A method and system for AED defibrillation rhythm classification and decision-making based on RF-SVM classification model

Abstract

This invention discloses an AED defibrillation rhythm classification decision method and system based on an RF-SVM classification model, comprising the following steps: S1: Acquire ECG data to be processed and sort the ECG data into N waveform types; S2: Divide the N waveform types into defibrillable waveforms and non-defibrillable waveforms; S3: Preprocess the defibrillable and non-defibrillable waveforms respectively to obtain data to be extracted, and then extract features from the data to be extracted; S4: Input the features extracted in S3 into the constructed RF-SVM classification model for training, and complete the optimization of the RF-SVM classification model; S5: Use the optimized RF-SVM classification model to build a classification inference function and output the ECG rhythm classification decision.

Description

Technical Field

[0001] This invention relates to the field of emergency medical device technology, and in particular to an AED defibrillation rhythm classification decision method and system based on the RF-SVM (random forest-Support Vector Machine) classification model. Background Technology

[0002] With the development of technology, the technology related to AEDs has become relatively mature, and their clinical application is very widespread abroad. In the use of AEDs, the automatic identification and correct classification of heart rhythm are its core and key technologies. That is, when a patient experiences VF or persistent VT, it is necessary to immediately and accurately identify and judge the condition, and immediately perform treatment measures such as electrical defibrillation and cardioversion.

[0003] In existing technologies, it is generally necessary to extract the frequency domain features of the heart rhythm to distinguish between shockable and non-shockable rhythms. However, this method is time-consuming, requires dedicated DSP devices for engineering implementation, and the classification decision reasoning and model training cannot be separated offline, resulting in long decision-making times; the rhythm classification specificity and sensitivity are also low.

[0004] For example, Chinese patent (CN104382590A) discloses an automatic identification and classification method for shockable cardioversion rhythms that combines electrocardiogram time-frequency domain feature analysis. It distinguishes between shockable and non-shockable cardioversion rhythms based on frequency domain signals such as the maximum amplitude ratio (Pa), average amplitude ratio (Pb), and average deviation ratio (Pc), thus enhancing the algorithm's efficiency. However, this method is time-consuming and has a long decision-making time, resulting in low rhythm classification specificity and sensitivity. Summary of the Invention

[0005] To address the issues of long classification time and low sensitivity in existing defibrillation rhythm classification technologies, this invention proposes an AED defibrillation rhythm classification decision-making method and system based on the RF-SVM classification model. By using the Random Forest (RF) algorithm to quickly optimize the input feature variables of the SVM classifier, the classification speed and accuracy are improved.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for classifying and deciding the defibrillation rhythm of an AED based on an RF-SVM classification model includes the following steps:

[0008] S1: Acquire the ECG data to be processed and sort the ECG data into N waveform types;

[0009] S2: Divide the N waveform types into defibrillable waveforms and non-defibrillable waveforms;

[0010] S3: Preprocess the defibrillable and non-defibrillable waveforms to obtain the data to be extracted, and then perform feature extraction on the data to be extracted.

[0011] S4: Input the features extracted in S3 into the constructed RF-SVM classification model for training, and complete the optimization of the RF-SVM classification model;

[0012] S5: Build a classification inference function using the optimized RF-SVM classification model and output the ECG rhythm classification decision.

[0013] Preferably, in S1, the waveform types include N, AF, SVTA, VF, VFL, and VT, and each waveform includes at least 100 samples, with each sample having at least 1000 sampling points.

[0014] Preferably, 1000 sampling points with an average heart rate greater than 150 beats / minute are selected from the VT waveform to construct the VT_hi waveform.

[0015] Preferably, N, AF, and SVTA are classified as non-defibrillable waveforms, while VF, VFL, and VT_hi are classified as defibrillable waveforms.

[0016] Preferably, in step S3, the preprocessing includes:

[0017] Step S31: Design a low-pass filter with a cutoff frequency of 20 Hz using a Butterworth first-order filter to filter out electromyographic interference from both defibrillable and non-defibrillable waveforms.

[0018] Step S32: Design a high-pass filter with a cutoff frequency of 2Hz using a Butterworth second-order filter to perform high-pass filtering on both defigurable and non-defigurable waveforms to filter out baseline drift.

[0019] Step S33: Design band-stop filters with cutoff frequencies of [49.8, 50.2] and [49.9, 50.1] using a Butterworth first-order filter to perform high-pass filtering on both defigurable and non-defigurable waveforms, thereby filtering out power frequency interference.

[0020] Preferably, in step S3, the extracted features include electrocardiogram signals, waveform signals, and time-frequency feature signals;

[0021] ECG signals include: SM, PD, bpm;

[0022] The waveform signals include: 'psdE', 'eE', 'ApEn', 'SpEn', 'FuzzyEn', 'PeEn', 'enveEn';

[0023] The time-frequency characteristic signals include: 'max', 'min', 'mean', 'peak', 'arv', 'var', 'std', 'kurtosis', 'skewness', 'rms', 'waveformF', 'peakF', 'impulseF', 'clearanceF', 'FC', 'MSF', 'RMSF', 'VF', 'RVF', 'SKMean', 'SKStd', 'SKSkewness', 'SKKurtosis'.

[0024] Preferably, S4 includes:

[0025] S4-1: The RF algorithm is used to calculate the importance coefficient of the features extracted in S3. Features with an importance coefficient lower than the preset coefficient are removed, that is, features with an importance coefficient not lower than the preset coefficient are retained to obtain the first feature group.

[0026] The formula for calculating the importance coefficient is:

[0027] X j =Σ(errOOB2-errOOB1) / Ntree (1)

[0028] In formula (1), X j The importance coefficient is represented by errOOB1, which is the out-of-bag error calculated for each decision tree in the random forest using the corresponding OOB data (out-of-bag data); errOOB2 is the out-of-bag error calculated again after randomly adding noise interference to the features extracted by S3; Ntree is the total number of decision trees in the random forest.

[0029] S4-2: Input the first feature group into the SVM for training to complete the optimization of the RF-SVM classification model. The optimization parameters include support vector machine coefficients, support vector machine dimension, support vector machine, intercept, C coefficient, and gamma coefficient.

[0030] Preferably, before inputting into the SVM, each feature group needs to be standardized using the following formula:

[0031] x′ i =(x i -x mean ) / x std (2)

[0032] In formula (2), x′ i Represents the standardized features; x i Indicates the features before standardization; x mean x represents the mean of each feature group; std This represents the standard deviation of each feature group.

[0033] The present invention also provides an AED defibrillation rhythm classification decision system based on an RF-SVM classification model, including a data acquisition unit, a waveform sorting unit, a waveform segmentation unit, a preprocessing unit, a feature extraction unit, an RF-SVM classification model construction unit, and a decision output unit;

[0034] The data acquisition unit is used to acquire ECG data to be processed.

[0035] The waveform sorting unit is used to sort the ECG data to be processed to obtain N waveforms;

[0036] Waveform division unit, used to divide N waveforms into defibrillable waveforms and non-defibrillable waveforms;

[0037] The preprocessing unit is used to preprocess the divided defibrillable and non-defibrillable waveforms to obtain the data to be extracted.

[0038] The feature extraction unit is used to extract features from the data to be extracted.

[0039] The RF-SVM classification model building unit is used to build the RF-SVM classification model and train and optimize it based on the extracted features.

[0040] The decision output unit is used to output ECG rhythm classification decisions.

[0041] In summary, by adopting the above technical solution, the present invention has at least the following beneficial effects compared with the prior art:

[0042] This invention optimizes the RF-SVM classification model by classifying the extracted data and using the Random Forest (RF) algorithm to filter the features of defibrillable waveforms, thereby improving classification speed and accuracy. Attached image description:

[0043] Figure 1 This is a schematic diagram of an AED defibrillation rhythm classification decision method based on an RF-SVM classification model according to an exemplary embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of an AED defibrillation rhythm classification decision system based on an RF-SVM classification model according to an exemplary embodiment of the present invention. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to embodiments and specific implementation methods. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0046] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0047] like Figure 1 As shown, this invention provides an AED defibrillation rhythm classification decision method based on the RF-SVM classification model, specifically including the following steps:

[0048] S1: Obtain ECG (electrocardiogram) data to be processed from authoritative databases such as MIT and CU, and sort the ECG data into N types of waveforms.

[0049] In this embodiment, the WFDB tool can be used on a Python or Matalab platform to sort the ECG data to be processed according to aux_note (labeled in the MIT and CU international authoritative ECG databases) to obtain N waveforms, including N, AF, SVTA, VF, VFL, and VT (which are the most important waveform types in the MIT and CU international authoritative ECG databases and are also the waveforms that require defibrillation classification as specified in the national standard AED defibrillation classification criteria). Each waveform includes at least 100 samples, and each sample has at least 1000 sampling points.

[0050] S2: Divide the N waveforms into defibrillable waveforms and non-defibrillable waveforms. Set the label of non-defibrillable waveforms to 1 and the label of defibrillable waveforms to 2. This is used to label the data samples, then train the model, and finally test the selected model.

[0051] In this embodiment, 1000 sampling points with an average heart rate of more than 150 beats / minute (i.e., greater than 150 beats / minute) are selected from the VT waveform to construct the VT_hi waveform.

[0052] In this embodiment, according to medical standards (within the scope of medical professions), N, AF, and SVTA are classified as non-defibrillable waveforms and labeled as 1; VF, VFL, and VT_hi are classified as defibrillable waveforms and labeled as 2.

[0053] In this embodiment, the waveform classification is defined by the AHA (American Heart Association) standard and is also a common industry practice. VT waveforms with a center rate of 150 beats / minute or higher are constructed as VT_hi and classified as defibrillable waveforms, labeled as 2. As for VT waveforms with a center rate of less than or equal to 150 beats / minute, the standard is not clearly defined and no processing is performed. For other waveforms, as long as the waveform type is identified, whether or not defibrillation is required is clearly defined by the AHA (American Heart Association) standard, and there is no need to judge the heart rate again.

[0054] S3: Preprocess the defibrillable and non-defibrillable waveforms to obtain the data to be extracted, and then perform feature extraction on the data to be extracted.

[0055] In this embodiment, both defibrillable and non-defibrillable waveforms undergo preprocessing. The processing steps for both waveforms are the same throughout the entire processing flow; only their labels remain independent.

[0056] Preprocessing includes three filtering steps, performed in the order of low-pass filtering, high-pass filtering, and band-stop filtering:

[0057] S31: A low-pass filter with a cutoff frequency of 20 Hz is designed using a Butterworth first-order filter to filter out electromyographic interference from both defibrillable and non-defibrillable waveforms.

[0058] Step S32: Design a high-pass filter with a cutoff frequency of 2 Hz using a Butterworth second-order filter to perform high-pass filtering on both defibrillable and non-defibrillable waveforms to filter out baseline drift.

[0059] Step S33: Design band-stop filters with cutoff frequencies of [49.8, 50.2] and [49.9, 50.1] using a Butterworth first-order filter to perform high-pass filtering on both defigurable and non-defigurable waveforms, thereby filtering out power frequency interference.

[0060] In this embodiment, the features extracted from the data to be extracted include electrocardiogram signals, waveform signals, and time-frequency feature signals.

[0061] In this embodiment, the electrocardiogram signal includes SM (slope peak), PD (amplitude probability density), and bpm (heart rate);

[0062] In this embodiment, the waveform signal includes 7 entropy values: 'psdE', 'eE', 'ApEn', 'SpEn', 'FuzzyEn', 'PeEn', 'enveEn';

[0063] In this embodiment, the time-frequency characteristic signal includes 19 time-frequency characteristics of signal processing: 'max', 'min', 'mean', 'peak', 'arv', 'var', 'std', 'kurtosis', 'skewness', 'rms', 'waveformF', 'peakF', 'impulseF', 'clearanceF', 'FC', 'MSF', 'RMSF', 'VF', 'RVF', 'SKMean', 'SKStd', 'SKSkewness', 'SKKurtosis'.

[0064] S4: Input the features extracted in S3 into the constructed RF-SVM classification model, and perform training to complete the optimization.

[0065] S4-1: The Random Forest (RF) algorithm is used to calculate the importance coefficient of the features extracted in S3. Features with an importance coefficient lower than the preset coefficient (e.g., 0.001) are removed, and features with an importance coefficient not lower than the preset coefficient are retained to obtain the first feature group.

[0066] In this embodiment, the importance coefficient can be calculated using the out-of-bag error rate, as shown in the formula:

[0067] X j =Σ(errOOB2-errOOB1) / Ntree (1)

[0068] In formula (1), X j The importance coefficient is represented by errOOB1, which is the out-of-bag error calculated for each decision tree in the random forest using the corresponding out-of-bag data; errOOB2 is the out-of-bag error calculated again after randomly adding noise interference to the features extracted by S3; Ntree is the total number of decision trees in the random forest, which can be set to 1000.

[0069] In this embodiment, the out-of-bag error is calculated using the features extracted by S3, and the process is as follows:

[0070] For each decision tree in the random forest, the first out-of-bag error, denoted as errOOB1, is calculated using the corresponding out-of-bag data.

[0071] Noise interference is randomly added to the features extracted by S3 (the purpose is to randomly change the value of the sample at feature X, so as to achieve the purpose of screening the importance of different samples), and the second out-of-bag error is calculated again, denoted as errOOB2;

[0072] If there are N trees in the random forest, then the importance of feature X is Σ(errOOB2-errOOB1) / Ntree. If adding random noise to a feature significantly reduces the accuracy of out-of-bag classification, it indicates that this feature has a significant impact on the classification results of the samples and is of high importance.

[0073] S4-2: Input the first feature group into the SVM for training, and optimize the RF-SVM classification model based on sensitivity and specificity.

[0074] The first feature group with labels is randomly divided into a training set (70%) and a test set (30%). A large number of C and gamma hyperparameters are set using the gridSearch() function in the Python sklearn library. The training set is input into an SVM for extensive iterative training. Finally, the test set is used to make decisions on the models trained under different hyperparameters using classification inference expressions. The decision results are compared with the actual results, and the model with the best sensitivity and specificity is selected.

[0075] In this embodiment, the SVM uses the ovo mode and the rbf kernel function. Before inputting into the SVM, each feature vector group needs to be standardized using z-score, as shown in the formula:

[0076] x′ i =(x i -x mean ) / x std (2)

[0077] In formula (2), x′ i Represents the standardized features; x i Indicates the features before standardization; x mean x represents the mean of each feature group; std This represents the standard deviation of each feature group.

[0078] In this embodiment, the optimization parameters of the RF-SVM classification model include support vector machine coefficients, support vector machine dimension, support vector machine, intercept, C coefficient, and gamma coefficient.

[0079] S5: Build a classification inference function using the optimized RF-SVM classification model and output the ECG rhythm classification decision.

[0080] In this embodiment, the expression for the classification inference function is:

[0081]

[0082] In formula (3), y represents the classification inference result; N represents the number of support vector machines (SVMs); Gamma represents the hyperparameters set during model training; X represents the first feature group; li Let represent the i-th support vector machine; rho represents the intercept obtained during model training.

[0083] Classification decision-making involves extracting features from the electrocardiogram signal to be analyzed and feeding them into a classification inference function expression for calculation. The result is calculated based on the classification inference expression. If the classification inference result is 1, it means that defibrillation is not possible; if the classification inference result is 2, it means that defibrillation is possible.

[0084] Based on the above methods, such as Figure 2 As shown, the present invention also provides an AED defibrillation rhythm classification decision system based on an RF-SVM classification model, comprising a data acquisition unit, a waveform sorting unit, a waveform division unit, a preprocessing unit, a feature extraction unit, an RF-SVM classification model construction unit, and a decision output unit connected in sequence; that is, the output end of the data acquisition unit is connected to the input end of the waveform sorting unit, the output end of the waveform sorting unit is connected to the input end of the waveform division unit, the output end of the waveform division unit is connected to the input end of the preprocessing unit, the output end of the preprocessing unit is connected to the input end of the feature extraction unit, the output end of the feature extraction unit is connected to the input end of the RF-SVM classification model construction unit, and the output end of the RF-SVM classification model construction unit is connected to the input end of the decision output unit.

[0085] In this embodiment, the data acquisition unit is used to acquire ECG (electrocardiogram) data to be processed from authoritative databases such as MIT and CU;

[0086] The waveform sorting unit is used to sort the ECG data to be processed according to aux_note using the WFDB tool on the Python or Matalab platform to obtain N waveforms, including N, AF, SVTA, VF, VFL, and VT. Each waveform includes at least 100 samples, and each sample has at least 1000 sampling points.

[0087] Waveform division unit, used to divide N waveforms into defibrillable waveforms and non-defibrillable waveforms;

[0088] The preprocessing unit is used to preprocess the divided defibrillable and non-defibrillable waveforms to obtain the data to be extracted.

[0089] The feature extraction unit is used to extract features from the data to be extracted.

[0090] The RF-SVM classification model building unit is used to build RF-SVM classification models and train and optimize them based on the extracted features.

[0091] The decision output unit is used to output classification decisions.

[0092] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.

Claims

1. A method for classifying and deciding the rhythm of AED defibrillation based on an RF-SVM classification model, characterized in that, Includes the following steps: S1: Acquire the ECG data to be processed and sort the ECG data into N waveform types; S2: Divide the N waveform types into defibrillable waveforms and non-defibrillable waveforms; S3: Preprocess the defibrillable and non-defibrillable waveforms to obtain the data to be extracted, and then perform feature extraction on the data to be extracted. S4: Input the features extracted in S3 into the constructed RF-SVM classification model for training, and complete the optimization of the RF-SVM classification model; S5: Use the optimized RF-SVM classification model to build a classification inference function and output the ECG rhythm classification decision; Wherein, S4 includes: S4-1: The RF algorithm is used to calculate the importance coefficient of the features extracted in S3. Features with an importance coefficient lower than the preset coefficient are removed, that is, features with an importance coefficient not lower than the preset coefficient are retained to obtain the first feature group. S4-2: Input the first feature group into the SVM for training to complete the optimization of the RF-SVM classification model.

2. The AED defibrillation rhythm classification and decision-making method based on the RF-SVM classification model as described in claim 1, characterized in that, In S1, the waveform types include N, AF, SVTA, VF, VFL, and VT. Each waveform includes at least 100 samples, and each sample has at least 1000 sampling points.

3. The AED defibrillation rhythm classification and decision-making method based on the RF-SVM classification model as described in claim 2, characterized in that, The VT_hi waveform is constructed by selecting 1000 sampling points with an average heart rate greater than 150 beats / minute from the VT waveform.

4. The AED defibrillation rhythm classification and decision-making method based on the RF-SVM classification model as described in claim 3, characterized in that, N, AF, and SVTA are classified as non-defibrillable waveforms, while VF, VFL, and VT_hi are classified as defibrillable waveforms.

5. The AED defibrillation rhythm classification and decision-making method based on the RF-SVM classification model as described in claim 1, characterized in that, In step S3, the preprocessing includes: Step S31: Design a low-pass filter with a cutoff frequency of 20 Hz using a Butterworth first-order filter to filter out electromyographic interference from both defibrillable and non-defibrillable waveforms. Step S32: Design a high-pass filter with a cutoff frequency of 2Hz using a Butterworth second-order filter to perform high-pass filtering on both defigurable and non-defigurable waveforms to filter out baseline drift. Step S33: Design band-stop filters with cutoff frequencies of [49.8, 50.2] and [49.9, 50.1] using Butterworth first-order filters to perform high-pass filtering on both defigurable and non-defigurable waveforms, thereby filtering out power frequency interference.

6. The AED defibrillation rhythm classification and decision-making method based on the RF-SVM classification model as described in claim 1, characterized in that, In S3, the extracted features include electrocardiogram signals, waveform signals, and time-frequency feature signals; ECG signals include: SM, PD, bpm; The waveform signals include: 'psdE', 'eE', 'ApEn', 'SpEn', 'FuzzyEn', 'PeEn', 'enveEn'; The time-frequency characteristic signals include: 'max', 'min', 'mean', 'peak', 'arv', 'var', 'std', 'kurtosis', 'skewness', 'rms', 'waveformF', 'peakF', 'impulseF', 'clearanceF', 'FC', 'MSF', 'RMSF', 'VF', 'RVF', 'SKMean', 'SKStd', 'SKSkewness', 'SKKurtosis'.

7. The AED defibrillation rhythm classification and decision-making method based on the RF-SVM classification model as described in claim 1, characterized in that, The formula for calculating the importance coefficient is: （1） In formula (1), The value represents the importance coefficient; errOOB1 is the out-of-bag error calculated using the corresponding OOB data for each decision tree in the random forest; errOOB2 is the out-of-bag error calculated again after randomly adding noise interference to the features extracted by S3; Ntree is the total number of decision trees in the random forest. The optimization parameters include support vector machine coefficients, support vector machine dimension, support vector machine, intercept, C coefficient, and gamma coefficient.

8. The AED defibrillation rhythm classification and decision-making method based on the RF-SVM classification model as described in claim 7, characterized in that, Before inputting the SVM, each feature vector group needs to be standardized, using the following formula: In formula (2), Represents the standardized features; Indicates the features before standardization; This represents the mean of each feature group; This represents the standard deviation of each feature group.

9. An AED defibrillation rhythm classification and decision system based on an RF-SVM classification model, based on the method of any one of claims 1-8, characterized in that, It includes a data acquisition unit, a waveform sorting unit, a waveform segmentation unit, a preprocessing unit, a feature extraction unit, an RF-SVM classification model construction unit, and a decision output unit; The data acquisition unit is used to acquire ECG data to be processed. The waveform sorting unit is used to sort the ECG data to be processed to obtain N waveforms; A waveform division unit is used to divide N waveforms into defibrillable waveforms and non-defibrillable waveforms; The preprocessing unit is used to preprocess the divided defibrillable and non-defibrillable waveforms to obtain the data to be extracted. The feature extraction unit is used to extract features from the data to be extracted; The RF-SVM classification model building unit is used to build the RF-SVM classification model and train and optimize it based on the extracted features. The decision output unit is used to output ECG rhythm classification decisions.

Images