KR20240030369A - System for Monitoring of Intelligent emergency and Driving method thereof - Google Patents

Abstract

The present invention relates to an intelligent emergency monitoring system applying an LSTM-based DRNN model using BIG SIGNAL and a method of driving the same. An information collection unit that collects biometric information signals from at least one biometric information monitoring device, and the information collection unit collects biometric information signals. A pre-processing unit that removes noise and normalizes the biometric information signal, a learning unit that uses the biometric information signal pre-processed in the pre-processing unit to apply it to an artificial intelligence neural network model and learns through machine learning, and when the patient's biometric information signal is received, It includes a situation identification unit that identifies whether an emergency situation has occurred by applying it to an artificial intelligence neural network model based on the learning results learned in the learning unit, wherein the artificial intelligence neural network model of the learning unit is an RNN model including at least one LSTM track. , the intelligent emergency monitoring system can provide an intelligent emergency monitoring system and its driving method through ECG recognition that can produce results faster and with better performance in a shorter sequence than the existing RNN learning model.

Description

Intelligent emergency monitoring system and driving method thereof {System for Monitoring of Intelligent emergency and Driving method thereof}

The present invention relates to an intelligent emergency monitoring system and its driving method. More specifically, it relates to an intelligent emergency monitoring system and its driving method applying an LSTM-based DRNN model using BIG SIGNAL.

In general, according to WHO, cardiovascular diseases (CVDs) are the leading cause of death today. More than 17.7 million people have died from CVD, accounting for approximately 31% of all deaths, and more than 75% of deaths occur in low- and middle-income countries.

Arrhythmia is a representative type of CVD that represents irregular changes in normal heart rhythm. Arrhythmias include atrial fibrillation, premature contraction, ventricular fibrillation, and tachycardia. Although a single arrhythmic heartbeat may not have a serious impact on life, sustained arrhythmic heartbeats can lead to fatal situations. For example, prolonged premature ventricular contractions (PVCs) beats can sometimes transition into ventricular tachycardia (VT) or ventricular tachycardia (VF) beats, which can lead to immediate heart failure.

Therefore, it is important to continuously monitor heart rate to manage and prevent CVDs. ECG is a non-invasive medical tool that displays heart rhythm and status. Therefore, automatically detecting irregular heart rhythms in ECG signals is very important in the field of cardiology.

Therefore, by monitoring the ECG signal, it is possible to determine whether an emergency situation has occurred.

Meanwhile, a recurrent neural network (RNN) model is a neural network structure that uses a single-layer or multi-layer recurrent connection configuration. It is generally applied to learning data such as speech, video, and strings.

This network model has the characteristic of memorizing the state of previous information and applying it to the current input data. The mechanism of the RNN model is particularly strong in learning sequential data.

KR 10-2019-0141326 A

The present invention was derived from this technical background, and its purpose is to provide an intelligent emergency monitoring system and driving method through ECG recognition that can produce results more quickly and with superior performance in a shorter sequence than existing RNN learning models. There is.

The present invention for achieving the above problems includes the following configuration.

That is, the intelligent emergency monitoring system according to an embodiment of the present invention includes an information collection unit that collects biometric information signals from one or more biometric information monitoring devices, and a preprocessor that removes noise and normalizes the biometric information signals collected by the information collection unit. , a learning unit that uses the biometric information signal pre-processed in the preprocessing unit and applies it to an artificial intelligence neural network model to learn through machine learning, and when the patient's biometric information signal is received, an artificial intelligence neural network is created based on the learning results learned in the learning unit. It includes a situation identification unit that is applied to the model to identify whether an emergency situation has occurred, and the artificial intelligence neural network model of the learning unit is an RNN model that includes at least one LSTM track.

Meanwhile, a method of driving an intelligent emergency situation monitoring system according to an embodiment includes an information collection step of collecting biometric information signals from at least one biometric information monitoring device, removing noise and normalizing the biometric information signals collected in the information collection step. A preprocessing step, a learning step of applying machine learning to an artificial intelligence neural network model using the biometric information signal preprocessed in the preprocessing step, and when the patient's biometric information signal is received, based on the learning results learned in the learning step. It includes a situation identification step to identify whether an emergency situation has occurred by applying it to an artificial intelligence neural network model, and the artificial intelligence neural network model in the learning step is an RNN model that includes at least one LSTM track.

According to the present invention, the effect of providing an intelligent emergency situation monitoring system and its operation method through ECG recognition that can produce results more quickly and with excellent performance in a short sequence compared to existing RNN learning models is derived.

1 is an exemplary diagram for explaining the overall operation of an intelligent emergency monitoring system according to an embodiment of the present invention.
Figure 2 is a flow chart showing a bio-signal analysis algorithm in an intelligent emergency monitoring system according to an embodiment of the present invention.
Figure 3 is a block diagram showing the configuration of an intelligent emergency situation monitoring system according to an embodiment of the present invention.
Figure 4 is an exemplary diagram to explain the bio-signal pre-processing process in the pre-processing unit in one embodiment of the present invention.
Figure 5 is an example diagram of a method of grouping input data for a real-time system.
Figure 6 is an example diagram for explaining a recurrent neural network (RNN) model according to an embodiment of the present invention.
Figure 7 is an example diagram illustrating an LSTM-based RNN model in which RNN nodes are replaced with LSTMs according to an embodiment of the present invention.
Figure 8 is an example diagram illustrating a bidirectional LSTM-based RNN model for private label classification.
Figure 9 is an example diagram for explaining an RNN model structure including an LSTM track according to an embodiment of the present invention.
10 and 11 are graphs to explain the accuracy of personal label classification for the emergency monitoring architecture performed in the intelligent emergency monitoring system according to an embodiment.
Figure 12 shows the private label classification accuracy of the bidirectional LSTM-based RNN model of the intelligent emergency monitoring system according to an embodiment.
Figure 13 is a flowchart of a method of driving an intelligent emergency situation monitoring system according to an embodiment of the present invention.

It should be noted that the technical terms used in the present invention are only used to describe specific embodiments and are not intended to limit the present invention. In addition, the technical terms used in the present invention, unless specifically defined in a different sense in the present invention, should be interpreted as meanings generally understood by those skilled in the art in the technical field to which the present invention pertains, and are not overly comprehensive. It should not be interpreted in a literal or excessively reduced sense.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings.

1 is an exemplary diagram for explaining the overall operation of an intelligent emergency monitoring system according to an embodiment of the present invention.

The intelligent emergency monitoring system according to an embodiment of the present invention uses preprocessed biometric information as input to a BIG-based biometric information classification model. And because biometric information includes time information, learning can be performed using an LSTM-based RNN model suitable for learning sequential data.

And we want to learn and test biometric information using a bidirectional LSTM-based RNN learning model. In one embodiment, the goal of learning is to receive preprocessed biometric information of objects A, B, C, ..., G and accurately identify and classify it as A's signal when A's signal is input.

The intelligent emergency monitoring system according to one embodiment can develop and test an AI network model based on Tensorflow-based PyTorch and MATLAB. Preprocessing and control of data can be performed in Matlab. And because it is possible to load and use neural networks and neural network architectures in TensorFlow and ONNX (Open Neural Network Exchange) model formats, you can freely select libraries when developing AI.

The intelligent emergency monitoring system according to an embodiment of the present invention is intended to propose a new LSTM-based DRNN architecture for ECG recognition.

As an example, an experimental evaluation of the proposed model was performed on two data sets. As a result, it was confirmed that the proposed model performed better than other existing RNN learning methods.

The improvement in the ability of the DRNN model to extract more features of the ECG using deep layers can show higher private label classification efficiency because temporal dependencies within the ECG can be further controlled.

By the intelligent emergency monitoring system and its driving method according to an embodiment of the present invention, the influence of the input sequence length can be evaluated, the relationship between the hidden unit and the gidden layer can be confirmed, and the ECG signal can be grouped using a detailed preprocessing procedure. It can have an effective impact on private label classification.

Since the RNN learning model according to one embodiment has better performance with shorter sequences than existing learning models, it can be useful for real-time personal ECG identification that requires quick results.

Figure 2 is a flow chart showing a bio-signal analysis algorithm in an intelligent emergency monitoring system according to an embodiment of the present invention.

As shown in Figure 2, the analysis algorithm of the intelligent emergency monitoring system according to one embodiment removes noise from biological signals through a pre-processing process and calculates the pressure distribution.

Then, posture analysis is performed according to the pressure distribution. At this time, the signal SNR within a certain pressure value is compared, the optimal signal is selected, the biosignal is extracted using a biosignal extraction algorithm such as heart rate and breathing, and the extracted biosignal is transmitted. In other words, biosignal measurement based on multi-sensor signal processing analysis is possible.

Figure 3 is a block diagram showing the configuration of an intelligent emergency situation monitoring system according to an embodiment of the present invention.

As shown in FIG. 3, the intelligent emergency monitoring system 10 according to an embodiment includes a communication unit 110, an information collection unit 120, a preprocessing unit 130, a learning unit 140, and a situation identification unit 150. do.

The communication unit 110 communicates with any internal component or at least one external terminal through a wired/wireless communication network. Here, wireless Internet technologies include Wireless LAN (WLAN), DLNA (Digital Living Network Alliance), Wibro (Wireless Broadband: Wibro), Wimax (World Interoperability for Microwave Access: Wimax), and HSDPA (High Speed Downlink Packet Access). ), HSUPA (High Speed Uplink Packet Access), IEEE 802.16, Long Term Evolution (LTE), LTE-A (Long Term Evolution-Advanced), Wireless Mobile Broadband Service (WMBS), etc. The communication unit 110 transmits and receives data according to at least one wireless Internet technology, including Internet technologies not listed above.

In addition, short-range communication technologies include Bluetooth, RFID (Radio Frequency Identification), Infrared Data Association (IrDA), UWB (Ultra Wideband), ZigBee, and Near Field Communication (NFC). , Ultrasound Communication (USC), Visible Light Communication (VLC), Wi-Fi, Wi-Fi Direct, etc. may be included. Additionally, wired communication technologies may include Power Line Communication (PLC), USB communication, Ethernet, serial communication, optical/coaxial cables, etc.

At least one external terminal may be the biometric information monitoring device 20. However, it is not limited to this and may include servers at medical institutions or emergency rescue centers capable of immediate response to emergency situations. In addition, it can be implemented as a user terminal carried by the patient or guardian so that the patient or guardian can be quickly notified in the event of an emergency.

The information collection unit 120 collects biometric information signals from at least one biometric information monitoring device. In one embodiment, the information collection unit 120 collects biometric information (big data) required for learning and biometric information for monitoring patient emergency situations from the biometric information monitoring device 20.

In one embodiment, the biometric information monitoring device 20 includes a proximity sensor, an illumination sensor, a motion detection sensor, a pressure sensor, GPS, a gyro sensor, an altitude sensor, a heart rate sensor, a blood pressure sensor, a blood oxygen saturation sensor, a blood sugar sensor, and a sound detection sensor. It may be configured to include one or more sensors selected from a group consisting of sensors, but is not limited thereto.

That is, the information collection unit 120 can receive biometric signals detected from the biometric information monitoring device 20.

The preprocessing unit 130 removes noise from the biometric information signal collected by the information collection unit 120, normalizes it, and preprocesses it.

Figure 4 is an exemplary diagram to explain the bio-signal pre-processing process in the pre-processing unit in one embodiment of the present invention.

In one aspect, the preprocessor 130 includes a band-pass filter that passes only alternating currents of the upper and lower cutoff frequencies, a differentiation filter that applies the signal passed by the band-pass filter to the Fourier spectrum, and a moving average that removes noise errors. Normalize by including filters.

Since the biometric information collected by the preprocessing unit 130 includes noise, difficulties may arise when learning from a BIG SIGNAL-based identification model without a preprocessing process. Therefore, the preprocessor 130 performs a signal processing process to remove noise.

The preprocessor 130 may apply a preprocessing process by applying a band-pass filter, a differentiation filter, and a moving average filter to remove noise from the raw data, and then normalize the data into a signal suitable for learning.

Specifically, Figure 4(a) is after applying a band-pass filter, Figure 4(b) is after applying a differential filter, Figure 4(c) is after applying a moving average filter, and Figure 4(d) is after applying normalization. This is a graph showing biological signals.

Additionally, since the type of input data must be considered in order to develop a learning model suitable for a real-time system, it is also possible to propose and apply a method for grouping input data.

Figure 5 is an example diagram of a method of grouping input data for a real-time system.

Specifically, Figure 5(a) groups input data based on time continuity, and Figure 5(b) groups input data based on peak values.

The learning unit 140 uses the biometric information signals preprocessed in the preprocessing unit 130 and applies them to an artificial intelligence neural network model to learn through machine learning.

In a characteristic aspect of the present invention, the artificial intelligence neural network model of the learning unit 140 is implemented as an RNN model including at least one LSTM track.

The RNN model is a neural network structure that uses a single-layer or multi-layer cyclic connection configuration, and is generally applied to learning data such as voice, video, and strings. This network model has the characteristic of memorizing the state of previous information and applying it to the current input data. This mechanism of the RNN model has strengths in learning sequential data.

Figure 6 is an example diagram for explaining a recurrent neural network (RNN) model according to an embodiment of the present invention.

, 출력

, 이전 hidden state

, 현재의 hidden state

는 수학식 1과 같이 정의된다. As shown in Figure 6, the RNN node currently inputs

, Print

, previous hidden state

, current hidden state

is defined as in Equation 1.

및

는 활성화함수로, hidden layer와 출력 layer에 작용한다. here

and

is an activation function, which acts on the hidden layer and output layer.

,

및

은 반복되는 가중치로 input-to-hidden, hidden to output, hidden-to-hidden의 연결 layer에 적용된다.

,

and

is a repeated weight applied to the connection layers of input-to-hidden, hidden to output, and hidden-to-hidden.

및

은 hidden state에 대한 편향 기간이고, 각각의 출력 상태에 적용된다. 여기서 활성화 함수는 요소별 선형성을 가짐과 동시에 시그몽드, 정류한 선형 유닛 또는 쌍곡선 함수 등의 비선형성 특성도 가진다. also

and

is the bias period for the hidden state and is applied to each output state. Here, the activation function has linearity for each element and also has non-linearity characteristics such as Sigmond, rectified linear unit, or hyperbolic function.

Figure 7 is an example diagram illustrating an LSTM-based RNN model in which RNN nodes are replaced with LSTMs according to an embodiment of the present invention.

Existing RNN models have the disadvantage of being difficult to train on long sequential data due to problems such as vanishing gradients, which interfere with the ability to backpropagate branching or long-term dependencies. Therefore, the LSTM-based RNN model of the intelligent emergency monitoring system 10 according to an embodiment of the present invention can solve existing problems by replacing existing RNN nodes with LSTM.

As shown in Figure 7, a memory cell called a 'gate' exists within the block in the repetitive hidden layer, and the 'gate' memory cell controls the state of new information. The output is determined by clearing the previous hidden state and updating new information. The component functions of each cell are as follows.

)는 새로운 정보의 메모리 셀입력 활성화를 제어한다. 이는 수학식 2와 같은 관계식을 갖는다. input gate (

) controls the activation of memory cell input of new information. This has the same relational expression as Equation 2.

)와, 출력 게이트(

), Forget gate(

), 입력 변조 게이트(

), internal state(

), hidden state(

)의 매개변수가 될 수 있다. Here, the U and W terms are the weight matrices, and the b term is the bias vector. The learning unit 140 of the intelligent emergency monitoring system 10 according to one embodiment, that is, the LSTM-based RNN model, trains a dataset for learning, and the model focuses on learning. b, U and W are input gates (

) and output gate (

), Forget gate(

), input modulation gate (

), internal state (

), hidden state(

) can be a parameter.

)는 출력 흐름을 제어한다. 이는 수학식 3과 같다. Output gate (

) controls the output flow. This is equivalent to equation 3.

The output is the product of the value passed through the _sigmoid function with the value of It's a gate.

)는 내부 상태 정보를 잊어버리는 기능을 제어한다. 이는 수학식 4와 같다.Forget gate(

) controls the function of forgetting internal state information. This is equivalent to equation 4.

)는 입력 데이터

, 이전 셀의 단기기억

을 가지고 시그모이드처리를 하는 부분이다. 시그모이드 함수는 0과 1 사이의 값을 출력하므로 출력값이 0.7이라하면, 0.7만큼의 기억이 장기기억(

)과 곱해진다.Forget gate(

) is the input data

, short-term memory of the previous cell

This is the part where sigmoid processing is performed. The sigmoid function outputs a value between 0 and 1, so if the output value is 0.7, 0.7 of the memory is stored in long-term memory (

) is multiplied by

)는 메모리 셀에 대한 주 입력을 제어한다. 이는 수학식 5와 같다.Input modulation gate (

) controls the main input to the memory cell. This is equivalent to equation 5.

)는 셀의 내부 반복을 제어한다. 이는 수학식 6과 같다.internal state(

) controls the internal repetition of the cell. This is equivalent to equation 6.

)는 이전 데이터 샘플의 정보를 제어한다. 이는 수학식 7과 같다. hidden state(

) controls the information of the previous data sample. This is equivalent to equation 7.

Figure 8 is an example diagram illustrating a bidirectional LSTM-based RNN model for private label classification.

The intelligent emergency monitoring system 10 according to one embodiment can receive personal ECG signals for a specific time, perform direct mapping, and classify personal labels.

)이고, 각 데이터 포인트

는 개인 ECG 신호의 벡터이다. 샘플은 n세그먼트 ECG 신호구성요소로 P의 기간 동안 구성된 크기 T의 범위이다. The input is equally spaced samples (

), and for each data point

is the vector of the individual ECG signal. A sample is a range of size T consisting of n-segment ECG signal components over a period of time P.

In one embodiment, 3, 6, or 9 ECG groups may be used (n=3, 6, 9).

). Scores are obtained by displaying private label predictions at the output of each time step (

).

는 분류 점수의 벡터이고, L은 계층, c는 개인 클래스의 수이다. 점수는 시간 t 에서 개인 레이블에 대한 각 시간 단계에 계산된다. 다중 예측 전체 창에 대해 T는 모든 점수를 하나의 예측으로 합산하여 얻는다. here

is the vector of classification scores, L is the hierarchy, and c is the number of individual classes. A score is computed at each time step for a private label at time t. For the entire window with multiple predictions, T is obtained by summing all scores into one prediction.

It uses a late fusion technique called the sum rule.

To convert the expected score into probability, the Y value, which is the softmax layer of the expected score, is applied as shown in Equation 8.

The bidirectional LSTM-based DRNN model consists of two parallel LSTM tracks. Personal labels are predicted using the past and future context of a specific time step.

Each layer has a forward track (LSTM _fl ) and a backward track (LSTM _bl ).

The two tracks read the ECG input from left to right and right to left, respectively.

및

는 예측결과,

및

의 출력 hidden layer,

및

는 각각의 현재 출력이다. In Equation 9 and Equation 10,

and

is the prediction result,

and

The output of the hidden layer,

and

is the current output of each.

는 개인 레이블의 예측점수를 나타낸다. 후기 융합은 수학식 11과 같이 표현된다. The rear layer (l=1,2,...,L), the top layer L, becomes the output of the sequence. At each time step, the scores are output from the forward LSTM and backward LSTM. Therefore, the combined score

represents the prediction score of the private label. Late fusion is expressed as Equation 11.

Figure 9 is an example diagram for explaining an RNN model structure including an LSTM track according to an embodiment of the present invention.

In one aspect of the present invention, the learning unit 140 may perform label classification using six RNN structures to evaluate model performance. Through experiments, it was confirmed that Arch 6 had the best private label classification performance and was selected as the best model.

Arch 1, 2, and 3 consist of a unidirectional network.

The number of hidden layers in arches 1, 2, and 3 is 1, 2, and 3, respectively.

Arches 4 and 5 are composed of a bidirectional network in the RNN model according to one embodiment, and the numbers of late-fusion layers and hidden layers are 1, 2, and 3, respectively.

Additionally, in the classification procedure of the two-way LSTM-based RNN model for private label classification,

Each ECG data set is separated into a training data set and a test data set. Each training and test sequence is of size 1 × N, where N is the number of samples of the ECG signal. After one-hot sequence encoding, the weight parameters of the bidirectional LSTM are determined using the training data set, and the classification probability value is obtained using the softmax function. After training the RNN, a test sequence evaluates the RNN model. A private label classification decision can be obtained by selecting the class with the highest probability from all classes for each test sequence.

When the patient's biometric information signal is received from the biometric information monitoring device 20, the situation identification unit 150 applies it to an artificial intelligence neural network model based on the learning results learned in the learning unit 140 to identify whether an emergency situation has occurred. .

As an example, the biometric information monitoring device 20 consists of a proximity sensor, an illumination sensor, a motion detection sensor, a pressure sensor, GPS, a gyro sensor, an altitude sensor, a heart rate sensor, a blood pressure sensor, a blood oxygen saturation sensor, a blood sugar sensor, and a sound detection sensor. It may be configured to include one or more sensors selected from the group, but is not limited to this.

Electrocardiogram (ECG) data includes heart rate, QRS period, PR interval, QT interval, type of T wave, time for which each wave is maintained, interval between each wave, and type of T wave obtained from the ECG signal. It may include characteristic values such as amplitude and kurtosis, and the ECG analysis data may include whether the ECG is abnormal based on the ECG data. In other words, the ECG analysis data may include an ECG score consisting of values from 0 to 1, where the ECG score can be determined whether an emergency situation occurs by how much the input ECG data is different or distorted compared to the normal ECG waveform. there is. For example, 0 may be normal and 1 may be abnormal.

The heartbeat, which is the cause of the electrocardiogram signal, is an impulse that originates from the sinus node located in the right atrium, first depolarizes the right and left atrium, and is briefly delayed in the atrioventricular node. Activates the ventricles. The right ventricle, which has the fastest septum and thin wall, activates before the left ventricle, which has a thick wall. The depolarization wave transmitted to the Purkinje fibers spreads from the endocardium to the outer pericardium like a wave in the myocardium, causing ventricular contraction. Because electrical impulses are normally conducted through the heart, the heart contracts approximately 60 to 100 times per minute. Each contraction is represented by one heart beat.

ECG signals include P waves, QRS complex, ST segment, T waves, etc. The isoelectric line is a standard line for measuring electrocardiogram waveforms. The space between the equipotential line and the ST segment is called the ST area. Among these, the important part for analyzing ECG signals is the QRS complex, which consists of Q, R, and S waves.

The QRS complex is the part of the heart where an electrical signal is generated when blood flows into and out of the ventricle and contracts, that is, when polarization occurs in the ventricle. It is the part where the signal is most clearly distinguished in the ECG signal, so it is known that the heart is beating. This is something you can check. The QRS complex occurs at approximately 0.06 to 0.12 seconds. These waves must have a standard shape for the heart's electrical activity to be considered normal. In order to determine whether it is a standard form or not, it is necessary to check whether characteristics such as the time each wave is maintained, the interval between each wave, the amplitude of each wave, and kurtosis are within the normal range.

In one embodiment, identification precision can be expressed and evaluated by equations 12 and 13.

Precision refers to positive predictive value. The closer it is to 100%, the more accurate the prediction may be. In other words, the number of identified biosignals (A, B, ...G) of actual people among the classes classified as positive is calculated.

The average precision of each individual class (POC) refers to the overall precision. c is the total number of patients, tp _c and fp _c are the true and false positive rates of human classification.

Additionally, identification accuracy can be expressed and evaluated by the following equation (14).

Accuracy is the accurately predicted value divided by the sum of all predicted values. The closer it is to 100%, the more accurate the prediction is. This means overall accuracy, which calculates the ratio of correctly predicted labels (the label is the unique identification information of the object) to the overall prediction. TP is the overall true positive rate of the classifier for all classes, TN is the overall true negative rate, FP is the overall false positive rate, and FN is the overall false negative rate.

Identification recall (Recall) can be expressed and evaluated by equations such as Equations 15 and 16.

Recall rate is a calculation of the number of patients correctly classified among all samples of a class, and the closer it is to 100%, the more accurate the prediction is. The average recall for each class (RFC) refers to the overall recall.

The weighted average of precision and recall (F1-score) is expressed and evaluated by a formula such as Equation 17.

은 C 클래스 집합에 있는 개별 예제의 전체수이다. The closer it is to 100%, the more accurate the prediction is. n _c is the number of samples in class c

is the total number of individual examples in the set of C classes.

The following describes the result analysis process of a bidirectional LSTM-based RNN model for label classification in an intelligent emergency monitoring system according to an embodiment.

10 and 11 are graphs to explain the accuracy of personal label classification for the emergency monitoring architecture performed in the intelligent emergency monitoring system according to an embodiment.

Figures 10 and 11 show the private label classification accuracy for the selected architectures.

Specifically, Figure 10 shows the classification accuracy of NSRDB using two selected parameters. The number of hidden units in FIG. 10(a) is 128, and the number of hidden units in FIG. 10(b) is 250. Input sequence length is 2-4 per heart beat.

Figure 11 shows the classification accuracy of MITDB using two selected parameters. The number of hidden units in (a) of FIG. 11 is 128 and the number of hidden units in (b) of FIG. 11 is 250. The input sequence length is 0-2 per heart beat.

In Figures 10(a) and 11(a), the number of hidden units in the hidden layer is 128, and the ECG signal period divided into fixed divisions was used.

As a result of Figures 10(a) and 11(a), the private label classification accuracy varied between 29.7-100% and 1.87-98.53%, respectively.

As a result of Figures 10(b) and 31(b), the classification accuracy varied between 5.5-100% and 2.21-99.73, respectively.

Figure 12 shows the private label classification accuracy of the bidirectional LSTM-based RNN model of the intelligent emergency monitoring system according to an embodiment.

Figure 12 shows the classification accuracy units of MITDB with 250 hidden parameters selected. Input sequence lengths (ISL) are 3, 6, and 9 per heart beat.

In Figure 12, the number of hidden units in the hidden layer is 250, and a signal divided by a fixed division time was used for the ECG.

Classification accuracy varied from 5.5-100% to 63.8-99.8%, and the presented results are for different input sequence lengths, including zero dropout and number of hidden units.

As a result, it showed better performance than the existing RNN, and it can be seen that the input length decreases randomly like a non-fixed group ECG.

Additionally, through this experiment, it can be seen that the private label classification accuracy increases as the number of hidden units decreases and the number of hidden layers increases.

Figure 13 is a flowchart of a method of driving an intelligent emergency situation monitoring system according to an embodiment of the present invention.

A method of driving an intelligent emergency monitoring system according to an embodiment collects information by collecting biometric information signals from at least one biometric information monitoring device (S200).

Then, the biometric information signal collected in the information collection step is preprocessed to remove noise and normalize (S210). At this time, the preprocessing step includes a band-pass filter that passes only alternating currents of the upper and lower cutoff frequencies, a differential filter that applies the signal passed by the band-pass filter to the Fourier spectrum, and a moving average filter that removes noise errors. It is carried out.

In the preprocessing step, the preprocessed biometric information signal is used and applied to an artificial intelligence neural network model to learn through machine learning (S220).

In one aspect of the present invention, the learning step classifies the personal label using samples at equal intervals among the patient's biometric information signals received during a specific time.

And to evaluate model performance, label classification is performed using six RNN structures.

Afterwards, when the patient's biometric information signal is received (S230), it is applied to the artificial intelligence neural network model based on the learning results learned in the learning stage to identify whether an emergency situation has occurred (S240).

At this time, the artificial intelligence neural network model in the learning stage is implemented as an RNN model that includes at least one LSTM track.

In one aspect of the present invention, the situation identification step calculates a pressure distribution from the biometric information signal pre-processed in the pre-processing step, performs posture analysis according to the calculated pressure distribution, and performs a signal-to-noise ratio (SNR) for signals within a certain pressure value. ) is compared and applied to the artificial intelligence neural network model using the selected signal.

The above-described method may be implemented as an application or in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc., singly or in combination.

The program instructions recorded on the computer-readable recording medium may be those specifically designed and configured for the present invention, or may be known and usable by those skilled in the computer software field.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specifically configured to store and perform program instructions, such as ROM, RAM, flash memory, etc.

Examples of program instructions include not only machine language code such as that created by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the invention and vice versa.

Although the above has been described with reference to embodiments, those skilled in the art will understand that various modifications and changes can be made to the present invention without departing from the spirit and scope of the present invention as set forth in the following patent claims. You will be able to.

10: Intelligent emergency monitoring system
20: Biometric information monitoring device 110: Communication Department
120: Information collection unit 130: Preprocessing unit
140: Learning unit 150: Situation identification unit

Claims (10)

an information collection unit that collects biometric information signals from at least one biometric information monitoring device;
a pre-processing unit that removes noise and normalizes the biometric information signal collected by the information collection unit;
A learning unit that uses the biometric information signal pre-processed in the pre-processing unit and applies it to an artificial intelligence neural network model to learn through machine learning; and
A situation identification unit that, when a patient's biometric information signal is received, applies it to an artificial intelligence neural network model based on the learning results learned in the learning unit to identify whether an emergency situation has occurred;
The artificial intelligence neural network model of the learning unit is,
Intelligent emergency monitoring system, which is a RNN model containing at least one LSTM track.
According to claim 1,
The learning department,
An intelligent emergency monitoring system that classifies personal labels using samples at equal intervals among the patient's biometric information signals received during a specific time.
According to claim 2,
The learning department,
An intelligent emergency monitoring system that performs label classification with six RNN structures to evaluate model performance.
According to claim 1,
The situation identification unit,
Calculate the pressure distribution from the biometric information signal pre-processed in the preprocessor, perform posture analysis according to the calculated pressure distribution, and compare the signal-to-noise ratio (SNR) for signals within a certain pressure value to use the selected signal to create artificial intelligence. An intelligent emergency monitoring system applied to an intelligent neural network model.
According to claim 1,
The preprocessor,
An intelligent emergency monitoring system comprising a band-pass filter that passes only alternating currents of the upper and lower cutoff frequencies, a differential filter that applies the signal passed by the band-pass filter to a Fourier spectrum, and a moving average filter that removes noise errors.
An information collection step of collecting biometric information signals from at least one biometric information monitoring device;
A pre-processing step of removing noise and normalizing the biometric information signal collected in the information collection step;
A learning step of applying machine learning to an artificial intelligence neural network model using the biometric information signals preprocessed in the preprocessing step; and
A situation identification step of identifying whether an emergency situation has occurred by applying it to an artificial intelligence neural network model based on the learning results learned in the learning step when the patient's biometric information signal is received;
The artificial intelligence neural network model in the learning stage is,
How to drive an intelligent emergency monitoring system, which is a RNN model containing at least one LSTM track.
According to claim 6,
The learning step is,
A method of operating an intelligent emergency monitoring system that classifies personal labels using samples at equal intervals among the patient's biometric information signals received during a specific time.
According to claim 7,
The learning step is,
How to drive an intelligent emergency monitoring system that performs label classification with 6 RNN structures to evaluate model performance.
According to claim 6,
The situation identification step is,
In the pre-processing step, a pressure distribution is calculated from the pre-processed biometric information signal, posture analysis is performed according to the calculated pressure distribution, and the signal-to-noise ratio (SNR) of the signal within a certain pressure value is compared and the selected signal is used to create artificial intelligence. How to drive an intelligent emergency monitoring system applied to an intelligent neural network model.
According to claim 6,
The preprocessing step is,
Intelligent, performed in a preprocessor including a band-pass filter that passes only alternating currents of the upper and lower cutoff frequencies, a differentiation filter that applies the signal passed by the band-pass filter to the Fourier spectrum, and a moving average filter that removes noise errors. How to operate an emergency monitoring system.