A risk assessment system based on myocardial ischemia and storage medium

Abstract

The present application relates to the technical field of electrocardiosignal processing, and particularly relates to a risk assessment system based on myocardial ischemia and a storage medium. A computer program in the storage medium executes the following steps: obtaining original electrocardiosignal discrete sequences of multiple leads, and determining optimal decomposition parameters according to the original electrocardiosignal discrete sequences; performing variational mode decomposition on the original electrocardiosignal discrete sequences based on the optimal decomposition parameters to obtain pure electrocardiosignal sequences of the leads; performing heartbeat segmentation on the pure electrocardiosignal sequences to obtain effective heartbeat sequences of the leads; performing feature extraction on the effective heartbeat sequences to obtain multi-dimensional feature vectors; inputting the multi-dimensional feature vectors into a pre-constructed classification model to obtain a classification result and a probability. The present application solves the problem that the risk assessment system for myocardial ischemia in the prior art is not accurate enough, and improves the accuracy of classification by performing variational mode decomposition on the original electrocardiosignal discrete sequences with optimal decomposition parameters and constructing multi-dimensional feature vectors.

Description

Technical Field

[0001] This invention relates to the field of electrocardiogram signal processing technology, specifically a risk assessment system and storage medium based on myocardial ischemia. Background Technology

[0002] As a key bioelectrical signal characterizing the electrophysiological activity of the human heart, electrocardiogram (ECG) signals are an important research object in the fields of biomedical information detection, human physiological state assessment, intelligent health monitoring, and medical electronic equipment development. They have wide application value in scenarios such as daily physiological monitoring, portable health testing devices, and remote physiological data acquisition.

[0003] However, the ECG signal acquisition process is susceptible to interference from various factors such as environment, equipment, and human movement, resulting in poor signal quality stability. The collaborative processing of multi-lead ECG signals is challenging, with insufficient adaptive processing capabilities in signal denoising and feature extraction, making it difficult to fully preserve the effective physiological information in the signal. Furthermore, existing ECG signal analysis systems have limited intelligence, insufficient mining and utilization of multi-dimensional signal features, weak model adaptability and generalization ability, and overall system processing accuracy, operational stability, and practicality are insufficient to meet current requirements for high efficiency, precision, and stability.

[0004] Therefore, there is an urgent need for a more accurate and stable myocardial ischemia risk assessment system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a risk assessment system and storage medium based on myocardial ischemia, which solves the problem that existing risk assessment systems for myocardial ischemia are not accurate enough.

[0006] To achieve the above objectives, one aspect of the present invention provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the following steps: acquiring raw discrete electrocardiogram (ECG) sequences of multiple leads and determining optimal decomposition parameters based on the raw ECG discrete sequences; performing variational mode decomposition on the raw ECG discrete sequences based on the optimal decomposition parameters to obtain a clean ECG sequence of the leads; performing heartbeat segmentation on the clean ECG sequence to obtain a valid heartbeat sequence of the leads; extracting features from the valid heartbeat sequence to obtain a multidimensional feature vector; and inputting the multidimensional feature vector into a pre-constructed classification model to obtain a classification result and probability.

[0007] This invention acquires multi-lead raw discrete ECG sequences and determines the optimal decomposition parameters to accurately adapt to the characteristics of ECG signals. Based on the optimal parameters, variational mode decomposition is performed to effectively remove noise and retain true ECG features. Heartbeat segmentation is performed on the pure ECG sequences to eliminate artifact heartbeats and ensure data quality. Multidimensional features of clinical anatomical location are extracted and integrated from the effective heartbeat sequences to comprehensively capture key information of myocardial ischemia and improve the accuracy of myocardial ischemia risk assessment.

[0008] Optionally, determining the optimal decomposition parameters based on the original discrete ECG sequence includes: obtaining a range of decomposition parameters; determining an initial population based on the range of decomposition parameters; performing variational mode decomposition on the original discrete ECG sequence based on individuals in the initial population and filtering to obtain denoised mode components of the individuals; calculating the maximum permutation entropy and maximum frequency aliasing of the individuals using the denoised mode components; performing non-dominated sorting on the individuals with the objectives of minimizing the maximum permutation entropy and the maximum frequency aliasing, respectively; iteratively obtaining the Pareto front using a multi-objective optimization algorithm based on the results of the non-dominated sorting; and determining the optimal decomposition parameters based on the Pareto front.

[0009] This invention obtains the range of decomposition parameters and determines the initial population to fully cover the parameter search space. Based on the variational mode decomposition of the original discrete ECG sequence by individuals, it obtains pure denoised mode components. It calculates the maximum permutation entropy and the maximum frequency aliasing to quantify the signal noise and mode aliasing levels. It completes the non-dominated sorting of individuals by minimizing the dual objectives. It obtains the Pareto front through multi-objective optimization iteration, and accurately determines the optimal decomposition parameters accordingly, thereby improving the accuracy of the optimal decomposition parameters.

[0010] Optionally, the step of performing variational mode decomposition and filtering on the original discrete ECG sequence based on individuals in the initial population to obtain the denoised mode components of the individuals includes: performing variational mode decomposition on the original discrete ECG sequence based on individuals in the initial population to obtain multiple initial mode components; obtaining the ECG data frequency range based on pre-acquired historical ECG data; and using the ECG data frequency range to filter the initial mode components to obtain denoised mode components.

[0011] This invention uses variational mode decomposition of the original discrete ECG sequence by an initial population of individuals to adaptively decompose the signal and fully preserve each mode component. It determines the frequency range of ECG data based on historical ECG data to conform to the inherent characteristics of clinical ECG and to clarify the standard screening interval. It uses this frequency range to screen the initial mode components to accurately remove noise components and retain pure and effective ECG components, thereby improving the data accuracy of the denoised mode components.

[0012] Optionally, the step of calculating the maximum permutation entropy and maximum frequency aliasing of the individual using the denoised mode components includes: calculating the initial permutation entropy of the denoised mode components and extracting the maximum value from the initial permutation entropy to obtain the maximum permutation entropy of the individual; calculating the initial frequency aliasing between two adjacent denoised mode components and extracting the maximum value from the initial frequency aliasing to obtain the maximum frequency aliasing of the individual.

[0013] This invention calculates the initial permutation entropy of the denoised mode components and extracts the maximum value to obtain the individual maximum permutation entropy, which accurately quantifies the signal noise and complexity level. It also calculates the initial frequency aliasing of adjacent denoised mode components and extracts the maximum value to obtain the individual maximum frequency aliasing, which intuitively reflects the degree of overlap in the frequency domain of modes and improves the objectivity and accuracy of the quantitative evaluation of variational mode decomposition effect.

[0014] Optionally, determining the optimal decomposition parameters based on the Pareto front includes: obtaining a reconstructed discrete ECG sequence based on each non-dominated individual in the Pareto front; calculating the reconstructed signal-to-noise ratio (SNR) based on the reconstructed discrete ECG sequence and the original discrete ECG sequence; extracting the minimum SNR from the reconstructed SNR; and determining the optimal decomposition parameters based on the minimum SNR.

[0015] This invention obtains reconstructed discrete ECG sequences based on Pareto front non-dominated individuals to retain high-quality decomposition results, calculates the reconstructed signal-to-noise ratio (SNR) by combining the original discrete ECG sequences to objectively measure the denoising and signal fidelity effects, and extracts the minimum SNR to determine the optimal decomposition parameters to accurately lock the optimal parameter combination for denoising, thereby further improving the accuracy of optimal decomposition parameter selection.

[0016] Optionally, the step of calculating the reconstructed signal-to-noise ratio based on the reconstructed discrete ECG sequence and the original discrete ECG sequence includes: calculating the signal energy of the original discrete ECG sequence; calculating the mean square error of the reconstructed discrete ECG sequence and the original discrete ECG sequence; and calculating the ratio of the signal energy to the mean square error to obtain the reconstructed signal-to-noise ratio.

[0017] This invention determines the signal energy benchmark by calculating the signal energy of the original discrete ECG sequence, calculates the mean square error between the reconstructed discrete ECG sequence and the original discrete ECG sequence to accurately quantify waveform distortion and amplitude deviation, and calculates the ratio of signal energy to mean square error to obtain the reconstructed signal-to-noise ratio to objectively measure the denoising and fidelity effect, thereby improving the accuracy and objectivity of the reconstructed signal-to-noise ratio calculation.

[0018] Optionally, the step of segmenting the pure ECG sequence to obtain the effective heartbeat sequence of the lead includes: performing R-wave detection on the pure ECG sequence to obtain multiple R-wave position sequences of the lead; fusing the R-wave position sequences to obtain multiple global R-wave positions; segmenting the pure ECG sequence based on the global R-wave positions to obtain multiple single heartbeat segments of the lead; obtaining a normal heartbeat duration range; filtering the single heartbeat segments based on the normal heartbeat duration range to obtain a filtered heartbeat sequence of the lead; and performing morphological matching on the filtered heartbeat sequence based on a preset heartbeat template correlation coefficient threshold to obtain the effective heartbeat sequence.

[0019] This invention obtains R-wave position sequences for each lead by separately detecting R-waves in a clean electrocardiogram sequence to accurately locate the beat point. It then merges the R-wave position sequences to obtain the global R-wave position, unifying the multi-lead beat benchmark and ensuring timing alignment. The global R-wave position is used to segment single beat segments to extract the complete beat waveform. Abnormal beats are eliminated by filtering according to the normal beat duration range. Finally, beat template morphology matching is used to filter out artifact beats and retain valid beats, thus improving the accuracy and reliability of beat segmentation.

[0020] Optionally, the feature extraction of the effective heartbeat sequence to obtain a multidimensional feature vector includes: the effective heartbeat sequence includes multiple effective heartbeat segments; extracting ST segment intervals and isoelectric line intervals from the effective heartbeat segments; calculating the amplitude difference between the ST segment intervals and the isoelectric line intervals to obtain the ST segment offset set of the leads; based on the pre-acquired anatomical location of myocardial ischemia, dividing the ST segment offset set into anterior wall lead group, inferior wall lead group, and lateral wall lead group; calculating the maximum offset value and offset range of the anterior wall lead group, the inferior wall lead group, and the lateral wall lead group respectively to form the intra-lead group features of the leads; calculating the offset difference between the pre-set clinically corresponding mirror lead pairs between the anterior wall lead group and the lateral wall lead group to form the inter-lead group features of the leads; and constructing a multidimensional feature vector based on the intra-lead group features and the inter-lead group features.

[0021] This invention extracts the ST segment and isoelectric line intervals from effective cardiac beat segments and calculates the amplitude difference to obtain the ST segment offset set, accurately capturing the core waveform changes of myocardial ischemia. Based on the anatomical location of myocardial ischemia, lead groups are divided to conform to clinical diagnostic logic. The maximum value and range of offset in each group are calculated to form intra-group features, and the offset difference of mirror leads is calculated to form inter-group features. The degree of ischemic area and spatial differences are comprehensively quantified. The intra-group and inter-group features are integrated to construct a multi-dimensional feature vector, which improves the pertinence, clinical relevance and comprehensiveness of ECG feature extraction.

[0022] Optionally, constructing a multidimensional feature vector based on the intra-lead group features and the inter-lead group features includes: calculating T-wave features and QTc intervals in the effective heartbeat segments; and constructing a multidimensional feature vector using the T-wave features, the QTc intervals, the intra-lead group features, and the inter-lead group features.

[0023] This invention supplements key ECG indicators for diagnosing myocardial ischemia by calculating T-wave characteristics and QTc interval in effective cardiac beat segments. Then, it integrates T-wave characteristics, QTc interval, intra-lead features, and inter-lead features to construct a multi-dimensional feature vector, which comprehensively enriches the feature dimensions, conforms to the logic of clinical ECG diagnosis, and improves the completeness and clinical adaptability of the multi-dimensional feature vector.

[0024] Another aspect of the present invention provides a risk assessment system based on myocardial ischemia, comprising an input device, a processor, an output device, and a memory, wherein the input device, processor, output device, and memory are interconnected, the memory includes the computer-readable storage medium described in the preceding aspect of the present invention, the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions.

[0025] The present invention provides a risk assessment system based on myocardial ischemia, which is compact, stable, highly integrated and simple in construction. It can stably execute the steps of the program instructions in the computer-readable storage medium provided in the preceding aspect of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0026] Figure 1 This is a flowchart of program instructions in a computer-readable storage medium according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a risk assessment system based on myocardial ischemia according to an embodiment of the present invention. Detailed Implementation

[0027] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0028] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0029] Please see Figure 1 In one embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the following steps: Step S1: Obtain the original discrete ECG sequences of multiple leads, and determine the optimal decomposition parameters based on the original discrete ECG sequences.

[0030] In this embodiment, the original discrete ECG sequence of multiple leads is a standard 12-lead discrete ECG sequence, which includes 6 limb leads and 6 chest leads. The limb leads are leads I, II, III, aVR, aVL, and aVF, which are used to reflect changes in electrical activity in the frontal plane of the heart. The chest leads are leads V1, V2, V3, V4, V5, and V6, which are used to reflect changes in electrical activity in the horizontal plane of the heart. The 12-lead continuous simulated ECG signal of the subject is synchronously acquired by the ECG acquisition device at a preset sampling rate. The continuous simulated signal is quantized into a discrete time sequence by analog-to-digital conversion, and finally the original discrete ECG sequence corresponding to each lead is obtained. This sequence completely preserves the original amplitude, temporal characteristics, and noise information such as baseline drift and electromyographic interference of the ECG signal.

[0031] The determination of the optimal decomposition parameters based on the original discrete electrocardiogram sequence specifically includes the following sub-steps: Step S101: Obtain the decomposition parameter range and determine the initial population based on the decomposition parameter range.

[0032] In this embodiment, firstly, for the core optimization parameters of Variational Mode Decomposition (VMD) (including the number of decomposition layers K, the penalty coefficient α, and the bandwidth constraint τ), combined with the inherent frequency domain characteristics of the 12-lead ECG signal and the engineering practice experience of ECG signal variational mode denoising, the upper and lower limits of the values ​​of each parameter are determined to construct a complete decomposition parameter range. Then, a random uniform initialization strategy is adopted to generate a preset number of parameter individuals within this parameter range. Each parameter individual is composed of a set of corresponding decomposition layer number, penalty coefficient, and bandwidth constraint combination. All parameter individuals together form an initial population that is evenly distributed and covers the complete search space.

[0033] Step S102: Based on the individuals in the initial population, perform variational mode decomposition on the original discrete ECG sequence and filter to obtain the denoised mode components of the individuals.

[0034] Specifically, the process of performing variational mode decomposition and screening on the original discrete electrocardiogram sequence based on individuals in the initial population to obtain the denoised mode components of the individuals includes the following sub-steps: Step S10201: Based on the individuals in the initial population, perform variational mode decomposition on the original discrete ECG sequence to obtain multiple initial mode components.

[0035] In this embodiment, the decomposition level, penalty coefficient, and bandwidth constraint corresponding to each parameter individual in the initial population are used as the core execution parameters of variational mode decomposition. Variational mode decomposition is performed on the original discrete ECG sequences of each of the 12 leads. Through the adaptive frequency domain separation and iterative optimization of the variational mode algorithm, the noisy and non-stationary original discrete ECG sequences are decomposed into multiple initial mode components that match the decomposition level and are independent of each other in frequency band. Each initial mode component is arranged in order from high to low signal frequency, thus completely preserving the effective feature components and noise components in the original ECG signal.

[0036] Step S10202: Obtain the frequency range of electrocardiogram data based on the previously acquired historical electrocardiogram data.

[0037] In this embodiment, a large amount of clinically collected historical ECG data, labeled as normal or typical ischemic by ECG experts, is first acquired. Each ECG data is preprocessed with bandpass filtering to remove obvious high-frequency electromyographic interference and low-frequency baseline drift. Then, Fourier transform or power spectrum estimation methods are used to calculate the energy spectrum of each preprocessed ECG signal. The energy spectra of all historical ECG data are then superimposed and averaged to obtain a statistical spectrum curve that reflects the overall energy distribution of the ECG signal. Finally, based on the frequency range where the energy proportion in the statistical spectrum curve exceeds a preset threshold (e.g., 95%), or in combination with the clinically recognized effective frequency range of ECG (usually 0.5Hz to 40Hz), the effective frequency range of ECG data required for variational mode decomposition is determined.

[0038] Step S10203: Use the frequency range of the electrocardiogram data to filter the initial modal components to obtain denoised modal components.

[0039] In this embodiment, the center frequency and dominant frequency distribution of each initial modal component obtained by variational mode decomposition are calculated respectively. The frequency parameters of each initial modal component are compared with the effective frequency range of the ECG data one by one. Modal components whose center frequency falls within the effective frequency range are retained, while high-frequency electromyographic interference components and low-frequency baseline drift components whose center frequency exceeds the range are removed. Finally, denoised modal components that retain only the effective ECG signal components are obtained, providing a pure signal basis for the subsequent calculation of permutation entropy and frequency aliasing.

[0040] Step S103: Calculate the maximum permutation entropy and maximum frequency aliasing of the individual using the denoised modal components.

[0041] The calculation of the maximum permutation entropy and maximum frequency aliasing of the individual using the denoised modal components specifically includes the following sub-steps: Step S10301: Calculate the initial permutation entropy of the denoised mode component, and extract the maximum value from the initial permutation entropy to obtain the maximum permutation entropy of the individual.

[0042] In this embodiment, for all denoised mode components corresponding to the current variational mode parameter individual, the initial permutation entropy of each component is calculated one by one; for any single-component time series, the embedding dimension is set. (Usually 3-7) and a time delay (usually 1), reconstructing the original time series into The set of delayed vectors in the phase space is then used to sort each delayed vector in ascending order of its value, thus obtaining the arrangement pattern of the vectors. All vectors are then counted. The probability distribution of each possible permutation pattern is obtained by calculating its frequency of occurrence. Then, calculate the initial permutation entropy value of the sequence according to the Shannon entropy formula. .

[0043] in, For the first The probability of occurrence of each permutation pattern is calculated, and an initial permutation entropy set is obtained after all components are calculated. The maximum value among them is selected as the maximum permutation entropy of the individual. Since permutation entropy can characterize signal complexity and noise level, taking the maximum value can constrain all modal components under this parameter individual to maintain a low complexity and low aliasing state, accurately characterizing the level of the component with the strongest noise and the highest complexity in the current decomposition result.

[0044] Step S10302: Calculate the initial frequency aliasing degree between two adjacent denoised mode components, and extract the maximum value from the initial frequency aliasing degree to obtain the maximum frequency aliasing degree of the individual.

[0045] In this embodiment, for all denoised mode components corresponding to the current individual, after sorting them from low to high center frequency, the initial frequency aliasing degree between each pair of adjacent mode components is calculated sequentially. Specifically, for any two adjacent denoised mode components, their power spectral density functions are obtained, the overlap area of ​​the two power spectral density curves is calculated, and the overlap area is divided by half of the sum of the total spectral areas of the two mode components to obtain the normalized frequency aliasing degree, which ranges from 0 to 1. The larger the value, the more severe the overlap between the two adjacent modes in the frequency domain, and the worse the mode separation effect. The above calculation is performed on all adjacent mode component pairs of the current individual to obtain a set of initial frequency aliasing degrees. Finally, the maximum value is extracted from the set of values ​​as the maximum frequency aliasing degree of the individual. Since variational mode decomposition requires that each mode be separated as much as possible in the frequency domain, the maximum frequency aliasing degree reflects the most severe mode overlap in the decomposition result of the individual.

[0046] Step S104: Perform non-dominated sorting on the individuals with the objectives of minimizing the maximum permutation entropy and the maximum frequency aliasing, respectively.

[0047] In this embodiment, for all individual parameters in the multi-objective optimization iteration process, the dual optimization objectives are minimizing the maximum permutation entropy and minimizing the maximum frequency aliasing. First, the two core objective values ​​(maximum permutation entropy and maximum frequency aliasing) corresponding to each individual are calculated. Then, based on the non-dominated sorting principle, the dominance relationship between any two individuals is determined one by one. That is, if the two objective values ​​of an individual are not greater than those of another individual, and at least one objective value is strictly less than that of another individual, then the individual dominates the other individual. Non-dominated individuals without any dominance are classified as the first non-dominated layer. The above dominance relationship determination is repeated for the remaining individuals, and the second, third and subsequent non-dominated layers are divided in turn. Finally, the non-dominated hierarchical sorting of all individual parameters is completed. This sorting method can accurately screen out high-quality variational modal parameter individuals that simultaneously meet the requirements of low complexity and low aliasing, providing a clear hierarchical basis for individual screening for subsequent acquisition of the Pareto front.

[0048] Step S105: Based on the results of the non-dominated sorting, the Pareto front is obtained by iteratively using a multi-objective optimization algorithm.

[0049] In this embodiment, based on the non-dominated ranking result of the current population, a multi-objective optimization algorithm with a crowding strategy is used to perform iterative optimization. In each iteration, individuals with better non-dominated levels are selected first, and individuals with excessively dense distribution are removed by combining crowding distance, retaining high-quality individuals with even distribution as parents. Then, simulated binary crossover and polynomial mutation operations are performed on the parent individuals to generate the corresponding offspring population. The parent and offspring populations are merged into a mixed population, and the non-dominated ranking and crowding calculation are re-performed on the mixed population. Then, a new population with the same size as the original population is selected according to the non-dominated level and distribution, completing one iteration. This iterative operation is repeated until the population no longer shows obvious objective function updates for several consecutive generations, or the preset maximum number of iterations is reached, at which point the optimization stops. Finally, the set of all the best individuals in the converged population that are located in the first non-dominated level and do not dominate each other is determined as the Pareto front obtained in this optimization.

[0050] Step S106: Determine the optimal decomposition parameters based on the Pareto front.

[0051] The determination of the optimal decomposition parameters based on the Pareto front specifically includes the following sub-steps: Step S10601: Obtain the reconstructed discrete ECG sequence based on each non-dominated individual in the Pareto front.

[0052] In this embodiment, for each non-dominated individual in the Pareto front, the individual has already completed signal decomposition and denoised mode component screening under the strain modal parameters during the previous multi-objective optimization iteration. Therefore, all the denoised mode components corresponding to the individual after screening are directly retrieved, and all effective denoised mode components are linearly superimposed and summed to directly obtain the reconstructed ECG discrete sequence corresponding to the non-dominated individual. There is no need to perform variational mode decomposition operation on the original ECG discrete sequence again.

[0053] Step S10602: Calculate the reconstructed signal-to-noise ratio based on the reconstructed discrete ECG sequence and the original discrete ECG sequence.

[0054] The calculation of the reconstructed signal-to-noise ratio based on the reconstructed discrete ECG sequence and the original discrete ECG sequence specifically includes the following sub-steps: Step S1060201: Calculate the signal energy of the original discrete electrocardiogram sequence.

[0055] In this embodiment, for the original discrete ECG sequence of each of the 12 leads, based on the statistical method of time-domain signal energy, the signal amplitude of all sampling times in the discrete sequence is traversed, and the square of each amplitude is calculated and then summed to obtain the total energy value of the original ECG signal of that lead. This value fully reflects the amplitude energy distribution of the original noisy ECG signal throughout the entire sampling period and serves as a benchmark reference value for subsequently measuring the fidelity and energy retention rate of the reconstructed signal.

[0056] Step S1060202: Calculate the mean square error between the reconstructed discrete ECG sequence and the original discrete ECG sequence.

[0057] In this embodiment, for each lead, the amplitude deviation between the reconstructed discrete ECG sequence and the original discrete ECG sequence is calculated point by point at the same sampling time. The mean square error index is used to characterize the waveform distortion and amplitude difference level of the reconstructed signal relative to the original signal. This index can objectively reflect the accuracy of the reconstruction of the effective features of the original ECG by the signals reconstructed by different variational mode parameters. It is a key quantitative basis for measuring the signal fidelity and distortion after denoising, and provides an important reference for selecting the optimal parameter individuals from the Pareto front.

[0058] Step S1060203: Calculate the ratio of the signal energy to the mean square error to obtain the reconstructed signal-to-noise ratio.

[0059] In this embodiment, the reconstructed signal-to-noise ratio (SNR) for a given parameter is obtained by dividing the total signal energy of the original discrete ECG sequence (NER) by the mean square error between the reconstructed discrete ECG sequence and the original sequence (MSE). Since the original signal energy is a constant, a smaller SNR indicates a greater difference between the reconstructed signal and the original noisy signal. Under the constraint of Pareto front-non-dominated individuals, this difference mainly originates from noise components such as effectively removed electromyographic interference and baseline drift. Therefore, a smaller SNR indicates a better denoising effect and more complete preservation of the true ECG signal.

[0060] Step S10603: Extract the minimum signal-to-noise ratio from the reconstructed signal-to-noise ratio, and determine the optimal decomposition parameters based on the minimum signal-to-noise ratio.

[0061] In this embodiment, the reconstructed signal-to-noise ratio (SNR) values ​​corresponding to all non-dominated individuals in the Pareto front are iterated. Since the reconstructed SNR is defined as the ratio of the signal energy of the original discrete ECG sequence to the mean square error of the reconstructed signal and the original signal in this scheme, and the original signal energy is a fixed constant, the smaller the reconstructed SNR, the greater the difference between the reconstructed signal and the original noisy signal. Under the constraint of non-dominated individuals in the Pareto front, this difference is only composed of noise components such as effectively removed electromyographic interference and baseline drift, and the denoising effect is better. Therefore, the result with the smallest reconstructed SNR value is selected from all non-dominated individuals, and the variational mode decomposition layer number, penalty coefficient, and bandwidth constraint carried by the individual with the smallest reconstructed SNR are determined as the optimal decomposition parameters for variational mode denoising of the ECG signal in this case.

[0062] Step S2: Perform variational mode decomposition on the original discrete ECG sequence based on the optimal decomposition parameters to obtain the pure ECG sequence of the lead.

[0063] In this embodiment, the optimal variational mode decomposition parameters obtained through multi-objective optimization are used to perform formal variational mode decomposition on the original discrete ECG sequences of each lead, resulting in multiple mode components with clear frequency domain division and minimal mutual interference. Then, based on the effective frequency range of the ECG, high-frequency noise and low-frequency drift components are removed, retaining the effective modes that reflect the true ECG characteristics. All effective mode components are linearly superimposed to finally obtain a pure ECG sequence with complete waveforms in each lead, low distortion, and significantly suppressed noise interference.

[0064] Step S3: Perform heartbeat segmentation on the pure ECG sequence to obtain the effective heartbeat sequence of the leads.

[0065] The specific steps for segmenting the pure electrocardiogram sequence to obtain the effective heartbeat sequence of the leads include the following: Step S301: R-wave detection is performed on the pure ECG sequence to obtain multiple R-wave position sequences of the leads.

[0066] In this embodiment, for the pure ECG sequence of each lead, a bandpass filter (usually set to 5Hz to 15Hz) is first used to preprocess the signal to enhance the frequency band energy of the QRS wave. Then, the Pan-Tompkins algorithm is used to calculate the slope, amplitude, and integral of the signal point by point, and the position of the R wave peak is determined by combining an adaptive threshold. For the detected candidate R wave peaks, false detection points are eliminated by double verification of the refractory period (usually set to 200ms) and the amplitude threshold. Finally, the R wave position sequence of each lead is output in chronological order. The R wave position sequence includes the R wave time corresponding to all valid heartbeats in the pure ECG sequence of that lead. Each lead is detected independently, thereby obtaining the R wave position sequence corresponding to each of the 12 leads.

[0067] Step S302: The R-wave position sequence is fused to obtain multiple global R-wave positions.

[0068] In this embodiment, after obtaining the R-wave position sequence of all leads, the R-wave positions of all leads are first merged into a total sequence in chronological order. Then, the total sequence is clustered, and multiple R-wave positions with time intervals less than a preset threshold (e.g., 50ms) are grouped into the same beat cluster. This threshold corresponds to the maximum allowable time deviation of the same beat between multiple leads. For each beat cluster, the median or average of all R-wave positions within the cluster is calculated as the global R-wave position of that beat. Finally, the global R-wave position sequence of all beats is output in chronological order. This sequence serves as a unified benchmark for subsequent beat segmentation, ensuring that the beats segmented between different leads are strictly aligned in time.

[0069] Step S303: Using the global R-wave position as a reference, the pure ECG sequence is segmented to obtain multiple single-beat segments of the lead.

[0070] In this embodiment, for each lead's clean ECG sequence, each R-wave position in the global R-wave position sequence is used as the central reference point for that beat. Based on this R-wave position, a preset first duration (e.g., 200ms) is truncated forward as the pre-QRS interval, and a preset second duration (e.g., 400ms) is truncated backward as the post-QRS interval, thereby extracting a single beat segment containing a complete P wave, QRS complex, and T wave. The above truncating operation is repeated for all leads and each global R-wave position to obtain multiple single beat segments corresponding to each lead. Segments belonging to the same beat in different leads are consistent in time, providing aligned single beat data for subsequent beat screening and feature extraction.

[0071] Step S304: Obtain the normal heartbeat duration range, and filter the single heartbeat segments based on the normal heartbeat duration range to obtain the filtered heartbeat sequence of the lead.

[0072] In this embodiment, firstly, based on clinical electrocardiogram diagnostic criteria or training set heart rate distribution statistics, a normal heartbeat duration range is pre-set (e.g., a lower limit of 0.6 seconds and an upper limit of 1.2 seconds, corresponding to a heart rate of 50-100 beats / min). Then, for each single heartbeat segment in each lead, its duration (i.e., the time difference between adjacent global R wave positions) is calculated, and it is determined whether the duration falls within the normal heartbeat duration range. If the duration of a single heartbeat segment is within the preset range, it is retained as a candidate valid heartbeat. If the duration is too short (shorter than the lower limit) or too long (longer than the upper limit), it is judged as an abnormal heartbeat (such as premature beats, pauses, or artifact interference) and is removed. The above judgment and screening operations are performed sequentially on all single heartbeat segments in all leads, and finally, the screened heartbeat sequence corresponding to each lead is obtained. The heartbeat segments in this sequence all have physiologically reasonable heart rate ranges, providing a quality-controllable data foundation for subsequent morphological matching and feature extraction.

[0073] Step S305: Based on the preset heartbeat template correlation coefficient threshold, perform morphological matching on the filtered heartbeat sequence to obtain a valid heartbeat sequence.

[0074] In this embodiment, firstly, several typical normal heartbeats are selected from the training set constructed from historical data (for example, 10 heartbeats with moderate amplitude and typical waveform morphology selected from the heartbeat sequences after duration screening and averaged) to construct a unified heartbeat template. Then, for each single heartbeat segment in the heartbeat sequence after screening for each lead, the R-wave peak point is used as a reference to align with the heartbeat template, and the Pearson correlation coefficient between the two is calculated as an index of heartbeat morphological similarity. To avoid mistakenly removing heartbeats with ST segment changes caused by myocardial ischemia, the morphological matching threshold is set to a low value (e.g., 0.3). This threshold is only used to identify and remove obvious spurious heartbeats caused by lead detachment, poor electrode contact, severe electromyography bursts, etc. (in which case the correlation coefficient is usually below 0.2), while ST segment elevation or depression caused by ischemia will not cause the correlation coefficient to drop below this threshold. If the correlation coefficient of a heartbeat segment is greater than or equal to this threshold, it is determined to be morphologically usable and retained; if it is less than this threshold, it is determined to be a severely spurious heartbeat and removed. After screening all leads, the heartbeat segments were sequentially matched and screened as described above to obtain the valid heartbeat sequences for each lead. These sequences have eliminated severe artifact interference while fully preserving ischemia-related morphological changes, providing a high-quality, highly sensitive data foundation for subsequent feature extraction such as ST segment shift.

[0075] Step S4: Extract features from the effective heartbeat sequence to obtain a multidimensional feature vector.

[0076] The specific steps for extracting features from the effective heartbeat sequence to obtain a multidimensional feature vector include the following: Step S401: The effective heartbeat sequence includes multiple effective heartbeat segments, and the ST segment interval and isoelectric line interval are extracted from the effective heartbeat segments.

[0077] In this embodiment, for each valid heartbeat segment in each lead's valid heartbeat sequence, the J point is first located based on the R wave peak point (usually a fixed offset of 40ms after the R wave peak is taken as the J point position). The ST segment interval is extended backward from the J point for a preset duration (e.g., 60ms to 80ms). At the same time, the PR stable segment between the QRS complex initiation point and the P wave initiation point is selected in the heartbeat segment, and this area is determined as the isoelectric line interval. The above interval positioning is completed for each heartbeat segment, and the corresponding ST segment interval and isoelectric line interval are extracted respectively, providing a standard waveform analysis interval for subsequent ST segment offset calculation.

[0078] Step S402: Calculate the amplitude difference between the ST segment interval and the equipotential line interval to obtain the ST segment offset set of the lead.

[0079] In this embodiment, for the ST segment interval and isoelectric line interval of each effective heartbeat segment in each lead, the amplitude of all sampling points in the isoelectric line interval is first averaged, and the average value is used as the stable ECG baseline amplitude of the heartbeat. Then, the average value of the amplitude of all sampling points in the ST segment interval of the same heartbeat is calculated. The ST segment offset corresponding to the heartbeat is obtained by subtracting the isoelectric line baseline amplitude from the average ST segment interval amplitude. The above method is used to calculate all effective heartbeat segments in the lead one by one. The offset obtained for each heartbeat is summarized in sequence to finally form a complete set of ST segment offsets for the lead.

[0080] Step S403: Based on the pre-acquired anatomical location of myocardial ischemia, the ST segment offset set is divided into anterior wall lead group, inferior wall lead group and lateral wall lead group.

[0081] In this embodiment, the anatomical localization of myocardial ischemia refers to the correspondence between myocardial ischemia and the classic anatomical localization of surface electrocardiogram leads in clinical electrocardiography: anterior myocardial ischemia is mainly reflected in precordial leads V1, V2, V3, and V4; inferior myocardial ischemia is mainly reflected in limb leads II, III, and aVF; and lateral myocardial ischemia is mainly reflected in limb leads I and aVL and precordial leads V5 and V6. According to this anatomical localization rule, the ST segment offset sets corresponding to each of the 12 leads are matched and assigned to the corresponding groups one by one. Among them, the ST segment offset sets of leads V1 to V4 are divided into the anterior wall lead group, the ST segment offset sets of leads II, III, and aVF are divided into the inferior wall lead group, and the ST segment offset sets of leads I, aVL, V5, and V6 are divided into the lateral wall lead group. Finally, the ST segment offset data corresponding to the three anatomical regions are grouped.

[0082] Step S404: Calculate the maximum offset and the range of offset for the anterior wall lead group, the inferior wall lead group, and the lateral wall lead group respectively to form the intra-lead group characteristics of the leads.

[0083] In this embodiment, for the anterior wall lead group, inferior wall lead group, and lateral wall lead group, the ST segment offset values ​​of all leads in each group are obtained. For each lead group, the maximum value of the offset values ​​in the group is first extracted as the maximum offset value of the group, which is used to characterize the most severe ST segment deviation in the anatomical region. Then, the range of the offset values ​​in the group is calculated, that is, the difference between the maximum and minimum values, which is used as the offset range of the group, to characterize the spatial variability or uniformity of ST segment offset in the anatomical region. The maximum value and range of each of the three lead groups, a total of six values, are combined to form the intra-lead group feature. This feature can quantitatively describe the most severe degree of myocardial ischemia and the consistency of the ischemic range in different myocardial ischemia areas, providing a quantitative basis for distinguishing between focal ischemia (small range) and diffuse ischemia (large range).

[0084] Step S405: Calculate the offset difference between the pre-set clinically corresponding mirror lead pairs between the anterior wall lead group and the lateral wall lead group to form the inter-lead group characteristics of the leads.

[0085] In this embodiment, firstly, based on the spatial correspondence of cardiac anatomy, clinically corresponding mirror lead pairs between the anterior wall lead group and the lateral wall lead group are pre-defined. These clinically corresponding mirror lead pairs refer to lead combinations that are positioned left and right in the transverse plane of the heart, including at least (V2, V5) and (V3, V6). V2 and V3 belong to the anterior wall lead group, and V5 and V6 belong to the lateral wall lead group. Clinically, ST segment changes in these corresponding leads are often compared to determine whether ischemia involves the lateral wall. For each pair of clinically corresponding mirror leads, the corresponding leads in the anterior wall lead group and the lateral wall lead group are obtained respectively. The ST segment offset of each lead is calculated, and the signed difference between the two leads (the offset of the anterior lead minus the offset of the lateral lead) is obtained to obtain the offset difference of the lead pair. The offset differences of each mirror lead pair are combined in a fixed order to form the inter-lead group feature. This feature can quantify the degree of spatial difference in ST segment offset between the anterior and lateral myocardial regions. In acute anterior wall ischemia, the ST segment of the anterior lead is elevated while the ST segment of the lateral lead is relatively normal or slightly depressed, and the difference increases accordingly. It can reflect the spatial extent of the ischemic area and the involvement of the lateral wall, providing a quantitative indicator for assessing the spatial distribution of myocardial ischemia.

[0086] Step S406: Construct a multidimensional feature vector based on the intra-lead group features and the inter-lead group features.

[0087] The construction of a multidimensional feature vector based on the intra-lead group features and the inter-lead group features specifically includes the following sub-steps: Step S40601: Calculate the T wave characteristics and QTc interval in the effective heartbeat segment.

[0088] In this embodiment, for each valid heartbeat segment in each lead, the T-wave interval is first located backward from the R-wave peak point. Within a preset time window, the peak and end points of the T-wave are detected, and the complete T-wave waveform is extracted. Then, the polarity (determining whether the T-wave direction is positive, negative, or bidirectional), symmetry (comparing the slope difference between the rising and falling branches of the T-wave), and T-wave amplitude (the vertical distance from the isoelectric line to the T-wave peak) of the T-wave segment are calculated as the T-wave characteristics of the heartbeat segment. At the same time, the QRS complex initiation and T-wave end point of the heartbeat segment are detected, and the time interval from the initiation to the end point is calculated to obtain the QT interval. The RR interval between adjacent R waves is used to perform heart rate correction on the QT interval to obtain the QTc interval. Finally, the T-wave characteristics and QTc interval are output as the electrocardiographic feature values ​​of the heartbeat segment for subsequent fusion with ST segment features.

[0089] Step S40602: Construct a multidimensional feature vector using the T-wave characteristics, the QTc interval, the intra-lead group characteristics, and the inter-lead group characteristics.

[0090] In this embodiment, for 12-lead ECG data, the T-wave characteristics (polarity, symmetry, amplitude) and QTc interval of the effective heartbeat segments in each lead are first fused at the lead level. The mean or median of the corresponding features in each of the 12 leads are calculated to form compressed global heartbeat supplementary features. Then, the intra-lead features generated in step S404 (including the maximum and range of ST segment offsets in the anterior, inferior, and lateral lead groups, a total of 6 fixed values) and the inter-lead features generated in step S405 (i.e., the ST segment offset differences between each anatomically mirrored lead pair between the anterior and lateral lead groups) are obtained. Finally, the above features are spliced ​​together in a preset fixed order: first, the intra-lead feature components are arranged, then the inter-lead feature components are arranged, and finally, the T-wave feature components and QTc interval components after 12-lead fusion are arranged, thereby constructing a complete multi-dimensional feature vector with fixed dimensions that is adapted to the input of the classification model.

[0091] Step S5: Input the multidimensional feature vector into the pre-built classification model to obtain the classification result and probability.

[0092] In this embodiment, a large amount of clinically labeled 12-lead ECG data was pre-collected to construct a training dataset. Multidimensional feature vectors were constructed for all original ECG samples using the same process, serving as the unified input samples for the model. Simultaneously, cardiovascular specialists completed multi-class labeling, classifying the samples into four categories: normal ECG, mild myocardial ischemia, moderate myocardial ischemia, and severe myocardial ischemia. The labeled feature vector sample set was then proportionally split into training, validation, and test sets. The XGBoost multi-class ensemble model, adapted to the structured multidimensional ECG features, was selected as the classification architecture. Multi-class cross-entropy was used as the loss function, and a Bayesian optimization algorithm was employed to optimize hyperparameters such as the learning rate, maximum tree depth, and regularization coefficient. Five-fold cross-validation was combined to complete iterative model training and optimal parameter selection. The multidimensional feature vector corresponding to the current ECG to be detected was input into the fully trained model. Forward inference outputs the corresponding probabilities of the four categories, and the category with the highest probability is taken as the final classification result.

[0093] like Figure 2 As shown, in another aspect, the present invention also provides a risk assessment system based on myocardial ischemia, including an input device, a processor, an output device, and a memory, wherein the input device, processor, output device, and memory are interconnected, the memory includes the computer-readable storage medium described above, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions.

[0094] In this embodiment, the input device is used to provide the system with relevant input data or instructions. In a risk assessment system based on myocardial ischemia, the input device may include common human-computer interaction interface devices such as keyboards, mice, and touchscreens. Through the input device, doctors or researchers can input necessary parameters.

[0095] The processor is the core component of the system, responsible for executing computer program instructions and performing data processing and analysis. In a risk assessment system based on myocardial ischemia, the processor analyzes and interprets the input test data by running pre-programmed algorithms and models. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), or other dedicated processing unit.

[0096] The memory is used to store computer programs, data, and parameters required by the system. It may include random access memory (RAM) for temporary data storage and processing, and persistent memory (such as hard disks or solid-state drives) for long-term data storage and preservation.

[0097] The output device is used to present the results of system processing and analysis to users or external devices. The output device can be a monitor, printer, charting device, etc. Through the output device, the system can display the prediction results, which can be used as a reference for doctors, researchers, or patients to assist in decision-making and communication.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the following steps: Obtain the raw discrete ECG sequences of multiple leads, and determine the optimal decomposition parameters based on the raw discrete ECG sequences, including: Obtain the range of decomposition parameters, and determine the initial population based on the range of decomposition parameters; Based on the individuals in the initial population, variational mode decomposition is performed on the original discrete electrocardiogram sequence to obtain multiple initial mode components; The frequency range of electrocardiogram (ECG) data is obtained based on pre-acquired historical ECG data; The initial modal components are filtered using the frequency range of the electrocardiogram data to obtain denoised modal components; Calculate the initial permutation entropy of the denoised mode components, and extract the maximum value from the initial permutation entropy to obtain the maximum permutation entropy of the individual; Calculate the initial frequency aliasing degree between two adjacent denoised mode components, and extract the maximum value from the initial frequency aliasing degree to obtain the maximum frequency aliasing degree of the individual. The individuals are sorted non-dominatedly with the objectives of minimizing the maximum permutation entropy and the maximum frequency aliasing, respectively. Based on the results of the non-dominated sorting, the Pareto front is obtained by iteratively using a multi-objective optimization algorithm. The reconstructed discrete electrocardiogram sequence is obtained based on each non-dominated individual in the Pareto front; Calculate the signal energy of the original discrete electrocardiogram sequence; Calculate the mean square error between the reconstructed discrete ECG sequence and the original discrete ECG sequence; The reconstructed signal-to-noise ratio is obtained by calculating the ratio of the signal energy to the mean square error. Extract the minimum signal-to-noise ratio from the reconstructed signal-to-noise ratio, and determine the optimal decomposition parameters based on the minimum signal-to-noise ratio; Based on the optimal decomposition parameters, variational mode decomposition is performed on the original discrete ECG sequence to obtain the pure ECG sequence of the lead; R-wave detection was performed on the pure electrocardiogram sequence to obtain multiple R-wave position sequences in the leads; The R-wave position sequence is fused to obtain multiple global R-wave positions; The pure electrocardiogram sequence is segmented based on the global R wave position to obtain multiple single-beat segments of the lead; Obtain the normal heartbeat duration range, and filter the single heartbeat segments based on the normal heartbeat duration range to obtain the filtered heartbeat sequence of the lead; Based on a preset heartbeat template correlation coefficient threshold, the filtered heartbeat sequences are morphologically matched to obtain valid heartbeat sequences. The effective heartbeat sequence includes multiple effective heartbeat segments, and the ST segment interval and isoelectric line interval are extracted from the effective heartbeat segments; Calculate the amplitude difference between the ST segment interval and the equipotential line interval to obtain the ST segment offset set of the lead; Based on the pre-acquired anatomical localization of myocardial ischemia, the ST segment offset set is divided into anterior wall lead group, inferior wall lead group and lateral wall lead group. The maximum offset and the range of offset for the anterior wall lead group, the inferior wall lead group, and the lateral wall lead group are calculated respectively to form the intra-lead group characteristics of the leads; Calculate the offset difference between the pre-set clinically corresponding mirror lead pairs between the anterior wall lead group and the lateral wall lead group to form the inter-lead group characteristics of the leads; Calculate the T-wave characteristics and QTc interval in the effective heartbeat segments; A multidimensional feature vector is constructed using the T-wave characteristics, the QTc interval, the intra-lead group characteristics, and the inter-lead group characteristics; The multidimensional feature vector is input into a pre-built classification model to obtain classification results and probabilities.

2. A risk assessment system based on myocardial ischemia, characterized in that, The device includes an input device, a processor, an output device, and a memory, wherein the input device, processor, output device, and memory are interconnected, the memory includes a computer-readable storage medium as described in claim 1, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions.

Images