Physiological signal analysis method and system based on multi-scale cross guidance neural network

Abstract

The invention belongs to the technical field of physiological signal analysis, and discloses a physiological signal analysis method and system based on a multi-scale cross guidance neural network, and the method comprises the steps: obtaining and segmenting a physiological signal into a plurality of data partitions; a plurality of parallel convolutional neural network sub-modules with different scales are adopted to carry out high-dimension mapping on the physiological signals of the data partitions to obtain nonlinear vectors, correlation among the different convolutional neural network sub-modules is learned and quantized into weight vectors, parallel operation is carried out on the nonlinear vectors based on the weight vectors, and the physiological signals of the multiple data partitions are obtained. Performing point convolution operation on a parallel operation result to obtain an information vector; decoding the information vector to obtain a feature vector; and carrying out physiological signal analysis by utilizing the feature vector. The method has classification and regression capabilities at the same time, does not need to re-optimize hyper-parameters when being generalized to different tasks, is light in design, can be deployed on edge equipment, and has good accuracy and effectiveness in different clinical scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of physiological signal analysis, and in particular to a physiological signal analysis method and system based on a multi-scale cross-guidance neural network. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Physiological signals are an important research subject in the medical field. By analyzing physiological signals, doctors can understand patients' specific activities, emotions, and health conditions, which has far-reaching significance for clinical diagnosis, treatment monitoring, and drug testing. The application of wearable devices has led to a longer recording time for physiological signals. At the same time, the popularity of electronic health records has diversified the types of tasks for physiological signal analysis. Physiological signal analysis in clinical practice mainly relies on manual annotations by experts, which has low analysis efficiency and frequent fatigue errors. The long duration and diversification of tasks further increase the burden of physiological signal analysis. Therefore, artificial intelligence (AI) is needed in clinical practice to improve the accuracy and efficiency of physiological signal analysis.

[0004] Although AI in physiological signal analysis involves clinical applications in multiple fields such as disease detection, physiological parameter monitoring, and disease prediction, from the perspective of task nature, AI mainly completes two basic tasks: classification and regression. Classification is to divide the input signal into predefined categories based on its characteristics, and regression is to find the relationship between the independent variable and the dependent variable. The difference between the two has led to the current research mainly developing different dedicated artificial intelligence (ANI). For example, analyzing electrocardiogram (ECG) signals to achieve atrial fibrillation (AF) identification and analyzing photoplethysmography (PPG) signals to achieve non-invasive blood pressure (BP) estimation require specific classification models and regression models respectively. In specific clinical scenarios, ANI alleviates the clinical burden of physiological signal analysis and provides patients with better treatment options.

[0005] However, in clinical practice, comorbidity is a common phenomenon, which means that patients suffer from two or more diseases at the same time, such as AF and hypertension. Faced with patients with comorbidities, AI needs to complete various types of physiological signal analysis tasks to provide comprehensive decision support, especially for patients in the intensive care unit. AI needs to complete regression tasks to achieve heart rate estimation, blood oxygen saturation estimation, and BP estimation, and needs to complete classification tasks to achieve organ failure warning, myocardial ischemia identification, and arrhythmia detection. This requires the algorithm to have both classification and regression capabilities. Therefore, existing ANIs have limitations that cannot be ignored: (1) When the same ANI is generalized to different tasks, the performance will be significantly reduced; (2) Integrating different ANIs to solve multiple tasks is a heavyweight solution and has limited flexibility in actual deployment; (3) ANIs need to continuously optimize hyperparameters when facing complex clinical environments. For example, the hyperparameter designs of machine learning (ML) algorithms such as random forest (RF), CatBoost, and extreme learning machine cannot be shared in different fields. Specifically, in AF identification and BP estimation, the trees, nodes, and depths of random forest (RF) are different. Deep learning (DL) algorithms such as Transformer, Mamba, and Kolmogorov-Arnold Networks are applied to natural language processing, image recognition, and time series prediction, but the network structures, optimization methods, and basic operators of different tasks are different. At the same time, these models are heavyweight and difficult to deploy on edge devices. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a physiological signal analysis method and system based on a multi-scale cross-guidance neural network, which has the capabilities of classification and regression, does not require re-optimization of hyperparameters when generalized to different tasks, has good robustness, has a lightweight design, can be deployed on edge devices, and has good accuracy and effectiveness in different clinical scenarios.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a physiological signal analysis method based on a multi-scale cross-guidance neural network, comprising the following steps:

[0009] Acquire physiological signals, and divide the physiological signals into multiple data partitions;

[0010] Use multiple parallel convolutional neural network submodules of different scales to perform high-dimensional mapping on the physiological signals of multiple data partitions to obtain nonlinear vectors, learn the correlation between different convolutional neural network submodules and quantize them into weight vectors, perform parallel operations on the nonlinear vectors based on the weight vectors, perform point convolution operations on the parallel operation results, and obtain information vectors;

[0011] Decode the information vector to obtain a feature vector that meets the specific physiological signal analysis task;

[0012] Physiological signal analysis is performed using feature vectors that meet specific physiological signal analysis tasks.

[0013] As an optional implementation, sparse synaptic connections are used to strengthen the information interaction between different convolutional neural network sub-modules, and the correlation between different convolutional neural network sub-modules is learned and quantized into a weight vector by embedding global information and adaptive correction.

[0014] As an optional implementation, a multi-scale deep separable convolutional neural network is used to perform parallel operations on nonlinear vectors, and a point convolution operation is performed on the parallel operation results.

[0015] As an optional implementation, the multi-scale deep convolutional neural network introduces skip connections, and the output of the shallow deep convolutional neural network is directly added to the input of the deep deep convolutional neural network to slow down gradient changes.

[0016] As an optional implementation, the multi-scale deep convolutional neural network is constructed by a DC operator, which applies the idea of ​​channel-by-channel convolution.

[0017] As an optional implementation, TCNN is used to decode the information vector, and the TCNN is constructed using a transposed convolution operator.

[0018] In a second aspect, the present invention provides a physiological signal analysis system based on a multi-scale cross-guidance neural network, comprising:

[0019] The data acquisition and partitioning module is configured to: acquire physiological signals and divide the physiological signals into a plurality of data partitions;

[0020] The encoding module is configured to: use multiple parallel convolutional neural network submodules of different scales to perform high-dimensional mapping on physiological signals of multiple data partitions respectively to obtain nonlinear vectors, learn the correlation between different convolutional neural network submodules and quantize it into weight vectors, perform parallel operations on the nonlinear vectors based on the weight vectors, perform point convolution operations on the parallel operation results, and obtain information vectors;

[0021] A decoding module is configured to: decode the information vector to obtain a feature vector that meets a specific physiological signal analysis task;

[0022] The analysis module is configured to perform physiological signal analysis using a feature vector that meets a specific physiological signal analysis task.

[0023] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0024] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0025] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present disclosure proposes a physiological signal analysis method and system based on a multi-scale cross-guidance neural network, which has the ability of classification and regression at the same time, and can complete the physiological signal analysis tasks of classification and regression with two different attributes without changing any hyperparameters. The present disclosure adopts a multi-scale mechanism for feature extraction, in which a small-scale receptive field is used to learn local detail information, and a large-scale receptive field is used to learn global semantic information. The multi-scale mechanism enables the present method to have the ability to extract diversified features, thereby extracting the features required for different tasks, and showing good versatility in different tasks.

[0028] The present disclosure proposes a physiological signal analysis method and system based on a multi-scale cross-guidance neural network, which adopts a distributed concept to solve the problem of high computational cost of multi-scale mechanisms. The backbone of this method is divided into multiple independent submodules. Each submodule is constructed by a lightweight convolution operator. The optimization of the backbone and operators reduces the computational cost of physiological signal analysis. The distributed architecture enables MSCG-Net to have lower computational cost and faster reasoning speed so that it can be deployed on edge devices.

[0029] The present invention discloses a physiological signal analysis method and system based on a multi-scale cross-guidance neural network, which introduces sparse synaptic connections to enhance the information interaction of submodules in a distributed neural network and solve the problem of information islands. The submodules of MSCG-Net are regarded as biological neurons, and the attention mechanism with negligible computational cost is used to simulate synapses. Any submodule is weighted by the remaining submodules to achieve cross-guidance. The cross-guidance mechanism based on sparse synaptic connections enables MSCG-Net to have stronger information interaction capabilities to select complementary and non-redundant features.

[0030] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0032] Figure 1 A flowchart of a physiological signal analysis method based on a multi-scale cross-guidance neural network provided in Example 1 of the present invention;

[0033] Figure 2 An experimental flow chart for verifying the physiological signal analysis method based on a multi-scale cross-guidance neural network provided in Example 1 of the present invention;

[0034] Figure 3 A comparison chart of ROC curves and AUC values ​​of the physiological signal analysis method based on a multi-scale cross-guidance neural network provided in Example 1 of the present invention and other methods in classification tasks;

[0035] Figure 4 A graph showing the regression results of the physiological signal analysis method based on a multi-scale cross-guidance neural network provided in Example 1 of the present invention;

[0036] Figure 5 Class activation map for multi-scale mechanism in classification task;

[0037] Figure 6 The present invention provides an experimental result of estimating SBP using the physiological signal analysis method based on a multi-scale cross-guidance neural network provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0040] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0041] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0042] Example 1

[0043] like Figure 1 As shown, this embodiment provides a physiological signal analysis method based on a multi-scale cross-guidance neural network, comprising the following steps:

[0044] S1 acquires physiological signals and divides the physiological signals into multiple data partitions;

[0045] S2 uses multiple parallel convolutional neural network submodules of different scales to perform high-dimensional mapping on the physiological signals of multiple data partitions, obtains nonlinear vectors, learns the correlation between different convolutional neural network submodules and quantizes them into weight vectors, performs parallel operations on nonlinear vectors based on weight vectors, and performs point convolution operations on the parallel operation results to obtain information vectors;

[0046] S3 decodes the information vector to obtain a feature vector that meets the specific physiological signal analysis task;

[0047] S4 performs physiological signal analysis using feature vectors that meet specific physiological signal analysis tasks.

[0048] The present invention proposes a universal multi-scale cross-guidance neural network (MSCG-Net), which does not require re-optimization of hyperparameters when applied to classification and regression and has good robustness when generalized to different tasks.

[0049] like Figure 1 As shown in (a), first, after the input physiological signal is partitioned, it is sequentially fed into a multi-scale convolutional neural network (CNN) for high-dimensional mapping, an encoder composed of a multi-scale deep separable convolutional neural network (DCNN), and a decoder constructed by a transposed convolutional neural network (TCNN) to achieve feature extraction based on a multi-scale mechanism. The multi-scale mechanism is used to parse physiological signals to extract the features required for different tasks.

[0050] Then, different CNNs, DCNNs and TCNNs were computed in parallel and multi-scale DCNNs were computed using channel-by-channel convolution to achieve computational cost optimization based on distributed thinking. Distributed architecture was used to optimize convolution operators and network backbones to reduce the computational cost of multi-scale mechanisms.

[0051] Finally, the squeeze and excitation network is used to learn the feature correlations between sub-modules and quantize them into weights to construct a cross-guidance mechanism based on the squeeze and excitation network. The squeeze and excitation network is used to strengthen the information interaction of sub-modules in the distributed architecture to improve robustness. The squeeze and excitation network is part of the sparse synaptic connection.

[0052] MSCG-Net uses multi-scale CNN to map physiological signals in high dimensions to facilitate DCNN analysis, such as Figure 1 (b). The calculation process of CNN is shown in formula (1).

[0053]

[0054] * represents convolution calculation, sk represents the scale of the receptive field, BN stands for batch normalization, y scale,sk Represents the output of CNN with scale sk. After projection by CNN, the physiological signal is converted into a high-dimensional nonlinear vector, where the dimension changes to

[0055] The role of multi-scale DCNN is to encode the temporal changes and spatial distribution of physiological signals into information vectors. MSCG-Net uses multi-scale DCNN to scale,sk Perform parallel operations, as shown in formula (2-3),

[0056]

[0057] represents the tensor product, w DC,sk represents the weight of DC, b DC,sk represents the DC bias, represents the weight of the point convolution, Represents the bias of the point convolution, y DC,sk represents the output of DC with scale sk, y DCNN,sk Represents the output of the DCNN with scale sk. Each scale of DCNN repeats the above operation to decompose the nonlinear vector output by CNN.

[0058] Each DCNN uses two sets of DC and two sets of point convolution, and the depth of MSCG-Net is 24. In order to avoid gradient vanishing and exploding, the present disclosure introduces skip connections. The output of the shallow DCNN is directly added to the input of the deep DCNN to slow down the gradient change. The details are as follows: Assume that the input of the mth DCNN is First, Z is outputted by formula (2-3) Then, Input the m+1th DCNN to get the output at last, Input the m+2th DCNN to get the output

[0059] TCNN decodes the information vector into a feature vector that meets the specific physiological signal analysis task. MSCG-Net uses the transposed convolution operator to construct TCNN. Assume that the output of DCNN is The set composed of TCNN is TCNN i right Decode to get the feature vector Different feature vectors are flattened and element-wise summed to obtain a feature set for physiological signal classification and regression. The process is shown in formula (4-6):

[0060]

[0061] is the transposed convolution, is the weight matrix, k inv is the size of the deconvolution kernel, θ(*) is the ELU, is element-wise addition. Different TCNNs use the same receptive field to decode vectors of different scales equally.

[0062] The multi-scale mechanism faces the problems of high computational cost and slow inference speed. This paper applies the distributed concept to data partitioning, optimization of convolution operators and design of backbone networks to solve the above problems. The idea of ​​distributed computing is to split a highly complex system into multiple simple sub-modules to achieve parallel computing. Specifically: (1) Based on Embedding, any physiological signal segment is divided into s data partitions, such as Figure 1 (b) (2) The DC operator is used to replace the standard convolution (RC) operator to construct a multi-scale DCNN. (3) CNNs, DCNNs, and TCNNs of different scales are calculated in parallel.

[0063] Assume that any physiological signal segment is After data partitioning, X is CNN analysis of X and X at s scales seg The FLOPs is shown in formula (7-8),

[0064]

[0065]

[0066] K i is the size of the receptive field, C iis the number of convolution operators, and n is the number of sampling points of the physiological signal. After data partitioning, FLOPs is reduced to 1 / s of the original.

[0067] DCNN is constructed by DC operator. DC operator applies the idea of ​​channel-by-channel convolution, and its parameters and FLOPs are smaller than those of RC operator. DC Computational FLOPs DC As shown in formula (9-10).

[0068] Par DC =K DC ×I DC +C Po ×I DC (9)

[0069] FLOPs DC =K DC ×n×I DC +I DC ×n×C Po (10)

[0070] K DC represents the receptive field size of DC, I DC Represents the number of input channels, C Po Represents the number of point convolution operators. RC parameter quantity Par when analyzing physiological signals RC Computational FLOPs RC As shown in formula (11-12).

[0071] Par RC =I RC ×K RC ×C RC (11)

[0072] FLOPs RC =I RC ×K RC ×n×C Rm (12)

[0073] K RC represents the receptive field size of RC, I RC Represents the number of input channels, C RC Represents the number of RC operators. When the hyperparameters of DC and RC are consistent, that is, K DC =K RC , I DC =I RC , C Po =C RC ,

[0074]

[0075] From formula (13-14), when analyzing the same physiological signal, the number of parameters and FLOPs of the DC operator is only (1 / C RC +1 / K RC ). In the present disclosure, C RC >>2, K RC ≥3.

[0076] Parallel computing is used between different scales. The FLOPs of parallel (PC) and non-parallel (NPC) are shown in formula (15-16).

[0077]

[0078] FLOPs NPC =I in ×K coder ×n×C out (16)

[0079] I in is the number of input channels, C out is the number of output channels, K coder is the size of the receptive field. After adopting the parallel method, FLOPs is reduced to 1 / s of the original.

[0080] The distributed computing architecture inevitably destroys the temporal correlation of physiological signals and weakens the information interaction within MSCG-Net. Inspired by the feedforward neural network in the brain nervous system, this paper designs a cross-guidance mechanism based on sparse synaptic connections to strengthen the information interaction between MSCG-Net submodules, such as Figure 1 (c) as shown.

[0081] The present disclosure cleverly applies the sparse synaptic connections of the brain nervous system to the sub-modules of the distributed network, thereby increasing the information interaction capability of the distributed network at a low computational cost. In MSCG-Net, different sub-modules are regarded as biological neurons, and the attention mechanism with negligible computational cost is used to simulate synapses. Specifically: any set of nonlinear transformations is sent to SE-Net after batch normalization and ELU processing. SE-Net learns the correlation between different sub-modules by embedding global information and adaptive correction and quantizes it into a weight vector. Each sub-module is weighted by the remaining sub-modules to achieve cross-guidance.

[0082] Assume that the nonlinear transformation output of any group of submodules is a represents the length of the vector, and b represents the dimension of the vector. SE-Net will Convert to weight vector

[0083] first, After compression operation F sq (·) is converted into a global information vector The process is shown in formula (17):

[0084]

[0085] After the above operations, we get

[0086] Then, SE-Net Perform the excitation operation F ex (·) to learn the importance of each vector dimension and obtain the weight vector of the corresponding submodule The process is shown in formula (18):

[0087]

[0088] and is a trainable parameter vector, σ(·) is Sigmiod, and δ(·) is Relu. After the above operations, we can obtain the weight set

[0089] Finally, the (s-1) weight vectors are summed and weighted to the output of the specified submodule, thereby obtaining the output of multiple submodule information interaction. The process is shown in formula (19):

[0090]

[0091] This paper verifies the performance of MSCG-Net in classification tasks and regression tasks based on AF recognition experiments and BP estimation experiments. The experimental process is as follows: Figure 2 As shown in the figure. The MIT-BIH Atrial Fibrillation Database (AFDB) and the standardized patient heart dataset (SPHDB) created by the present invention were used for AF recognition experiments. The Medical Information Martfor Intensive Care Database (MIMICDB) was used for BP estimation experiments. AFDB, SPHDB, and MIMICDB have three problems: class imbalance, scarce labeled data, and noise, which are consistent with clinical scenarios.

[0092] The evaluation indicators of the AF recognition experiment are accuracy (Acc), specificity (Spe), sensitivity (Sen), precision (Pre) and F1 score. The calculation process is shown in formula (20-24).

[0093]

[0094] TP is true positive, TN is true negative, FN is false negative, and FP is false positive. The present disclosure uses Receiver Operating Characteristic (ROC) curve and Area Under Curve (AUC) value to evaluate the clinical significance of MSCG-Net in classification.

[0095] The evaluation indicators of BP estimation experiment are MAE and RMSE, and the calculation process is shown in formulas (25) and (26). This paper uses the standards of British Hypertension Society (BHS) and Association for the Advancement of Medical Instrumentation (AAMI) to evaluate the clinical application value of MSCG-Net in regression.

[0096]

[0097] BP truei Represents the real BP value, BP outi Represents the estimated BP value.

[0098] The experimental data results of MSCG-Net in classification tasks show that MSCG-Net is less affected by category imbalance and has good stability. MSCG-Net can complete the classification of physiological signals in clinical scenarios with category imbalance and maintain high generalization in clinical scenarios with scarce labeled data. This paper evaluates the clinical application value of MSCG-Net in physiological signal classification through ROC curve and AUC value. The results are as follows: Figure 3 The ROC curve of AFDB is shown in Figure 3 As shown in (a), the AUC value range is [99.89%, 99.94%]. The ROC curve of SPHDB is shown in Figure 3 As shown in (bf), the AUC value ranges from [97.55%, 99.81%]. The ROC curve and AUC value prove that MSCG-Net has high clinical application value in the classification of physiological signals and can adapt to complex and changing clinical environments.

[0099] The experimental data results of MSCG-Net in the regression task show that MSCG-Net can maintain accurate and stable regression performance when facing diverse subjects in clinical environments, and meets clinical standards in the regression of physiological signals. This paper conducts statistical analysis on the regression results, such as Figure 4As shown in the figure, the mean error (ME) is as low as 0.19 mmHg, and the 95% consistency limit (LoA) is [-9.89, 10.26] mmHg. This shows that the regression results of MSCG-Net are highly consistent with the true values. The Pearson Correlation Coefficient (PCC) between the regression results and the true values ​​reached 0.9570 and the p-value was less than 10. -15 This shows that the regression results of MSCG-Net have a high correlation with the true value and are statistically significant.

[0100] Class activation diagram of multi-scale mechanism in classification task Figure 5 As shown in the figure. Scales 3 and 5 focus only on one RR interval or a complete heartbeat, while scales 7 and 9 capture two adjacent heartbeats and more than three consecutive heartbeats, respectively. The identification of the P wave requires a complete heartbeat, and the premise for judging that the RR interval is absolutely irregular is to observe at least 3 or more heartbeats. The multi-scale mechanism has a complementary role in classification, which enables MSCG-Net to mimic the diagnostic mechanism of clinical experts.

[0101] This paper analyzes the effects of different scales on various regression tasks. The most representative example is the experiment of estimating SBP, such as Figure 6 As shown. SBP is located at the junction of the ascending branch and the descending branch, which requires MSCG-Net to capture both the rapid rise and slow fall of the slope. Therefore, this requires small scales and large scales to jointly analyze physiological signals. Among all subjects, ablation of small scale K=5 and large scale K=9 resulted in errors of 29.91% and 30.62%, respectively. Similarly, DBP has similar results to SBP, and ablation scales K=3 and K=7 resulted in errors of 30.98% and 27.22%, respectively. MAP represents the average level of BP, which requires MSCG-Net to capture outliers and periodicity of BP. Therefore, the four scales have similar importance, and the errors caused by ablating the four scales are all around 25%.

[0102] In addition, among the four categories of subjects, the higher the severity of hypertension, the greater the degree of oscillation of BP, and the higher the effect of small scales. Conversely, the effect of large scales is higher. For example, the ablation scales K=3 and K=5 together lead to a 77.22% error in the SBP of the S2 category subjects. The ablation scales K=7 and K=9 together lead to a 65.09% error in the SBP of the N category subjects. At the same time, there are certain differences in the error distributions generated by the four scales in each category of subjects, which shows that the multi-scale mechanism can capture the specificity of different subjects. The multi-scale mechanism has a complementary role in regression, which enables MSCG-Net to better cope with the diverse clinical needs.

[0103] The MSCG-Net proposed in the present disclosure is a versatile lightweight algorithm that can complete physiological signal analysis tasks of two different attributes, classification and regression, without changing any hyperparameters. The multi-scale mechanism enables MSCG-Net to extract diversified features and thus show good versatility in different tasks. The distributed architecture enables MSCG-Net to have lower computing costs and faster reasoning speed so that it can be deployed on edge devices. The cross-guidance mechanism based on Sparsesynaptic connectivity enables MSCG-Net to have stronger information interaction capabilities to select complementary and non-redundant features. Compared with existing ANI, MSCG-Net has better versatility and robustness in physiological signal analysis, and has higher flexibility and cost-effectiveness in integrated deployment. The present disclosure contributes to the development of general artificial intelligence and is of great clinical significance.

[0104] Example 2

[0105] This embodiment provides a physiological signal analysis system based on a multi-scale cross-guidance neural network, including:

[0106] The data acquisition and partitioning module is configured to: acquire physiological signals and divide the physiological signals into a plurality of data partitions;

[0107] The encoding module is configured to: use multiple parallel convolutional neural network submodules of different scales to perform high-dimensional mapping on physiological signals of multiple data partitions respectively to obtain nonlinear vectors, learn the correlation between different convolutional neural network submodules and quantize it into weight vectors, perform parallel operations on the nonlinear vectors based on the weight vectors, perform point convolution operations on the parallel operation results, and obtain information vectors;

[0108] A decoding module is configured to: decode the information vector to obtain a feature vector that meets a specific physiological signal analysis task;

[0109] The analysis module is configured to perform physiological signal analysis using a feature vector that meets a specific physiological signal analysis task.

[0110] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer executable instructions.

[0111] In further embodiments, there is also provided:

[0112] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

[0113] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0114] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0115] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0116] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

[0117] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in embodiment 1 is implemented.

[0118] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0119] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0120] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, and the like.

[0121] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0122] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A physiological signal analysis method based on a multi-scale cross-guidance neural network, characterized in that: The following steps are involved: Acquire physiological signals, and divide the physiological signals into multiple data partitions; Use multiple parallel convolutional neural network submodules of different scales to perform high-dimensional mapping on the physiological signals of multiple data partitions to obtain nonlinear vectors, learn the correlation between different convolutional neural network submodules and quantize them into weight vectors, perform parallel operations on the nonlinear vectors based on the weight vectors, perform point convolution operations on the parallel operation results, and obtain information vectors; Decode the information vector to obtain a feature vector that meets the specific physiological signal analysis task; Physiological signal analysis is performed using feature vectors that meet specific physiological signal analysis tasks.

2. The physiological signal analysis method based on a multi-scale cross-guidance neural network according to claim 1, characterized in that: Sparse synaptic connections are used to strengthen the information interaction between different convolutional neural network sub-modules. The correlation between different convolutional neural network sub-modules is learned and quantized into a weight vector by embedding global information and adaptive correction.

3. The physiological signal analysis method based on multi-scale cross-guidance neural network according to claim 1, characterized in that: A multi-scale deep separable convolutional neural network is used to perform parallel operations on nonlinear vectors, and point convolution operations are performed on the parallel operation results.

4. The physiological signal analysis method based on multi-scale cross-guidance neural network according to claim 3, characterized in that: The multi-scale deep convolutional neural network introduces skip connections, and the output of the shallow deep convolutional neural network is directly added to the input of the deep deep convolutional neural network to slow down the gradient change.

5. The physiological signal analysis method based on multi-scale cross-guidance neural network according to claim 4, characterized in that: The multi-scale deep convolutional neural network is constructed by a DC operator, which applies the idea of ​​channel-by-channel convolution.

6. The physiological signal analysis method based on multi-scale cross-guidance neural network according to claim 1, characterized in that: The information vector is decoded using TCNN, which is constructed using a transposed convolution operator.

7. A physiological signal analysis system based on a multi-scale cross-guidance neural network, characterized in that: include: The data acquisition and partitioning module is configured to: acquire physiological signals and divide the physiological signals into a plurality of data partitions; The encoding module is configured to: use multiple parallel convolutional neural network submodules of different scales to perform high-dimensional mapping on physiological signals of multiple data partitions respectively to obtain nonlinear vectors, learn the correlation between different convolutional neural network submodules and quantize it into weight vectors, perform parallel operations on the nonlinear vectors based on the weight vectors, perform point convolution operations on the parallel operation results, and obtain information vectors; A decoding module is configured to: decode the information vector to obtain a feature vector that meets a specific physiological signal analysis task; The analysis module is configured to perform physiological signal analysis using a feature vector that meets a specific physiological signal analysis task.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images