center study | |
Moridani et al. | Automatic Epileptic Seizure Detection Based on EEG Signals Using Deep Learning | |
Bangani et al. | Personalized Stress Monitoring AI System For Healthcare Workers |