Method and apparatus for multi-stage arrhythmia prediction based on multi-time series input

Abstract

Disclosed are a method and an apparatus for multi-stage arrhythmia prediction based on multi-time series input. The present embodiment provides a method and an apparatus for multi-stage arrhythmia prediction based on multi-time series input, which globally and locally extract a hemodynamic pattern of arrhythmia for each preset time unit (10 seconds, 5 seconds) from an electrocardiogram signal of a preset time (30 seconds), fuse continuous features for each preset time unit (10 seconds, 5 seconds), and learn dynamic changes in the fused data to predict whether arrhythmia occurs for each N-time interval, thereby enabling rapid treatment and response.

Description

Multi-time series input-based multi-stage arrhythmia prediction method and device

[0001] One embodiment of the present invention relates to a multi-stage arrhythmia prediction method and apparatus based on multiple time series inputs.

[0002] The following description merely provides background information related to the present embodiment and does not constitute prior art.

[0003] Arrhythmia prediction algorithms have been proposed and developed in various forms.

[0004] CNN (Convolutional Neural Network)-based algorithms propose a framework capable of extracting various features and developed an arrhythmia prediction algorithm.

[0005] We developed an arrhythmia classification algorithm that extracts the features of each arrhythmia by stacking multiple convolutional layers. However, the proposed algorithm focused primarily on classifying atrial fibrillation, and the algorithm proposed to classify various arrhythmias did not achieve high classification performance because it extracted only local features.

[0006] To overcome the aforementioned problem, we proposed an algorithm that extracts and learns global features of arrhythmias using a Residual Network and a Squeeze-and-Excitation Block (SE), which improved performance for classifying various arrhythmias. However, since it was only possible to classify a single arrhythmia for an input signal, it did not consider the possibility of multiple arrhythmias appearing in the input signal.

[0007] LSTM (Long Short Term Memory)-based algorithms were proposed to overcome the problem of CNN-based algorithms failing to learn continuous relationships. However, when the input length increases and related elements are far apart, long-term dependencies occur, leading to a problem where the continuous relationships between samples cannot be learned. If a large number of LSTMs are configured to overcome the aforementioned long-term dependencies, the learning and inference speeds slow down due to the increase in the number of parameters.

[0008] There is only a single-point prediction algorithm for arrhythmia prediction.

[0009] The atrial fibrillation prediction algorithm provides predictive information for specific single points in time, such as 1 day, 7 days, or one month later. By utilizing this predictive information, it is possible to manage atrial fibrillation, which was difficult to manage because it could not be detected by simple electrocardiograms or long-term Holter monitoring, and to prevent stroke and cardiovascular diseases at an early stage.

[0010] Malignant arrhythmia prediction algorithms have been developed that provide single-point-time prediction information for 5, 10, 30 minutes, and 1 hour in advance. Since malignant arrhythmia is a critical situation that can lead to death immediately upon onset, early preparation is the most effective. However, the longest possible single-point-time prediction of 1 hour in advance is not a sufficient amount of time to make early preparations, such as visiting a nearby hospital or the emergency room. Furthermore, there is a problem in that specific plans for intervention or preparation cannot be established because the trajectory of the disease from the past to the future time point is unknown, making it impossible to determine the specific level of urgency.

[0011] The present embodiment aims to provide a multi-time series input-based arrhythmia multi-stage prediction method and apparatus that can rapidly treat and respond by extracting the cardiodynamic pattern of arrhythmia globally and locally from an electrocardiogram signal of a preset time (30 seconds) for preset time units (10 seconds, 5 seconds), fusing continuous features for each preset time unit (10 seconds, 5 seconds), and learning the dynamic change of the fused data to predict whether arrhythmia will occur at N-hour intervals.

[0012] According to one aspect of the present embodiment, an electrocardiogram waveform input unit that receives an electrocardiogram waveform in N-second intervals in a multi-time series format; a feature extraction unit that extracts electrocardiogram features for each of the electrocardiogram waveforms in preset time intervals; a feature fusion unit that generates feature fusion data by fusing the electrocardiogram features in time series order; a pattern learning unit that learns an arrhythmia occurrence pattern by learning the feature fusion data; and an arrhythmia prediction unit that predicts whether the arrhythmia will occur at N-hour intervals based on the arrhythmia occurrence pattern.

[0013] As explained above, according to the present embodiment, the cardiodynamic pattern of arrhythmia is extracted globally and locally from an electrocardiogram signal of a preset time (30 seconds) for preset time units (10 seconds, 5 seconds), continuous features are fused for each preset time unit (10 seconds, 5 seconds), and the dynamic change of the fused data is learned to predict whether arrhythmia occurs at N-hour intervals, thereby enabling rapid treatment and response.

[0014] FIG. 1 is a diagram showing an electrocardiogram waveform according to the present embodiment.

[0015] FIG. 2 is a diagram showing a classification process for determining arrhythmia of an electrocardiogram waveform according to the present embodiment.

[0016] Figure 3 is a diagram showing the classification results of electrocardiogram waveforms according to the present embodiment.

[0017] FIG. 4 is a diagram showing the P, Q, R, S, and T waves (P wave, QRS complex, T wave) and characteristic indicators of the electrocardiogram according to the present embodiment.

[0018] FIG. 5 is a diagram showing a multi-stage arrhythmia prediction device according to the present embodiment.

[0019] Figure 6 is a diagram showing a structural comparison between a single-point-time prediction of arrhythmia and a multi-stage prediction framework according to the present embodiment.

[0020] FIG. 7 is a diagram showing a multi-stage arrhythmia prediction model according to the present embodiment.

[0021] FIG. 8 is a flowchart illustrating a multi-stage arrhythmia prediction method according to the present embodiment.

[0022] Hereinafter, the present embodiment will be described in detail with reference to the attached drawings.

[0023] FIG. 1 is a diagram showing an electrocardiogram waveform according to the present embodiment.

[0024] The arrhythmia multi-stage prediction device (500) according to the present embodiment can be applied to one-dimensional (1D) biosignal data processing in the field of biosignal processing. The arrhythmia multi-stage prediction device (500) uses an arrhythmia determination algorithm to divide the P waveform, QRS-complex (N,S,V), T waveform, and noise waveform included in the electrocardiogram waveform into waveform units.

[0025] FIG. 2 is a diagram showing a classification process for determining arrhythmia of an electrocardiogram waveform according to the present embodiment.

[0026] The arrhythmia multi-stage prediction device (500) according to the present embodiment applies a one-dimensional arrhythmia judgment algorithm to the input electrocardiogram waveform (raw data) and applies a value that classifies the waveform point by point to output the result.

[0027] Figure 3 is a diagram showing the classification results of electrocardiogram waveforms according to the present embodiment.

[0028] The arrhythmia multi-stage prediction device (500) according to the present embodiment can apply an arrhythmia determination algorithm to the input electrocardiogram waveform (raw data) and apply a value that classifies the waveform point by point, so that if it is noise, it is indicated as 1, and if it is a normal waveform that is not noise, it is indicated as 0.

[0029] , beats can be broadly divided into normal beats (N), supraventricular beats (S), and ventricular beats (V). A beat in an electrocardiogram waveform basically includes a P wave, a QRS wave, and a T wave.

[0030] It includes noise generated by a non-invasive biosignal collection method in the electrocardiogram waveform. The arrhythmia multi-stage prediction device (500) detects the P waveform, Q waveform, R waveform, S waveform, and T waveform included in the electrocardiogram waveform, and classifies the normal beat (N), supraventricular beat (S), and ventricular beat (V). The arrhythmia multi-stage prediction device (500) calculates feature indicator values ​​from the localization of the P waveform, Q waveform, R waveform, S waveform, and T waveform.

[0031] The arrhythmia multi-stage prediction device (500) distinguishes the PR interval, QRS interval, QT interval, ST segment, and RR interval for the input electrocardiogram waveform.

[0032] The arrhythmia multi-stage prediction device (500) can classify beats based on characteristic information of P waveform, Q waveform, R waveform, S waveform, and T waveform based on PR interval, QRS interval, QT interval, ST segment, and RR interval.

[0033] The arrhythmia multi-stage prediction device (500) classifies heart beats into normal beats (N), supraventricular beats (S), and ventricular beats (V). When reading an electrocardiogram waveform, the arrhythmia multi-stage prediction device (500) can detect abnormal conditions based on localization and classification information.

[0034] The arrhythmia multi-stage prediction device (500) performs localization on an electrocardiogram waveform and calculates a feature indicator value, and can design various types of arrhythmia detection machine learning models using the feature indicator value.

[0035] The arrhythmia multi-stage prediction device (500) performs precise localization using segmentation and can obtain reliablely precise feature indicator values ​​based on the localization results. The arrhythmia multi-stage prediction device (500) uses the feature indicator values ​​to detect other abnormal beats, arrhythmias, etc.

[0036] The arrhythmia multi-stage prediction device (500) performs classification on the electrocardiogram waveform to classify the heart beat into a normal beat (N), a supraventricular beat (S), and a ventricular beat (V), and detects the arrhythmia.

[0037] The arrhythmia multi-stage prediction device (500) applies a segmentation technique to the electrocardiogram waveform to identify the characteristic indicator values ​​of each section of the P waveform, Q waveform, R waveform, S waveform, T waveform, and noise waveform included in the electrocardiogram waveform. Based on the characteristic indicator values, the arrhythmia multi-stage prediction device (500) rapidly classifies the heartbeat into normal beats (N), supraventricular beats (S), ventricular beats (V), and noise, and removes the noise to enable a fast and accurate medical judgment.

[0038] FIG. 5 is a diagram showing a multi-stage arrhythmia prediction device according to the present embodiment.

[0039] An electrocardiogram (ECG) interprets the electrical activity of the heart and is recorded by electrodes attached to the skin and external equipment. It is used to measure the rate and regularity of heartbeats and is utilized for the diagnosis or research of abnormal cardiac activity.

[0040] Arrhythmia refers to the heart beating irregularly, meaning the pulse beats excessively fast, slow, or irregularly. Arrhythmia includes emergency arrhythmias (such as ventricular arrhythmia and asystole) and non-emergency arrhythmias such as atrial fibrillation and supraventricular arrhythmia; it is important because early detection can prevent the onset of secondary diseases and critical situations.

[0041] The arrhythmia multi-stage prediction device (500) uses a multi-time series input-based method in which, when one input (a 30-second electrocardiogram waveform) is input, features are extracted by separating them into preset time units (10 seconds, 5 seconds), and then patterns are recognized, and the unique features of each pattern are transmitted to the next time point.

[0042] The arrhythmia multi-stage prediction device (500) predicts the probability of arrhythmia occurrence (time point) at multiple time points based on a single input-based framework. The arrhythmia multi-stage prediction device (500) predicts the probability of arrhythmia occurrence at multiple time points in emergency situations, thereby enabling the establishment of time-specific priorities. The arrhythmia multi-stage prediction device (500) enables the establishment of treatment strategies based on multiple time points in emergency situations.

[0043] The arrhythmia multi-stage prediction device (500) solves the problem of organ-dependent information loss by using an artificial intelligence model. The arrhythmia multi-stage prediction device (500) learns the local / global cardiodynamic patterns of signal data for each unit input sequentially to predict whether an arrhythmia will occur at N-hour intervals. The arrhythmia multi-stage prediction device (500) uses a framework to predict whether an arrhythmia will occur and the risk of occurrence at various hour intervals in a clinical setting, allowing medical staff to monitor the time-dependent changes in the arrhythmia.

[0044] The arrhythmia multi-stage prediction device (500) according to the present embodiment includes an electrocardiogram waveform input unit (510), a feature extraction unit (520), a feature fusion unit (530), a pattern learning unit (540), and an arrhythmia prediction unit (550). The components included in the arrhythmia multi-stage prediction device (500) are not necessarily limited thereto.

[0045] Each component included in the arrhythmia multi-stage prediction device (500) can be connected to a communication path connecting a software module or a hardware module inside the device and operate organically with one another. These components communicate using one or more communication buses or signal lines.

[0046] Each component of the arrhythmia multi-stage prediction device (500) illustrated in FIG. 5 represents a unit that processes at least one function or operation, and can be implemented as a software module, a hardware module, or a combination of software and hardware.

[0047] The electrocardiogram waveform input unit (510) receives an electrocardiogram waveform for each of the N-second intervals (30 seconds). The electrocardiogram waveform input unit (510) collects N sets of electrocardiogram waveforms in N-second intervals (30 seconds).

[0048] The electrocardiogram waveform input unit (510) according to the present embodiment receives an electrocardiogram waveform in N-second intervals in the form of multiple time series.

[0049] The feature extraction unit (520) extracts electrocardiogram features for each electrocardiogram waveform in preset time units. The feature extraction unit (520) inputs each electrocardiogram waveform into a Convolutional Neural Network (CNN) model or a Long Short-Term Memory (LSTM) model in preset time units to extract electrocardiogram features for each time unit.

[0050] The feature extraction unit (520) according to the present embodiment extracts electrocardiogram features for each electrocardiogram waveform at preset time intervals.

[0051] The feature extraction unit (520) includes a first electrocardiogram pattern extraction unit (GSPU) (552) and a second electrocardiogram pattern extraction unit (LSPU) (524).

[0052] The first electrocardiogram pattern extraction unit (GSPU) (552) extracts first electrocardiogram features for each electrocardiogram waveform according to a preset first time unit (10 seconds). Here, the preset first time unit has a longer time unit than the preset second time unit. The first electrocardiogram pattern extraction unit (GSPU) (552) extracts a first time series change pattern based on the electrocardiogram features for each electrocardiogram waveform extracted in the preset first time unit (10 seconds).

[0053] The second electrocardiogram pattern extraction unit (LSPU) (524) extracts second electrocardiogram features for each electrocardiogram waveform for each of the pre-set second time units (5 seconds). Here, the pre-set second time unit has a shorter time unit than the pre-set first time unit. The second electrocardiogram pattern extraction unit (LSPU) (524) extracts a second time series change pattern based on the electrocardiogram features for each electrocardiogram waveform extracted for the pre-set second time unit (5 seconds).

[0054] The feature fusion unit (530) generates feature fusion data by sequentially fusing each electrocardiogram feature by time period. The feature fusion unit (530) propagates and fuses the electrocardiogram features extracted in chronological order.

[0055] The feature fusion unit (530) fuses the cardiac dynamic information extracted from the previous pre-set time unit (10 seconds, 5 seconds) data and the cardiac dynamic information extracted from the current pre-set time unit (10 seconds, 5 seconds) data to generate feature fusion data.

[0056] The feature fusion unit (530) propagates the feature fusion data to the model of the next time point's preset time unit (10 seconds, 5 seconds) data.

[0057] The feature fusion unit (530) generates feature fusion data by sequentially fusing the first electrocardiogram feature received from the first electrocardiogram pattern extraction unit (GSPU) (552) and the second electrocardiogram feature received from the second electrocardiogram pattern extraction unit (LSPU).

[0058] The feature fusion unit (530) transmits the feature fusion data to the first electrocardiogram pattern extraction unit (GSPU) (552) and the second electrocardiogram pattern extraction unit (LSPU) (524) at the next time point, at a preset first time unit (10 seconds). The feature fusion unit (530) continuously performs feature fusion data from the input time to the end time and propagates it to the arrhythmia multi-stage prediction model.

[0059] The feature fusion unit (530) according to the present embodiment generates feature fusion data by fusing electrocardiogram features in a time series order. The feature fusion unit (530) generates feature fusion data by sequentially fusing the first electrocardiogram feature and the second electrocardiogram feature.

[0060] The feature fusion unit (530) generates first feature fusion data by fusing the N-1th electrocardiogram feature of the previous time point and the Nth electrocardiogram feature of the current time point for each electrocardiogram waveform at a preset first time unit (10 seconds), and propagates it to the first electrocardiogram pattern extraction unit (GSPU) (522) of the next time point.

[0061] The feature fusion unit (530) generates second feature fusion data by fusing the N-1 electrocardiogram feature of the previous time point and the Nth electrocardiogram feature of the current time point for each electrocardiogram waveform at a preset second time unit (5 seconds), and propagates it to the second electrocardiogram pattern extraction unit (LSPU) (524) of the next time point.

[0062] The feature fusion unit (530) continuously fuses the first electrocardiogram feature and the second electrocardiogram feature sequentially from the input time to the end time to generate feature fusion data.

[0063] The pattern learning unit (540) finally inputs the fused feature data into a fully connected layer neural network to learn the arrhythmia occurrence pattern. The pattern learning unit (540) propagates and updates the fused feature data.

[0064] The pattern learning unit (540) learns global / dynamic time series patterns by passing feature fusion data through fully connected layers. The pattern learning unit (540) performs iterative connections up to time N seconds to extract and learn global / dynamic pattern changes that appear over a long interval.

[0065] The pattern learning unit (540) according to the present embodiment learns the arrhythmia occurrence pattern by learning feature fusion data. The pattern learning unit (540) learns the arrhythmia occurrence pattern by inputting the feature fusion data into a fully connected layer neural network.

[0066] The arrhythmia prediction unit (550) outputs multi-stage arrhythmia prediction information using feature fusion data. The arrhythmia prediction unit (550) predicts whether arrhythmia will occur at N-hour intervals based on the arrhythmia occurrence pattern.

[0067] The arrhythmia prediction unit (550) according to the present embodiment predicts whether an arrhythmia will occur at N-hour intervals based on the arrhythmia occurrence pattern.

[0068] Figure 6 is a diagram showing a structural comparison between a single-point-time prediction of arrhythmia and a multi-stage prediction framework according to the present embodiment.

[0069] The arrhythmia multi-stage prediction device (500) observes the cardiodynamic patterns of arrhythmia globally and locally from electrocardiogram waveforms of preset time units (each 30 seconds) and electrocardiogram waveforms of preset time units (each 10 seconds). The arrhythmia multi-stage prediction device (500) extracts electrocardiogram features. The arrhythmia multi-stage prediction device (500) learns dynamic changes by fusing and combining continuous features for each electrocardiogram waveform of preset time units (10 seconds). The arrhythmia multi-stage prediction device (500) rapidly predicts whether an arrhythmia will occur at N-hour intervals.

[0070] As illustrated in FIG. 6, the arrhythmia multi-stage prediction device (500) comprises a temporary block (Temporal Block X) of a preset time unit (each 10 seconds). v-2 , Temporal Block X v-1 , Temporal Block X v ) is input into the prediction model. The arrhythmia multi-stage prediction device (500) predicts whether arrhythmia will occur in N-hour intervals (1 hour, 2 hours, 3 hours, 4 hours, 5 hours).

[0071] FIG. 7 is a diagram showing a multi-stage arrhythmia prediction model according to the present embodiment.

[0072] The arrhythmia multi-stage prediction device (500) identifies the change pattern of arrhythmia in a specific input interval and extracts features for each input time period to accurately predict the severity for each time period. The arrhythmia multi-stage prediction device (500) continuously learns cardiac dynamic characteristics and global characteristics and then fuses the features.

[0073] FIG. 8 is a flowchart illustrating a multi-stage arrhythmia prediction method according to the present embodiment.

[0074] The arrhythmia multi-stage prediction device (500) configures an N-second electrocardiogram waveform to be used as input to the framework (S810).

[0075] The arrhythmia multi-stage prediction device (500) performs preprocessing to correct baseline wander for N-second electrocardiogram waveforms and generates preprocessed data (S820).

[0076] The arrhythmia multi-stage prediction device (500) inputs preprocessed data for each pre-set time unit (10 seconds, 5 seconds) into each framework (S830). The arrhythmia multi-stage prediction device (500) extracts electrocardiogram features from each of the preprocessed data for each pre-set time unit (10 seconds, 5 seconds) (S840).

[0077] The arrhythmia multi-stage prediction device (500) propagates and updates electrocardiogram features in a preset time unit (10 seconds, 5 seconds) of the next time point (S850). In step S850, the arrhythmia multi-stage prediction device (500) extracts electrocardiogram features and then propagates and updates the electrocardiogram features of the current time point in a preset time unit (10 seconds, 5 seconds) of the next time point.

[0078] The arrhythmia multi-stage prediction device (500) checks whether progress has been made to a preset data interval (N-10 to N-second data interval) (S860). If, as a result of the check in step S860, progress has not been made to a preset data interval (N-10 to N-second data interval), steps S840 and S850 are repeated.

[0079] If, as a result of the verification in step S860, the process proceeds to a pre-set data interval (N-10 to N second data interval), the arrhythmia multi-stage prediction device (500) predicts and outputs whether arrhythmia occurs at N time intervals based on the arrhythmia occurrence pattern (S870).

[0080] Although steps S810 to S870 are described as being executed sequentially in FIG. 8, they are not necessarily limited thereto. In other words, since it is possible to modify and execute the steps described in FIG. 8 or to execute one or more steps in parallel, FIG. 8 is not limited to a chronological order.

[0081] As described above, the multi-stage arrhythmia prediction method according to the present embodiment described in FIG. 8 can be implemented as a program and recorded on a computer-readable recording medium. A computer-readable recording medium on which a program for implementing the multi-stage arrhythmia prediction method according to the present embodiment is recorded includes all types of recording devices in which data that can be read by a computer system is stored.

[0082] The above description is merely an illustrative explanation of the technical concept of the present embodiment, and a person skilled in the art to which the present embodiment belongs would be able to make various modifications and variations within the scope of the essential characteristics of the present embodiment. Accordingly, the present embodiments are intended to explain, not limit, the technical concept of the present embodiment, and the scope of the technical concept of the present embodiment is not limited by these embodiments. The scope of protection of the present embodiment shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present embodiment.

Claims

1. An electrocardiogram waveform input unit that receives electrocardiogram waveforms in N-second intervals in a multi-time series format; A feature extraction unit that extracts electrocardiogram features for each of the above electrocardiogram waveforms in preset time units; A feature fusion unit that generates feature fusion data by fusing the above electrocardiogram features in chronological order; A pattern learning unit that learns arrhythmia occurrence patterns by learning the above-mentioned feature fusion data; An arrhythmia prediction unit that predicts whether the arrhythmia occurs at N-hour intervals based on the above arrhythmia occurrence pattern; A multi-stage arrhythmia prediction device characterized by including 2. In Paragraph 1, The above feature extraction unit is, A first electrocardiogram pattern extraction unit that extracts first electrocardiogram features for each of the above electrocardiogram waveforms according to a first time unit set in advance; A second electrocardiogram pattern extraction unit that extracts second electrocardiogram features for each of the above electrocardiogram waveforms according to a pre-set second time unit; A multi-stage arrhythmia prediction device characterized by including 3. In Paragraph 2, A multi-stage arrhythmia prediction device characterized in that the above-mentioned second time unit has a shorter time unit than the above-mentioned first time unit.

4. In Paragraph 2, The above feature fusion section is, A multi-stage arrhythmia prediction device characterized by sequentially fusing the first electrocardiogram feature and the second electrocardiogram feature to generate the feature fusion data.

5. In Paragraph 3, The above feature fusion section is, A multi-stage arrhythmia prediction device characterized by generating first feature fusion data by fusing the N-1th electrocardiogram feature of a previous time point and the Nth electrocardiogram feature of a current time point for each of the above electrocardiogram waveforms according to a first time unit set, and propagating the data to the above first electrocardiogram pattern extraction unit of a next time point.

6. In Paragraph 3, The above feature fusion section is, A multi-stage arrhythmia prediction device characterized by generating second feature fusion data by fusing the N-1 electrocardiogram feature of a previous time point and the Nth electrocardiogram feature of a current time point for each of the above electrocardiogram waveforms according to a pre-set second time unit, and propagating the second electrocardiogram pattern extraction unit of a next time point.

7. In Paragraph 3, The above feature fusion section is, A multi-stage arrhythmia prediction device characterized by continuously sequentially fusing the first electrocardiogram feature and the second electrocardiogram feature from the input time to the end time to generate the feature fusion data.

8. In Paragraph 1, The above pattern learning unit is, A multi-stage arrhythmia prediction device characterized by inputting the above-mentioned feature fusion data into a fully connected layer neural network to learn the above-mentioned arrhythmia occurrence pattern.

Images