Atrial fibrillation identification method based on multi-feature-value and multi-model fusion

Abstract

The invention relates to a multi-feature-value multi-model fusion atrial fibrillation recognition method, and relates to the technical field of medical signal processing, and the method comprises the steps: obtaining an RR interval sequence; in parallel, extracting a multi-dimensional short-time feature set containing the RR interval fuzzy measure entropy from the RR interval sequence of the preset short time period, and generating a short-time recognition result through a short-time model; meanwhile, a multi-dimensional long-time feature set is extracted from the RR interval sequence of the preset long time period, and a long-time recognition result is generated through a long-time model; and the model result fusion module outputs a final atrial fibrillation recognition result based on the long-time recognition result by combining the proportion of the atrial fibrillation conclusions in the plurality of short-time recognition results and applying dual threshold judgment logic. According to the method, non-linear dynamic characteristics of different time scales are combined, and an intelligent fusion strategy of hierarchical decision is adopted, so that the detection sensitivity of paroxysmal atrial fibrillation and the suppression capability of interference noise can be effectively considered, and the comprehensive accuracy of automatic identification of atrial fibrillation is improved.

Description

Technical Field

[0001] This application relates to the field of medical signal processing technology, and in particular to a method for atrial fibrillation identification by multi-feature value and multi-model fusion. Background Technology

[0002] Atrial fibrillation (AF) is one of the most common arrhythmias in clinical practice, characterized by disordered atrial electrical activity and an absolutely irregular ventricular rate. AF not only causes discomfort such as palpitations and chest tightness, but is also a significant risk factor for serious cardiovascular events such as ischemic stroke and heart failure. Holter monitoring provides continuous ECG recordings for up to 24 hours or even longer, making it a crucial technique for diagnosing, especially screening, paroxysmal AF. However, the amount of data generated by Holter monitoring is extremely large, and relying entirely on manual interpretation for AF diagnosis is not only time-consuming and labor-intensive, but also prone to missed or misdiagnosed cases due to visual fatigue. Therefore, developing efficient and accurate automated AF identification algorithms has significant clinical value.

[0003] Currently, automatic atrial fibrillation identification algorithms based on RR interval sequences have been extensively studied. Their basic principle is to utilize the absolutely irregular ventricular rate characteristic during an atrial fibrillation episode. However, existing technologies still have several limitations in the pursuit of high-precision identification.

[0004] Many traditional methods rely on relatively basic time-domain or frequency-domain statistical features, which are insufficient for describing the highly irregular and nonlinear dynamics of the RR interval sequence in atrial fibrillation, driven by complex intrinsic mechanisms. When faced with atypical atrial fibrillation, atrial flutter with irregular conduction, or interference from factors such as frequent premature beats and artifacts, the discriminative power of these features decreases significantly, leading to room for improvement in the accuracy and robustness of the algorithms.

[0005] Furthermore, existing technologies often suffer from limitations in their analysis across different time scales. Some methods focus on detailed analysis of short-term RR intervals, which, while capturing instantaneous changes in heart rate, are prone to misdiagnosis due to isolated artifacts, arrhythmias, and other temporary disturbances. They lack a macroscopic understanding of the persistent state of arrhythmias, affecting the specificity of identification. Other methods focus on the overall statistical characteristics of long-term data. While this helps confirm persistent atrial fibrillation, the averaging effect over long periods may smooth out transient, paroxysmal atrial fibrillation events, thus reducing detection sensitivity and leading to missed diagnoses.

[0006] When attempting to combine results from different analytical dimensions, existing technologies often employ relatively simple fusion strategies, such as using fixed voting mechanisms or basic logic gates. These rigid fusion methods struggle to make dynamic, weighted decisions based on the confidence levels of different model outputs, failing to effectively balance the sensitivity and specificity of detection in complex clinical scenarios and limiting further improvements in the overall diagnostic performance of the algorithm. Summary of the Invention

[0007] The purpose of this application is to provide a multi-feature value and multi-model fusion method for atrial fibrillation identification, which solves the problem that existing technologies in automatic atrial fibrillation identification have limited feature description capabilities at a single time scale and simple model fusion strategies, making it difficult to simultaneously ensure high sensitivity and high specificity in complex cardiac rhythm scenarios.

[0008] Firstly, this application provides a method for atrial fibrillation identification using multi-feature value and multi-model fusion, the method comprising: Obtain the RR interval sequence of the electrocardiogram signal to be analyzed; Based on the RR interval sequence within a preset short time period, a short-time feature set is extracted, and the short-time feature set is input into a trained short-time model to obtain a short-time recognition result for the preset short time period; Based on the RR interval sequence within a preset long time period, a long-term feature set is extracted, and the long-term feature set is input into a trained long-term model to obtain a long-term recognition result for the preset long time period. The long-term identification results and the multiple short-term identification results contained therein are fused using model results to determine the final atrial fibrillation identification result.

[0009] In one optional embodiment, the preset short period lasts for 20 seconds and the preset long period lasts for 10 minutes.

[0010] In an optional embodiment, the short-time feature set includes: the mean of the RR interval, the minimum of the RR interval, the sample entropy of the RR interval, and the entropy of the RR interval fuzzy measure.

[0011] In an optional embodiment, the RR-interval fuzzy measure entropy (FuzzyME) is calculated as the sum of the local fuzzy measure entropy (FuzzyLME) and the global fuzzy measure entropy (FuzzyGME). The calculation steps include: For RR interval vectors with embedding dimension m, construct local baseline removal vector sets and global baseline removal vector sets respectively; Calculate the Chebyshev distance between any two vectors in the vector group to obtain the local distance and the global distance; By using a preset fuzzy membership function, the local distance and global distance are converted into local similarity and global similarity, respectively. Calculate the average local similarity when the embedding dimension is m. and average global similarity Increase the embedding dimension to m+1 and repeat the above steps to obtain the average local similarity. and average global similarity The local fuzzy measure entropy and the global fuzzy measure entropy are calculated based on the following formula: Finally, the entropy of the RR interval fuzzy measure is obtained: FuzzyME=FuzzyLME+FuzzyGME; In an optional embodiment, the step of extracting the long-term feature set includes: Normalization is performed on the RR interval sequence over the long period; based on the normalized RR interval sequence, the coordinates of the data points are generated as (x′). i ,x′ i+1 Poincaré scatter plot; The coordinate plane of the Poincaré scatter plot is divided into a plurality of preset triangular regions; The number of data points falling into each triangular region is counted, and the proportion of this number in the total number of data points is calculated. All proportions are combined into a regional probability feature, which is used as the long-term feature set.

[0012] In an optional embodiment, both the short-time model and the long-time model are decision tree models.

[0013] In an optional embodiment, the step of fusing the model results includes: If the long-term identification result is determined to be atrial fibrillation, then the final atrial fibrillation identification result corresponding to the long period of time is determined to be atrial fibrillation. If the long-term identification result is determined to be non-atrial fibrillation, then the proportion of short-term identification results that are determined to be atrial fibrillation within the long-term period is calculated. The final atrial fibrillation identification result is determined based on the comparison between the ratio and the first preset threshold θ1 and the second preset threshold θ2.

[0014] Specifically, if the ratio is higher than θ1, then the final atrial fibrillation identification result corresponding to the long period of time is determined to be atrial fibrillation; If the ratio is between θ2 and θ1, then the short-term identification result of each short period within the long period is directly used as its corresponding final atrial fibrillation identification result.

[0015] Secondly, this application provides a multi-feature value multi-model fusion atrial fibrillation identification system, which includes: an RR interval sequence acquisition module for acquiring the RR interval sequence of the electrocardiogram signal to be analyzed; The short-time model processing module is used to extract a short-time feature set based on the RR interval sequence within a preset short time period, and input the short-time feature set into the short-time model to obtain a short-time recognition result; The long-term model processing module is used to extract a long-term feature set based on the RR interval sequence within a pre-discarded long time period, and input the long-term feature set into the long-term model to obtain the long-term recognition result; The model result fusion module is used to fuse the long-term identification results and the multiple short-term identification results contained therein to determine the final atrial fibrillation identification result.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This application extracts the entropy of the RR interval fuzzy measure as a key feature of the short-time model. This feature simultaneously considers the local dynamic changes and global trends of the RR interval sequence, enabling a more detailed characterization of the rhythm irregularity of the heart rhythm. Compared to methods that only use single-scale statistical features, the introduction of this multi-dimensional feature helps improve the short-time model's ability to distinguish between atrial fibrillation and non-atrial fibrillation rhythms, thus providing a foundation for the accuracy of the overall identification method. 2. This application constructs a long-term model based on a Poincaré scatter plot, dividing the scatter plot into multiple specific regions and using the distribution probability of data points in these regions as features. This method can capture the overall pattern of long-term heart rate variability from a macroscopic perspective, effectively identifying the characteristics of persistent atrial fibrillation. The introduction of the long-term model provides a stable global basis for the identification process, helping to avoid misjudgments caused by transient arrhythmias or artifacts; 3. The result fusion strategy of this application for long-term and short-term models means that when the long-term model determines that the condition is not atrial fibrillation, the results of the short-term model will be further referenced. This hierarchical fusion mechanism allows this method to utilize the stability of the long-term model to confirm sustained rhythms, while also leveraging the sensitivity of the short-term model to identify paroxysmal and short-term atrial fibrillation events. This reduces the overall misjudgment rate while maintaining the ability to identify different types of atrial fibrillation. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method used in this application; Figure 2 This is the system architecture diagram of this application; Figure labeling: 10, RR interval sequence acquisition module; 20, short-term model processing module; 30, long-term model processing module; 40, model result fusion module. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1 -Appendix Figure 2 This application will be described in further detail below.

[0019] Example 1: A multi-feature value, multi-model fusion method for atrial fibrillation identification, see attached figure. Figure 1 The method includes: S100: Obtain the RR interval sequence of the ECG signal to be analyzed; The input for this step is a digitized dynamic electrocardiogram (ECG) signal, which is a time-amplitude sequence acquired at a preset sampling frequency. To accurately obtain the RR interval, this step specifically includes signal preprocessing, QRS complex detection, R-wave peak localization, and interval calculation.

[0020] First, the raw input ECG signal is preprocessed to suppress noise and enhance the characteristics of the QRS complex. Preprocessing includes baseline drift correction and high-frequency noise removal. Baseline drift correction is achieved by applying a digital high-pass filter to remove low-frequency components introduced by the subject's breathing or body movement. High-frequency noise removal is achieved by applying a digital low-pass filter to attenuate high-frequency components such as electromyographic interference.

[0021] Next, QRS group detection is performed on the preprocessed ECG signal. This process utilizes the steep morphological characteristics of the QRS group. Specifically, the signal is differentially analyzed to obtain its first derivative, thereby highlighting the slope change of the signal; the differentially analyzed signal is then squared point by point to make all data points positive and further amplify the high-slope portion; a moving window integrator is applied to the squared signal for smoothing, generating a characteristic signal in which the QRS group of the original ECG signal appears as a concentrated pulse wave.

[0022] Next, R-wave peak localization is performed on the characteristic signal. This process uses a dynamic thresholding method to identify valid QRS complex pulses. The system maintains a signal threshold and a noise threshold, which are dynamically updated based on the amplitudes of recently detected signal and noise pulses. When a pulse in the characteristic signal exceeds the current signal threshold, that pulse is identified as a candidate QRS complex. To eliminate interference from artifacts such as T waves, the system also checks whether the time interval between the candidate pulse and the previously confirmed R wave is within a physiologically reasonable range.

[0023] Finally, within each original EG signal segment confirmed as a valid QRS complex, a local maximum point is searched, and the sampling index of this point is recorded as the timestamp of the R-wave peak. Through complete processing of an ECG signal segment, an R-wave peak timestamp sequence T = {t1, t2, ..., t...} is obtained.N+1 The difference x between adjacent timestamps in the sequence is calculated. i =t i+1 -t i This will give us the final RR interval sequence X = {x1, x2, ..., x...} N The unit is typically milliseconds (ms). This RR interval sequence will serve as input data for all subsequent feature extraction steps.

[0024] S200. Based on the RR interval sequence within a preset short time period, extract a short-time feature set and input the short-time feature set into a trained short-time model to obtain short-time recognition results for the preset short time period. This process extracts data segments sequentially from the complete RR interval sequence in preset 20-second intervals. For each 20-second RR interval segment X = {x1, x2, ..., x...} N Module 20 performs short-time feature set extraction and inputs the extracted feature set into a pre-trained short-time model to obtain the short-time recognition result of the segment. The short-time feature set extraction process includes calculating four independent features: the first feature is the mean RR interval μ. RR The calculation formula is as follows: Where N is the total number of RR intervals within the 20-second segment, x i It is the duration of the i-th RR interval.

[0025] The second characteristic is the minimum RR interval x. min It is obtained by iterating through all the RR interval values ​​within the 20-second segment and finding the minimum value among them.

[0026] The third feature is the RR interval sample entropy, SampEn. In this embodiment, its calculation parameters are set as follows: embedding dimension m = 2, similarity tolerance r = 0.2. The calculation steps are as follows: First, the one-dimensional RR interval segment X is reconstructed into N-m+1 m-dimensional vectors. in

[0027] Secondly, define any two vectors and Distance between Chebyshev distance is the maximum absolute value of the difference between corresponding elements of a vector.

[0028] Then, for each vector Count other vectors whose distance is not greater than r. The quantity, denoted as C i Based on C i calculate

[0029] Next, for all Calculate the average to obtain the template matching probability B. m (r).

[0030] Increase the embedding dimension to m+1=3 and repeat the above process to obtain the template matching probability A when the dimension is 3. m (r).

[0031] Finally, the sample entropy value of this RR interval segment is calculated using the following formula: The fourth feature is the RR interval fuzzy measure entropy (FuzzyME). Its calculation also assumes an embedding dimension of m = 2. The calculation steps are as follows: First, reconstruct an m-dimensional vector from the RR interval segment X. To address this, local baseline removal vector sets and global baseline removal vector sets are constructed separately. Local baseline removal is performed by extracting data from each vector... Subtract its own mean from the middle Implementation. Global baseline removal is achieved by extracting data from each vector. Subtract the global mean μ of the entire 20-second segment from the middle RR accomplish.

[0032] Secondly, within both the local and global baseline-removed vector sets, the Chebyshev distance between any two vectors is calculated to obtain the local distance d. L (i,j) and global distance d G (i,j).

[0033] Then, these distance values ​​are input into a preset exponential fuzzy membership function to calculate the local similarity S. L (i,j) and global similarity S G (i,j): and The function parameter is set as follows: n L =3,n G =2. Similarity radius r L and r G It is related to the standard deviation σ of the current RR interval, specifically r L =0.2σ and r G =0.2σ, where σ is the standard deviation of all RR interval values ​​within the 20-second RR interval segment currently being analyzed.

[0034] Next, the average similarity of all vector pairs is calculated to obtain the average local similarity when the embedding dimension is m. and average global similarity

[0035] Increase the embedding dimension to m+1=3, and repeat the process of removing the baseline, calculating the distance, calculating the similarity, and averaging to obtain... and

[0036] Finally, the local fuzzy measure entropy (FuzzyLME) and the global fuzzy measure entropy (FuzzyGME) are calculated separately, and then added together to obtain the final RR interval fuzzy measure entropy value: FuzzyME=FuzzyLME+FuzzyGME; After completing the above four features (μ) RR ,x min After calculating SampEn and FuzzyME, the short-time model processing module 20 combines these four values ​​into a four-dimensional feature vector. This feature vector is input into a pre-trained decision tree classifier. The classifier judges the input feature vector according to its internal decision nodes and rules, and outputs a binary classification result, which is the short-time identification result of the 20-second RR interval segment. For example, the value "1" represents the atrial fibrillation state, and the value "0" represents the non-atrial fibrillation state.

[0037] S300. Based on the RR interval sequence within a preset long time period, extract a long-term feature set and input the long-term feature set into a trained long-term model to obtain long-term recognition results for the preset long time period. This process extracts data segments from the complete RR interval sequence in preset 10-minute intervals. For each 10-minute RR interval data segment X = {x1, x2, ..., x...} N}, Module 30 performs the extraction of long-term feature sets and inputs the extracted feature sets into a pre-trained long-term model to obtain the long-term recognition result of the data segment.

[0038] The extraction of long-term feature sets begins with normalization of the 10-minute RR interval data segment. This operation aims to eliminate the influence of individual baseline heart rate differences. Normalization is performed by x-intervals for each RR interval value in the data segment. i This is achieved by dividing by the maximum RR interval value max(X) in the data segment: This operation maps all RR interval values ​​to the interval [0,1], generating a normalized RR interval sequence X′.

[0039] Next, a Poincaré scatter plot is constructed based on the normalized sequence X′. This plot is a two-dimensional scatter plot, where each data point P... i The coordinates are determined by a pair of continuous normalized RR interval values ​​(x′). i ,x′ i+1 It is determined that a total of N-1 data points are generated.

[0040] Then, the [0,1]×[0,1] coordinate plane containing the Poincaré scatter plot is partitioned. This plane is divided into 20 pre-defined, non-overlapping triangular regions. To make this partitioning method more explicit, a specific partitioning method is as follows: Define 10 lower triangular regions with the x-axis as their base. For the k-th region (k = 1, ..., 10), the coordinates of its three vertices are ((k-1)×0.1, 0), (k×0.1, 0) and ((k-0.5)×0.1, (k-0.5)×0.1).

[0041] Define 10 left-angled triangular regions with the y-axis as their base. For the k-th region (k = 1, ..., 10), the coordinates of its three vertices are (0, (k-1) × 0.1), (0, k × 0.1), and ((k-0.5) × 0.1, (k-0.5) × 0.1).

[0042] The 20 triangular regions defined in this way, with their vertices located on the x-axis, y-axis, or main diagonal y=x, can completely and non-overlap cover the main distribution area in the scatter plot.

[0043] Finally, the regional probability features are calculated. The long-term model processing module iterates through all N-1 data points in the Poincaré scatter plot, determining the probability characteristics of each data point (x′). i ,x′ i+1 The triangular region to which ) belongs. For each region R... k (Where k = 1, ..., 20) Set a counter, and when a data point falls into this region, its corresponding counter N k Add one. After traversing all data points, calculate the proportion P of data points in each region using the following formula. k : These 20 percentage values ​​{P1, P2, ..., P 20 Together, they form a 20-dimensional feature vector, which is the long-term feature set used to describe the 10-minute RR interval data segment.

[0044] After extracting the long-term feature set, the long-term model processing module 30 inputs the 20-dimensional feature vector into a pre-trained decision tree classifier. This classifier uses its internal decision rules to classify the input vector and outputs a single binary classification result as the long-term identification result for the 10-minute data segment. For example, a value of "1" represents atrial fibrillation, and a value of "0" represents non-atrial fibrillation.

[0045] S400 performs model result fusion on the long-term identification results and the multiple short-term identification results contained therein to determine the final atrial fibrillation identification result.

[0046] The model result fusion module 40 performs a fusion operation on the recognition results generated in steps S200 and S300. For each 10-minute data segment, the input to this module includes: The system generates one long-term identification result from the long-term model processing module 30, and thirty short-term identification results corresponding to the long time period, generated by the short-term model processing module 20. The purpose of this module is to determine the final atrial fibrillation identification result for each heartbeat within the 10-minute data segment based on a set of preset rules.

[0047] In this embodiment, the fusion rule relies on two preset numerical thresholds: the first preset threshold θ1 is set to 0.8, and the second preset threshold θ2 is set to 0.4. The specific logic of the fusion process is as follows: First, the model result fusion module 40 checks the long-term recognition results of the 10-minute data segment.

[0048] If the long-term identification result is determined to be atrial fibrillation (for example, the value is "1"), then the fusion logic directly determines that the final atrial fibrillation identification result of all heartbeats in the 10-minute data segment is atrial fibrillation.

[0049] If the long-term identification result determines that the patient is not in atrial fibrillation (e.g., the value is "0"), the fusion logic proceeds to a secondary evaluation process based on the short-term identification result. In this process, module 40 first calculates the proportion R of the 20-second segments within the 10-minute data segment that were identified as atrial fibrillation by the short-term model. af This ratio is obtained by summing thirty short-term identification results (with values ​​of 0 or 1) and then dividing by 30.

[0050] Next, the calculated ratio R af Compare with two preset thresholds: If R af Greater than the first preset threshold θ1 (i.e., R) af If the value is greater than 0.8, it indicates that the vast majority of short-term segments exhibit atrial fibrillation characteristics. In this case, the model result fusion module 40 determines the final atrial fibrillation identification result of all heartbeats within the 10-minute data segment as atrial fibrillation.

[0051] If R af Less than or equal to the first preset threshold θ1 but greater than or equal to the second preset threshold θ2 (i.e., 0.4 ≤ R) af If the value is ≤0.8, it indicates the presence of a significant but not overwhelming proportion of short-term atrial fibrillation characteristic segments. In this case, module 40 sets the final atrial fibrillation identification result of each 20-second segment within the 10-minute data segment as its original short-term identification result. That is, if the short-term identification result of a short segment is "1", then its final identification result is "1"; if it is "0", then the final identification result is "0".

[0052] If R af Less than the second preset threshold θ2 (i.e., R) af If the value is less than 0.4, it indicates that the proportion of short-term segments with atrial fibrillation characteristics is low. In this case, the model result fusion module 40 determines the final atrial fibrillation identification result of all heartbeats within this 10-minute data segment as non-atrial fibrillation.

[0053] By executing the above fusion logic, the model result fusion module 40 generates the final atrial fibrillation identification result sequence for each 10-minute data segment, either per heartbeat or per short time period, thereby completing the entire atrial fibrillation identification process.

[0054] Example 2: Refer to Appendix Figure 2 This application provides a multi-feature value, multi-model fusion atrial fibrillation identification system, which includes: The RR interval sequence acquisition module 10 is used to acquire the RR interval sequence of the electrocardiogram signal to be analyzed. The RR interval sequence acquisition module 10 is the system's input interface. Its function is to receive a continuous raw dynamic electrocardiogram signal from an external source and extract the RR interval sequence from it. The output of this module, namely the RR interval sequence, is simultaneously transmitted to the short-time model processing module 20 and the long-time model processing module 30.

[0055] The short-time model processing module 20 is used to extract a short-time feature set based on the RR interval sequence within a preset short time period, and input the short-time feature set into the short-time model to obtain short-time recognition results; The short-time model processing module 20 segments the sequence using a preset short time period (e.g., 20 seconds) as the processing unit. For each short time period, the module extracts a multi-dimensional short-time feature set, including the entropy of the RR interval fuzzy measure, and generates a short-time recognition result using an internal short-time model. Therefore, for a long time period, the module generates a sequence consisting of multiple short-time recognition results.

[0056] The long-term model processing module 30 is used to extract long-term feature sets based on the RR interval sequence within a pre-rounded long time period, and input the long-term feature sets into the long-term model to obtain long-term recognition results. The long-term model processing module 30 uses a preset long time period (e.g., 10 minutes) as the processing unit. For each long time period, the module extracts a multi-dimensional long-term feature set that is different from the short-term feature set in both dimension and composition, based on the probability distribution of the Poincaré scatter plot region, and uses an internal long-term model to generate a single long-term recognition result.

[0057] The model result fusion module 40 is used to fuse the long-term identification results and the multiple short-term identification results contained therein to determine the final atrial fibrillation identification result.

[0058] The model result fusion module 40 functions to aggregate and process the results from the first two processing modules. This module receives a single long-term identification result generated by the long-term model processing module 30, and multiple short-term identification result sequences corresponding to the long time period, generated by the short-term model processing module 20. The model result fusion module 40 contains a set of preset fusion rules and numerical thresholds. Based on the statistical ratio of the long-term and short-term identification results, it executes decision logic to ultimately generate and output the final atrial fibrillation identification result for the long time period.

[0059] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for atrial fibrillation identification using multi-feature value and multi-model fusion, characterized in that, Includes the following steps: Obtain the RR interval sequence of the electrocardiogram signal to be analyzed; Based on the RR interval sequence within a preset short time period, a short-time feature set is extracted, and the short-time feature set is input into a short-time model to obtain a short-time recognition result; Based on the RR interval sequence within a preset long time period, a long-term feature set is extracted, and the long-term feature set is input into a long-term model to obtain a long-term recognition result; The short-term and long-term identification results are fused using model results to determine the final atrial fibrillation identification result.

2. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, The short-time feature set includes: the mean of the RR interval, the minimum of the RR interval, the sample entropy of the RR interval, and the fuzzy measure entropy of the RR interval.

3. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, The RR interval fuzzy measure entropy is calculated as the sum of the local fuzzy measure entropy and the global fuzzy measure entropy: FuzzyME=FuzzyLME+FuzzyGME; The local fuzzy measure entropy and the global fuzzy measure entropy are calculated based on the local debasement vector group and the global debasement vector group of the RR interval sequence, respectively.

4. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, The steps for extracting long-term feature sets include: The RR interval sequence over the long period is normalized. Construct a Poincaré scatter plot based on the normalized RR interval sequence; The Poincaré scatter plot is divided into multiple preset regions; The percentage of data points falling into each preset region in the Poincaré scatter plot is statistically analyzed to form regional probability features, which are then used as the long-term feature set.

5. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, The step of dividing the Poincaré scatter plot into multiple preset regions is as follows: The coordinate plane of the Poincaré scatter plot is divided into a plurality of triangular regions.

6. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, The steps for fusing the model results include: When the long-term identification result is atrial fibrillation, the final atrial fibrillation identification result is determined to be atrial fibrillation; When the long-term identification result is not atrial fibrillation, the final atrial fibrillation identification result is determined based on the proportion of atrial fibrillation among the multiple short-term identification results included in the long-term period, combined with a preset threshold.

7. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, When the long-term identification result is not atrial fibrillation, the specific steps to determine the final atrial fibrillation identification result are as follows: If the ratio is higher than the first preset threshold θ1, then the final atrial fibrillation identification result is determined to be atrial fibrillation; If the ratio is between the second preset threshold θ2 and the first preset threshold θ1, then the short-term identification result is taken as the final atrial fibrillation identification result.

8. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, The short-time model or the long-time model is a decision tree model.

9. The atrial fibrillation identification method based on multi-feature value and multi-model fusion according to claim 1, characterized in that, The preset short period lasts for 20 seconds, and the preset long period lasts for 10 minutes.

10. A multi-feature value multi-model fusion atrial fibrillation identification system, based on the multi-feature value multi-model fusion atrial fibrillation identification method according to any one of claims 1-9, characterized in that, include: The RR interval sequence acquisition module is used to acquire the RR interval sequence of the electrocardiogram signal to be analyzed. The short-time model processing module is used to extract a short-time feature set based on the RR interval sequence within a preset short time period, and input the short-time feature set into the short-time model to obtain a short-time recognition result; The long-term model processing module is used to extract a long-term feature set based on the RR interval sequence within a preset long time period, and input the long-term feature set into the long-term model to obtain the long-term recognition result; The model result fusion module is used to fuse the short-term identification results and the long-term identification results to determine the final atrial fibrillation identification result.

Images