Sudden death risk real-time evaluation system and method based on multi-mode physiological signal fusion

Abstract

The invention discloses a sudden death risk real-time assessment system and method based on multi-modal physiological signal fusion, relates to the field of human physiological state monitoring and early warning, and solves the problems of low accuracy and poor real-time performance of sudden death risk assessment by single-modal signals. The system comprises a signal acquisition module, a preprocessing module, a high-dimensional feature extraction module, a multi-modal feature fusion module, a sudden death risk quantification module, a model updating engine module and an early warning feedback module, and each module integrates a self-adaptive filtering unit, a time sequence convolutional network unit, an improved multi-head self-attention mechanism unit and the like. According to the scheme, multi-mode signals such as electrocardio are synchronously collected through multiple channels, a dynamic risk index is calculated through self-adaptive noise reduction, parallel extraction of time-frequency domain nonlinear features and mutual information weighted fusion in combination with kernel density estimation, and online incremental updating and multi-stage early warning of a model are achieved; the sudden death risk can be accurately evaluated in real time, the anomaly detection sensitivity and the early warning timeliness are improved, and the method is suitable for daily health monitoring and high-risk group risk management and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical health monitoring, and more particularly to a sudden death risk real-time evaluation system and method based on multi-modal physiological signal fusion. BACKGROUND

[0002] As a serious acute condition threatening human life and health, sudden death has the characteristics of sudden onset, rapid progression and high mortality. Real-time and accurate risk assessment is the key to effective prevention and intervention. In the field of human physiological state monitoring, traditional sudden death risk assessment techniques mostly rely on single modal physiological signals, such as detecting arrhythmia indicators through electrocardiogram signals alone. However, the occurrence of sudden death is the result of coordinated abnormalities of multiple physiological systems, and single modal signals cannot fully reflect the complex physiological state of the body, resulting in insufficient evaluation dimensions and susceptibility to noise interference, which limits the accuracy and timeliness of risk assessment.

[0003] At the same time, the existing technology lacks efficient synchronous acquisition and time alignment mechanism in multi-modal physiological signal processing, and the time synchronization and data integrity of multi-source signals are difficult to guarantee. In the feature extraction link, the parallel extraction capability of time domain, frequency domain and nonlinear features of the signal is insufficient, and the sudden death related information contained in the physiological signal cannot be fully mined. In the multi-modal feature fusion, there is a lack of intelligent weighting strategy based on the correlation between modalities, and it is difficult to realize the complementary advantages of multi-source features. The risk quantization model is mostly a static model, which cannot be updated and optimized online according to real-time data, and the grading accuracy and feedback real-time of the early warning mechanism also need to be improved. These technical defects seriously restrict the clinical application and actual effect of sudden death risk real-time evaluation, and therefore an sudden death risk real-time evaluation system based on multi-modal physiological signal fusion is urgently needed to solve the above problems. SUMMARY

[0004] In view of the above technical deficiencies, the present application discloses an sudden death risk real-time evaluation system based on multi-modal physiological signal fusion, which acquires electrocardiogram, skin electricity, respiration and blood volume pulse signals through multiple channels in a synchronous manner, and completes preprocessing through adaptive signal decomposition and dynamic time warping. High-dimensional time-frequency and nonlinear features are extracted in parallel by fusing time convolution network and gated recurrent unit. Intelligent weighted fusion of multi-modal features is achieved with the help of improved multi-head self-attention mechanism and K-neighborhood based mutual information entropy estimation. Dynamic risk index is generated based on kernel density estimation and state space model. Online sequence learning strategy is used to realize model incremental update. Multi-level warning is triggered according to risk index grading and real-time feedback. The system can comprehensively, accurately and real-time evaluate sudden death risk, improve abnormal detection sensitivity and warning timeliness, and provide effective technical support for human physiological state monitoring and sudden death prevention.

[0005] In order to achieve the above technical effects, the present application adopts the following technical solutions:

[0006] A sudden death risk real-time evaluation system based on multi-modal physiological signal fusion is applied to the field of human physiological state monitoring and early warning, and the system comprises a signal acquisition module, a signal preprocessing module, a high-dimensional feature extraction module, a multi-modal feature fusion module, a sudden death risk quantification module, a model updating engine and an early warning feedback module.

[0007] The signal acquisition module adopts a multi-channel synchronous acquisition technology architecture to synchronously acquire electrocardio signals, skin electricity signals, respiratory signals and blood volume pulse signals, and maintain the time synchronization and data integrity of the multi-modal physiological signals. The signal acquisition module comprises a bioelectric signal electrode array, an optical sensing unit and an impedance detection circuit, and completes the synchronous acquisition of multi-source signals through a special interface protocol.

[0008] The signal preprocessing module integrates an adaptive filter and a quality evaluation timing unit, realizes adaptive noise reduction of signals through an adaptive signal decomposition algorithm, and completes the timing alignment of multi-modal signals through a dynamic time warping technology. The signal preprocessing module comprises a signal input interface, a variational modal decomposer, a dynamic threshold processor, a timing calibrator, a signal quality evaluator and a preprocessing output interface, and constitutes a complete signal preprocessing pipeline.

[0009] The high-dimensional feature extraction module fuses a timing convolution network and a gated recurrent unit, and extracts the time domain features, frequency domain features and nonlinear features of signals in parallel to generate a high-dimensional feature vector space. The high-dimensional feature extraction module comprises a feature extraction controller, a timing convolution processor, a recurrent neural network unit and a feature vector synthesizer.

[0010] The multi-modal feature fusion module adopts an improved multi-head self-attention mechanism, takes electrocardio signal features as a benchmark reference, realizes intelligent weighted fusion of multi-modal features through calculation of inter-modal mutual information entropy, and comprises an attention weight calculation unit, a feature projection transformer, a mutual information entropy analyzer and a fusion feature synthesizer to constitute a multi-level feature fusion architecture.

[0011] The sudden death risk quantification module integrates a timing anomaly detector and a risk regressor, calculates an abnormal score based on a kernel density estimation algorithm, and generates a dynamic risk index in combination with a historical risk state. The sudden death risk quantification module comprises an anomaly detection controller, a kernel density estimation unit, a risk state storage and a regression predictor. The model updating engine adopts an online sequence learning strategy, realizes incremental updating and optimization of model parameters through a regularized least squares algorithm, and comprises an incremental learning controller, a weight updater, a parameter optimizer and a model verification unit.

[0012] The early warning feedback module triggers a multi-level early warning protocol according to the risk index classification, and completes real-time feedback and display of the evaluation results; the early warning feedback module includes a risk level judge, an early warning protocol executor, a feedback information generator, and a man-machine interaction interface.

[0013] The signal preprocessing module includes a signal input interface, a variational modal decomposer, a dynamic threshold processor, a timing calibrator, a signal quality evaluator, and a preprocessing output interface.

[0014] The variational modal decomposer adopts an adaptive bandwidth optimization algorithm architecture, including a wave generator, a modal decomposer, and a reconstruction filter; the wave generator calculates signal envelope features through Hilbert transform; the modal decomposer solves variational constraint problems by using an alternating direction multiplier method; the reconstruction filter reconstructs high-quality signals after denoising through inverse transform;

[0015] The timing calibrator is implemented by using a dynamic time warping algorithm, including a distance matrix calculation unit, a warping path searcher, and a time scaler; the distance matrix calculation unit measures signal similarity features by using Euclidean distance; the warping path searcher finds the optimal alignment path by using a dynamic programming algorithm; the time scaler realizes signal resampling by using a cubic spline interpolation algorithm;

[0016] The signal quality evaluator integrates a multi-dimensional quality index evaluation system, including a signal-to-noise ratio calculation unit, a waveform distortion detection unit, and a motion artifact identification unit; the signal-to-noise ratio calculation unit calculates signal quality indicators by using a power spectral ratio method; the waveform distortion detection unit calculates waveform correlation coefficients by using a template matching algorithm; the motion artifact identification unit performs artifact compensation processing based on acceleration signal features.

[0017] The high-dimensional feature extraction module includes a feature extraction controller, a timing convolution processor, a recurrent neural network unit, and a feature vector synthesizer.

[0018] The timing convolution processor adopts an expanded causal convolution architecture with an exponential expansion factor, including a convolution kernel optimization unit, a feature mapping unit, and a pooling layer; the convolution kernel optimization unit optimizes convolution kernel parameters by using an adaptive gradient descent algorithm; the feature mapping unit realizes feature transformation by using a nonlinear activation function; the pooling layer adopts a multi-scale pooling strategy combining maximum pooling and average pooling;

[0019] The recurrent neural network unit adopts a gated attention recurrent unit structure, including a time attention mechanism, a gated update unit, and a state memory unit; the time attention mechanism calculates attention weights by using historical state similarity; the gated update unit integrates the calculation processes of update gates and reset gates; the state memory unit stores timing feature information by using a long short-term memory structure.

[0020] The multi-modal feature fusion module comprises an attention weight calculation unit, a feature projection transformer, an mutual information entropy analyzer, and a fusion feature synthesizer.

[0021] The mutual information entropy analyzer adopts a non-parametric mutual information estimation structure based on K-nearest neighbors, comprising a probability density estimation unit, an entropy value calculation unit, and a bias correction unit. The probability density estimation unit calculates the joint probability distribution using the kernel density estimation method. The entropy value calculation unit calculates the mutual information entropy value through the log-likelihood ratio. The bias correction unit corrects the estimation bias using the Jackknife resampling technique.

[0022] The fusion feature synthesizer adopts a weighted feature fusion architecture, comprising a weight distribution unit, a feature weighting unit, and a fusion output unit. The weight distribution unit distributes the weight of each modal feature according to the mutual information entropy value. The feature weighting unit realizes feature weighting using point multiplication operation. The fusion output unit generates the final fusion feature tensor through splicing operation.

[0023] The sudden death risk quantification module comprises an anomaly detection controller, a kernel density estimation unit, a risk state memory, and a regression predictor.

[0024] The anomaly detection controller adopts a multi-index anomaly detection architecture, comprising a threshold self-adaptive unit, an anomaly scorer, and an alarm triggering unit. The threshold self-adaptive unit dynamically adjusts the anomaly threshold according to the historical data distribution. The anomaly scorer calculates the anomaly score using the density estimation method. The alarm triggering unit triggers the hierarchical early warning mechanism according to the scoring result.

[0025] The risk state memory adopts a spatio-temporal state storage architecture, comprising a state encoder, a memory matrix, and a state retriever. The state encoder encodes the historical state information through recurrent neural network. The memory matrix stores the key state features using attention mechanism. The state retriever retrieves the relevant historical state through similarity matching.

[0026] The regression predictor adopts a hybrid architecture of state space model and deep learning, comprising a state transition unit, an observation model unit, and a parameter optimization unit. The state transition unit models the state evolution process using linear dynamic system. The observation model unit establishes the mapping relationship between the state and the observation through deep learning network. The parameter optimization unit optimizes the model parameters using the expectation maximization algorithm.

[0027] The model update engine comprises an incremental learning controller, a weight updater, a parameter optimizer, and a model verification unit.

[0028] The incremental learning controller adopts an online sequence extreme learning machine architecture, and includes a hidden node management unit, an output weight calculation unit, and a model updating unit; the hidden node management unit dynamically adds and deletes hidden layer nodes and optimizes node parameters; the output weight calculation unit calculates output layer weights by using a recursive least square method; and the model updating unit implements incremental model updating through an online learning algorithm.

[0029] The weight updater adopts a recursive least square architecture with a forgetting factor, and includes a forgetting factor adjustment unit, a covariance matrix updating unit, and a weight correction unit; the forgetting factor adjustment unit dynamically adjusts the size of the forgetting factor according to a prediction error; the covariance matrix updating unit updates the covariance matrix by using a rank-one updating algorithm; and the weight correction unit corrects weight parameters by using a gradient descent method.

[0030] A sudden death risk real-time evaluation method based on multi-modal physiological signal fusion is applied to the field of human physiological state monitoring and early warning, and the method comprises the following steps:

[0031] Step one, multi-modal physiological signal synchronous acquisition and quality evaluation:

[0032] Through the multi-channel synchronous acquisition architecture formed by the bioelectric signal electrode array, the optical sensing unit and the impedance detection circuit, the electrocardiosignal, the skin electricity signal, the respiration signal and the blood volume pulse signal of the user are synchronously acquired; the signal-to-noise ratio of each signal is calculated in real time and the motion artifact is identified by using the power spectrum ratio method and the acceleration signal feature analysis method, and the signal quality is evaluated in multiple dimensions, and only the signal segments with qualified quality are reserved for subsequent processing;

[0033] Step two, adaptive noise reduction and multi-modal time sequence alignment:

[0034] Adaptive signal decomposition algorithm is used to perform adaptive noise reduction processing on various physiological signals, the signal envelope feature is calculated through Hilbert transform, and the alternating direction multiplier method is used to solve the variational constraint problem, and high-quality signals are reconstructed; the dynamic time warping algorithm is used to calculate the Euclidean distance matrix of the remaining signals and the reference signal with the electrocardiosignal as the time sequence reference, the optimal alignment path is searched through dynamic programming, and the accurate resampling and time sequence alignment of the multi-modal signals are completed through cubic spline interpolation;

[0035] Step three, parallel feature extraction in high-dimensional time-frequency domain and nonlinearity:

[0036] Local time-domain and frequency-domain features of the signal are extracted using the time convolution network, which adopts a dilated causal convolution structure with an exponential dilation factor and combines a multi-scale pooling strategy of maximum pooling and average pooling; long-term dependencies and nonlinear features of the signal are extracted in parallel using the gated attention recurrent unit structure, attention weights are calculated through historical state similarity, and state information flow is controlled using update gates and reset gates; and the feature vector synthesizer finally synthesizes the above features into a unified high-dimensional feature vector;

[0037] Step four, multi-modal feature fusion based on mutual information weighting:

[0038] Based on the electrocardiogram signal features as the benchmark reference, the correlation between different modal features is calculated through the improved multi-head self-attention mechanism; the mutual information entropy between the modes is estimated using a non-parametric method based on K-nearest neighbors, the joint probability distribution is calculated using kernel density estimation, and the estimation bias is corrected using the Jackknife resampling technique; based on the mutual information entropy value, appropriate weights are assigned to each modal feature, point multiplication is used to realize feature weighting, and the final multi-modal fusion feature tensor is generated through splicing operation;

[0039] Step five, calculation and prediction of dynamic risk index:

[0040] The abnormal score of the current fusion feature vector is calculated based on the kernel density estimation algorithm, and the adaptive bandwidth optimization technique is used to improve the accuracy of probability density estimation; combined with the historical state information saved in the risk state memory, the state space model and the deep learning hybrid architecture are used to model the time evolution process of risk, the state transition is described using a linear dynamic system, and the model parameters are optimized using the expectation maximization algorithm, and the dynamic risk index is output;

[0041] Step six, multi-level early warning triggering and real-time feedback:

[0042] According to the size of the dynamic risk index, the corresponding warning level is triggered according to the preset multi-level warning protocol; through the human-computer interaction interface, real-time generation and visualization of feedback information is realized, including risk index curve, warning level prompt and health suggestion, forming a complete monitoring-evaluation-warning closed loop.

[0043] In the fourth step, the mutual information entropy I(X;Y) between the modes is calculated as follows:

[0044] Discretize the continuous feature vectors X and Y based on K-nearest neighbors, and divide the feature space into multiple local regions;

[0045] The joint probability distribution is calculated using a non-parametric kernel density estimation method based on K-nearest neighbors and the marginal probability distribution The estimation formula is:

[0046]

[0047] In formula (1), n ​​represents the total number of samples used for estimation, K(·) is the Gaussian kernel function, and h x and h y These are the adaptive bandwidths for features X and Y, respectively, where x represents the target estimation point for feature X. i Let |x| represent the actual observed value of feature X in the i-th sample. i || represents the Euclidean distance between the target estimated point of feature X and the i-th sample observation. Represents the joint bandwidth normalization coefficient;

[0048] The mutual information entropy is ultimately calculated using the following deviation correction formula:

[0049]

[0050] In formula (2), This is a double summation operator. Indicates that feature X takes x i Feature Y takes y i The joint probability distribution estimate at time , Indicates that feature X takes x i Marginal probability distribution estimate at time, Indicates that feature Y takes y i The marginal probability distribution estimate at time , ∈ is a small positive value, log is the logarithmic operator, λ represents the bias correction coefficient, MI corr The deviation correction term represents the mutual information entropy;

[0051] The calculated mutual information entropy value is used for weight allocation in subsequent feature weighted fusion. The weight calculation satisfies:

[0052] In step five, the anomaly score S(x) is calculated through the following process:

[0053] The feature distribution of normal historical samples is modeled using a Gaussian kernel density estimation method based on local adaptive bandwidth. The probability density estimation function is:

[0054]

[0055] In formula (3), This represents the probability density estimate of the characteristics of normal historical samples, where n is the total number of normal historical samples. It is the summation operator, d represents the dimension of the eigenvector, and H... i H represents the local adaptive bandwidth matrix corresponding to the i-th normal sample.i is the determinant of the bandwidth matrix H i , denotes the inverse matrix of the bandwidth matrix H i , x is the target feature vector to be evaluated, x i is the feature vector of the i-th normal historical sample, (x-x i ) T is the transpose of the difference vector (x-x i ), exp(·) is the exponential function, is the normalization coefficient of the Gaussian kernel function;

[0056] The calculation formula of the local adaptive bandwidth matrix H

[0057] H i = h·∑ i (4)

[0058] In formula (4), h is a global bandwidth parameter, ∑ i is a local covariance matrix centered at the sample point x i ;

[0059] The anomaly score is defined as the normalized negative log probability density combined with the historical risk trend:

[0060]

[0061] In formula (5), α and β are weight coefficients, R t , R t-1 are the risk indexes at the current and previous moments respectively, σ R is the standard deviation of the historical risk index.

[0062] The present application has the following beneficial technical effects:

[0063] The sudden death risk real-time evaluation system based on multi-modal physiological signal fusion disclosed in the application can effectively solve the problems of poor time synchronization and insufficient data integrity of multi-source signals, realize high-precision synchronous acquisition and timestamp calibration of multi-modal physiological data, and lay a foundation for multi-source information fusion through multi-modal physiological signal synchronous acquisition and time alignment technology; the multi-dimensional feature parallel extraction technology can synchronously extract time domain, frequency domain and nonlinear features of each modal signal, overcome the defects of single feature dimension and insufficient information mining of traditional technology, and improve risk correlation feature coverage and recognition; the dynamic weighted fusion algorithm based on modal correlation can adjust the fusion weight according to the contribution difference of each modal signal in risk evaluation, realize the complementary advantages of multi-source features, solve the problems of fixed weight and poor adaptability of existing fusion technology, and improve the rationality of fusion; the online adaptive update risk quantization model and grading early warning mechanism can dynamically optimize model parameters based on real-time data and realize multi-level early warning, solve the limitations of insufficient adaptability of traditional static models, significantly improve the real-time performance, accuracy and early warning refinement degree of risk evaluation, provide reliable support for early identification and timely intervention of sudden death risk, effectively improve the clinical application effect of traditional technology, and have important clinical practical value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings:

[0065] Figure 1 The flow chart of the sudden death risk implementation evaluation system based on multi-modal physiological signal fusion of the application;

[0066] Figure 2 The signal acquisition and preprocessing flow chart of the sudden death risk implementation evaluation system based on multi-modal physiological signal fusion of the application;

[0067] Figure 3 The high-dimensional feature extraction and multi-modal fusion flow chart of the sudden death risk implementation evaluation system based on multi-modal physiological signal fusion of the application;

[0068] Figure 4 The sudden death risk quantization and early warning feedback flow chart of the sudden death risk implementation evaluation system based on multi-modal physiological signal fusion of the application;

[0069] Figure 5 The application scene diagram of the sudden death risk implementation evaluation system based on multi-modal physiological signal fusion of the application. DETAILED EMBODIMENT

[0070] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, it should be understood that the embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application;

[0071] Example 1:

[0072] A sudden death risk real-time evaluation system based on multi-modal physiological signal fusion is applied to the field of human physiological state monitoring and early warning: the system comprises a signal acquisition module 100, a signal preprocessing module 200, a high-dimensional feature extraction module 300, a multi-modal feature fusion module 400, a sudden death risk quantification module 500, a model updating engine 600 and an early warning feedback module 700;

[0073] The signal acquisition module 100 adopts a multi-channel synchronous acquisition technology architecture to synchronously acquire electrocardio signals, skin electricity signals, respiratory signals and blood volume pulse signals, and maintain the time synchronization and data integrity of multi-modal physiological signals; the signal acquisition module 100 comprises a bioelectricity signal electrode array 110, an optical sensing unit 120 and an impedance detection circuit 130, and completes the synchronous acquisition of multi-source signals through a special interface protocol;

[0074] The signal preprocessing module 200 integrates an adaptive filter 210 and a quality evaluation timing unit 220, realizes adaptive noise reduction of signals through an adaptive signal decomposition algorithm, and completes the timing alignment of multi-modal signals through a dynamic time warping technology; the signal preprocessing module 200 comprises a signal input interface 230, a variational modal decomposer 240, a dynamic threshold processor 250, a timing calibrator 260, a signal quality evaluator 270 and a preprocessing output interface 280, and constitutes a complete signal preprocessing pipeline;

[0075] The high-dimensional feature extraction module 300 fuses a timing convolution network 310 and a gated recurrent unit 320, and extracts the time domain features, frequency domain features and nonlinear features of signals in parallel to generate a high-dimensional feature vector space; the high-dimensional feature extraction module 300 comprises a feature extraction controller 330, a timing convolution processor 340, a recurrent neural network unit 350 and a feature vector synthesizer 360;

[0076] The multi-modal feature fusion module 400 adopts an improved multi-head self-attention mechanism 410, takes electrocardio signal features as a benchmark reference, realizes intelligent weighted fusion of multi-modal features through calculation of inter-modal mutual information entropy; the multi-modal feature fusion module 400 comprises an attention weight calculation unit 420, a feature projection transformer 430, a mutual information entropy analyzer 440 and a fusion feature synthesizer 450, and constitutes a multi-level feature fusion architecture;

[0077] The sudden death risk quantification module 500 integrates the time sequence anomaly detector 510 and the risk regressor 520, calculates an anomaly score based on a kernel density estimation algorithm, and generates a dynamic risk index in combination with a historical risk state; the sudden death risk quantification module 500 comprises an anomaly detection controller 530, a kernel density estimation unit 540, a risk state storage 550, and a regression predictor 560;

[0078] The model updating engine 600 adopts an online sequence learning strategy, and realizes incremental updating and optimization of model parameters through a regularized least square algorithm; the model updating engine 600 comprises an incremental learning controller 610, a weight updater 620, a parameter optimizer 630, and a model verification unit 640;

[0079] The early warning feedback module 700 triggers a multi-level early warning protocol according to a risk index classification, and completes real-time feedback and display of an evaluation result; the early warning feedback module 700 comprises a risk level determiner 710, an early warning protocol executor 720, a feedback information generator 730, and a human-computer interaction interface 740.

[0080] To solve the technical bottlenecks of traditional sudden death risk assessment, such as single mode dependence, lack of synchronization, and insufficient feature mining, the system modules cooperatively achieve technical breakthroughs: the signal acquisition module 100 adopts a multi-channel synchronous acquisition architecture, integrates a bioelectric signal electrode array 110, an optical sensing unit 120, and an impedance detection circuit 130, synchronously acquires electrocardiogram, skin electricity, respiration, and blood volume pulse signals through a special interface protocol, ensures the time synchronization and data integrity of multi-modal signals, and breaks through the limitations of single signal evaluation dimension; the signal preprocessing module 200 integrates an adaptive filter 210 and a quality evaluation timing unit 220, realizes adaptive noise reduction with a variational modal decomposer 240, completes signal timing alignment with a dynamic time warping technology of a timing calibrator 260, cooperates with a signal quality evaluator 270 to build a preprocessing pipeline for multi-dimensional evaluation, and solves the problems of multi-source signal noise interference and time misalignment; the high-dimensional feature extraction module 300 fuses a timing convolution network 310 and a gated recurrent unit 320, expands causal convolution with an exponential expansion factor of a timing convolution processor 340, and uses a gating attention mechanism of a recurrent neural network unit 350 to extract signal time domain, frequency domain, and nonlinear features in parallel, and generates a high-dimensional feature vector space; the multi-modal feature fusion module 400 takes an improved multi-head self-attention mechanism 410 as the core, calculates inter-modal mutual information entropy through a mutual information entropy analyzer 440, intelligently weights through an attention weight calculation unit 420, and completes multi-level fusion through a feature projection transformer 430 and a fusion feature synthesizer 450 to realize the complementary advantages of multi-source features; the sudden death risk quantification module 500 integrates a timing anomaly detector 510 and a risk regressor 520, calculates an abnormal score based on a kernel density estimation unit 540, combines the historical state stored in a risk state storage 550, and generates a dynamic risk index through a regression predictor 560; the model updating engine 600 adopts an online sequence learning strategy, realizes incremental updating of model parameters through a regularization least squares algorithm of an incremental learning controller 610 and a weight updater 620, and solves the problem of insufficient adaptability of static models; the early warning feedback module 700 triggers a multi-level early warning protocol through a risk level judge 710, realizes real-time feedback of evaluation results through a human-computer interaction interface 740, and improves the early warning grading accuracy and response timeliness.

[0081] Embodiment 2:

[0082] The signal preprocessing module 200 includes a signal input interface 230, a variational modal decomposer 240, a dynamic threshold processor 250, a timing calibrator 260, a signal quality evaluator 270, and a preprocessing output interface 280;

[0083] The variational mode decomposer 240 adopts an adaptive bandwidth optimization algorithm architecture, including a wave generator 241, a mode decomposer 242, and a reconstruction filter 243; the wave generator 241 calculates signal envelope characteristics through Hilbert transform; the mode decomposer 242 solves a variational constraint problem through an alternating direction multiplier method; and the reconstruction filter 243 reconstructs a high-quality signal after denoising through inverse transform.

[0084] The timing calibrator 260 is implemented by using a dynamic time warping algorithm, including a distance matrix calculation unit 261, a warping path searcher 262, and a time scaler 263; the distance matrix calculation unit 261 measures signal similarity characteristics by using Euclidean distance; the warping path searcher 262 finds an optimal alignment path by using a dynamic programming algorithm; and the time scaler 263 realizes signal resampling by using a cubic spline interpolation algorithm.

[0085] The signal quality evaluator 270 integrates a multi-dimensional quality index evaluation system, including a signal-to-noise ratio calculation unit 271, a waveform distortion detection unit 272, and a motion artifact identification unit 273; the signal-to-noise ratio calculation unit 271 calculates a signal quality index by using a power spectral ratio method; the waveform distortion detection unit 272 calculates a waveform correlation coefficient by using a template matching algorithm; and the motion artifact identification unit 273 performs artifact compensation processing based on acceleration signal characteristics.

[0086] The signal preprocessing module 200 receives the multi-modal physiological signal through the signal input interface 230, and realizes adaptive noise reduction by the variational modal decomposer 240. The variational modal decomposer 240 adopts an adaptive bandwidth optimization algorithm architecture. The wave generator 241 calculates the signal envelope characteristics through Hilbert transform to capture the signal amplitude variation law. The modal decomposer 242 iteratively solves the variational constraint problem by using the alternating direction multiplier method, and decomposes the original signal into a plurality of intrinsic modal functions. The reconstruction filter 243 integrates the effective modal functions through inverse transform to reconstruct the high-quality signal after noise reduction. The dynamic threshold processor 250 performs threshold screening on the denoised signal to eliminate redundant interference components. The time sequence calibrator 260 adopts a dynamic time warping algorithm. The distance matrix calculation unit 261 quantizes the similarity between different modal signal sampling points by using the Euclidean distance, and generates a distance matrix. The warping path searcher 262 traverses the distance matrix by using a dynamic programming algorithm to find the optimal time alignment path that minimizes the cumulative distance. The time scaler 263 resamples the signal based on a cubic spline interpolation algorithm to realize the time axis synchronization of the multi-modal signal. The signal quality evaluator 270 constructs a multi-dimensional evaluation system. The signal-to-noise ratio calculation unit 271 calculates the power spectral density ratio of the signal and the noise by using the power spectral ratio method. The waveform distortion detection unit 272 calculates the correlation coefficient of the actual signal and the standard template by using a template matching algorithm. The motion artifact identification unit 273 locates the artifact interval according to the acceleration signal characteristics and compensates. Finally, the qualified signal is output through the preprocessing output interface 280.

[0087] Embodiment 3:

[0088] The high-dimensional feature extraction module 300 comprises a feature extraction controller 330, a time sequence convolution processor 340, a recurrent neural network unit 350, and a feature vector synthesizer 360.

[0089] The time sequence convolution processor 340 adopts an expansion causal convolution architecture with an exponential expansion factor, and comprises a convolution kernel optimization unit 341, a feature mapping unit 342, and a pooling layer 343. The convolution kernel optimization unit 341 optimizes the convolution kernel parameters by using an adaptive gradient descent algorithm. The feature mapping unit 342 realizes feature transformation by using a nonlinear activation function. The pooling layer 343 adopts a multi-scale pooling strategy combining maximum pooling and average pooling.

[0090] The recurrent neural network unit 350 adopts a gated attention recurrent unit structure, and comprises a time attention mechanism 351, a gated update unit 352, and a state memory unit 353. The time attention mechanism 351 calculates the attention weight through the historical state similarity. The gated update unit 352 integrates the calculation processes of the update gate and the reset gate. The state memory unit 353 stores the time sequence feature information by using a long short-term memory structure.

[0091] The high-dimensional feature extraction module 300 works in coordination through the feature extraction controller 330, the timing convolution processor 340 adopts an expansion causal convolution architecture with an exponential expansion factor, the convolution kernel optimization unit 341 dynamically optimizes the convolution kernel parameters with an adaptive gradient descent algorithm, and the feature extraction is enhanced in pertinence; the feature mapping unit 342 realizes nonlinear transformation of features through a nonlinear activation function, and improves the feature discrimination; the pooling layer 343 combines the multiscale strategy of maximum pooling and average pooling to retain key features while reducing the dimension. The recurrent neural network unit 350 adopts a gated attention recurrent structure, the time attention mechanism 351 calculates the weight through the similarity of historical states, and highlights important time sequence information; the gated update unit 352 integrates the update gate and reset gate operations to dynamically adjust information transmission; the state memory unit 353 stores the time sequence features in a long short-term memory structure to capture long-term dependencies. Finally, the feature vector synthesizer 360 fuses the two types of features to construct a high-dimensional feature vector space.

[0092] Embodiment 4:

[0093] The multi-modal feature fusion module 400 includes an attention weight calculation unit 420, a feature projection transformer 430, an mutual information entropy analyzer 440, and a fusion feature synthesizer 450.

[0094] The mutual information entropy analyzer 440 adopts a non-parametric mutual information estimation structure based on K-nearest neighbors, including a probability density estimation unit 441, an entropy value calculation unit 442, and a bias correction unit 443; the probability density estimation unit 441 calculates the joint probability distribution using the kernel density estimation method; the entropy value calculation unit 442 calculates the mutual information entropy value through the log-likelihood ratio; and the bias correction unit 443 corrects the estimation bias using the Jackknife resampling technique.

[0095] The fusion feature synthesizer 450 adopts a weighted feature fusion architecture, including a weight distribution unit 451, a feature weighting unit 452, and a fusion output unit 453; the weight distribution unit 451 distributes the weight of each modal feature according to the mutual information entropy value; the feature weighting unit 452 realizes feature weighting through point multiplication operation; and the fusion output unit 453 generates the final fusion feature tensor through splicing operation.

[0096] The present application realizes cross-modal information integration through a multi-modal feature fusion module 400, an attention weight calculation unit 420 initializes the weight distribution, and a feature projection transformer 430 maps each modal feature to a unified dimensional space. The mutual information entropy analyzer 440 adopts a K-nearest neighbor non-parametric estimation architecture, the probability density estimation unit 441 models the inter-modal joint probability distribution using a kernel density estimation method, the entropy value calculation unit 442 quantifies the modal correlation by deriving the mutual information entropy value through the log-likelihood ratio, and the bias correction unit 443 reduces the estimation bias using the Jackknife resampling technique. The fusion feature synthesizer 450 is based on a weighted architecture, the weight distribution unit 451 dynamically adjusts the weight of each modal feature according to the mutual information entropy value, the feature weighting unit 452 realizes feature importance modulation through point multiplication operation, and the fusion output unit 453 generates a fusion feature tensor containing multi-modal collaborative information through splicing operation.

[0097] Embodiment 5:

[0098] The sudden death risk quantification module 500 includes an anomaly detection controller 530, a kernel density estimation unit 540, a risk state memory 550, and a regression predictor 560.

[0099] The anomaly detection controller 530 adopts a multi-index anomaly detection architecture, including a threshold adaptive unit 531, an anomaly scorer 532, and an alarm triggering unit 533; the threshold adaptive unit 531 dynamically adjusts the anomaly threshold according to the historical data distribution; the anomaly scorer 532 calculates the anomaly score using the density estimation method; and the alarm triggering unit 533 triggers a hierarchical warning mechanism according to the scoring result;

[0100] The risk state memory 550 adopts a spatio-temporal state storage architecture, including a state encoder 551, a memory matrix 552, and a state retriever 553; the state encoder 551 encodes historical state information through a recurrent neural network; the memory matrix 552 stores key state features using an attention mechanism; and the state retriever 553 retrieves relevant historical states through similarity matching;

[0101] The regression predictor 560 adopts a hybrid architecture of state space model and deep learning, including a state transition unit 561, an observation model unit 562, and a parameter optimization unit 563; the state transition unit 561 models the state evolution process using a linear dynamic system; the observation model unit 562 establishes the mapping relationship between the state and the observation through a deep learning network; and the parameter optimization unit 563 optimizes the model parameters using the expectation maximization algorithm.

[0102] The application realizes dynamic risk assessment through the sudden death risk quantification module 500, the abnormality detection controller 530 adopts a multi-index architecture, the threshold self-adaptive unit 531 dynamically adjusts the abnormal threshold according to historical data distribution, the abnormality scorer 532 calculates the abnormal score by using a density estimation method, and the alarm triggering unit 533 starts a graded early warning according to the score. The risk state memory 550 adopts a space-time architecture, the state encoder 551 encodes the historical state through a recurrent neural network, the memory matrix 552 stores key features by using an attention mechanism, and the state retriever 553 retrieves relevant historical states through similarity matching. The regression predictor 560 adopts a hybrid architecture, the state transition unit 561 models state evolution by using a linear dynamic system, the observation model unit 562 constructs a state-observation mapping by using deep learning, the parameter optimizer 563 optimizes parameters by using an expectation maximization algorithm, and finally a dynamic risk index is generated.

[0103] Embodiment 6:

[0104] The model updating engine 600 comprises an incremental learning controller 610, a weight updater 620, a parameter optimizer 630 and a model verification unit 640.

[0105] The incremental learning controller 610 adopts an online sequence extreme learning machine architecture and comprises a hidden node management unit 611, an output weight calculation unit 612 and a model updating unit 613. The hidden node management unit 611 dynamically adds or deletes hidden layer nodes and optimizes node parameters. The output weight calculation unit 612 calculates output layer weights by using a recursive least squares method. The model updating unit 613 realizes incremental updating of the model by using an online learning algorithm.

[0106] The weight updater 620 adopts a recursive least squares architecture with a forgetting factor and comprises a forgetting factor adjustment unit 621, a covariance matrix updating unit 622 and a weight correction unit 623. The forgetting factor adjustment unit 621 dynamically adjusts the size of the forgetting factor according to the prediction error. The covariance matrix updating unit 622 updates the covariance matrix by using a rank-one update algorithm. The weight correction unit 623 corrects the weight parameters by using a gradient descent method.

[0107] The application realizes parameter dynamic optimization through the model updating engine 600, the incremental learning controller 610 adopts an online sequence extreme learning machine architecture, the hidden node management unit 611 dynamically adjusts the number of hidden layer nodes and optimizes parameters, the output weight calculation unit 612 solves the output layer weight by the recursive least square method, and the model updating unit 613 completes incremental updating of the model through an online learning algorithm. The weight updater 620 adopts a recursive least square architecture with a forgetting factor, the forgetting factor adjustment unit 621 dynamically adjusts the forgetting factor according to a prediction error, the covariance matrix updating unit 622 iteratively updates the covariance matrix by a rank-one updating algorithm, the weight correction unit 623 realizes accurate correction of the weight parameter by the gradient descent method, and the parameter optimizer 630 and the model verification unit 640 form a complete model optimization closed loop.

[0108] Embodiment 7

[0109] A sudden death risk real-time evaluation method based on multi-modal physiological signal fusion is applied to the field of human physiological state monitoring and early warning, and the method comprises the following steps:

[0110] Step one, multi-modal physiological signal synchronous acquisition and quality evaluation:

[0111] Through the multi-channel synchronous acquisition architecture composed of the bioelectric signal electrode array 110, the optical sensing unit 120 and the impedance detection circuit 139, the electrocardiosignal, the skin electricity signal, the respiration signal and the blood volume pulse signal of the user are synchronously acquired; the signal-to-noise ratio of each signal is calculated in real time and the motion artifact is identified by adopting the analysis method based on the power spectrum ratio method and the acceleration signal feature, the signal quality is evaluated in multiple dimensions, and only the signal segment with qualified quality is reserved for subsequent processing;

[0112] Step two, adaptive noise reduction and multi-modal time sequence alignment:

[0113] Adaptive signal decomposition algorithm is adopted to perform adaptive noise reduction processing on various physiological signals, the signal envelope feature is calculated through Hilbert transform, and the alternating direction multiplier method is used to solve the variational constraint problem, and high-quality signals are reconstructed; the dynamic time warping algorithm is adopted, the electrocardiosignal is taken as the time sequence reference, the Euclidean distance matrix of the remaining signals and the reference signal is calculated, the optimal alignment path is searched through dynamic programming, and the accurate resampling and time sequence alignment of the multi-modal signals are completed by using the cubic spline interpolation;

[0114] Step three, parallel feature extraction in high-dimensional time-frequency domain and nonlinearity:

[0115] The local time domain features and frequency domain features of the signal are extracted using the timing convolution network 310, which adopts a dilated causal convolution structure with an exponential expansion factor and combines a multi-scale pooling strategy of maximum pooling and average pooling; the long-term dependence and nonlinear features of the signal are extracted in parallel using the gated attention recurrent unit 320 structure, the attention weight is calculated through the historical state similarity, and the state information flow is controlled by the update gate and the reset gate; finally, the feature vector synthesizer 360 combines the above features into a unified high-dimensional feature vector;

[0116] Step four, multi-modal feature fusion based on mutual information weighting:

[0117] Based on the electrocardiogram signal features as the benchmark reference, the correlation between different modal features is calculated through the improved multi-head self-attention mechanism; the mutual information entropy between modalities is estimated using a non-parametric method based on K nearest neighbors, the joint probability distribution is calculated using kernel density estimation, and the estimation bias is corrected using the Jackknife resampling technique; based on the mutual information entropy value, appropriate weights are assigned to each modal feature, point multiplication is used to realize feature weighting, and the final multi-modal fusion feature tensor is generated through splicing operation;

[0118] Step five, calculation and prediction of dynamic risk index:

[0119] The abnormal score of the current fusion feature vector is calculated based on the kernel density estimation algorithm, and the adaptive bandwidth optimization technique is used to improve the accuracy of probability density estimation; combined with the historical state information saved in the risk state memory, the state space model and deep learning hybrid architecture are used to model the time evolution process of risk, the state transition is described by a linear dynamic system, and the model parameters are optimized using the expectation maximization algorithm, and the dynamic risk index is output;

[0120] Step six, multi-level early warning triggering and real-time feedback:

[0121] According to the size of the dynamic risk index, the corresponding warning level is triggered according to the preset multi-level warning protocol; through the human-computer interaction interface, real-time generation and visualization of feedback information is realized, including risk index curve, warning level prompt and health suggestion, forming a complete monitoring-evaluation-warning closed loop.

[0122] To verify the effectiveness of the sudden death risk real-time evaluation method based on multi-modal physiological signal fusion, a control experiment is designed, and the physiological signals of healthy people and high-risk people are collected, and the whole process is completed according to the method steps, and the effect of each step and the final risk evaluation accuracy are presented by data, and the experiment is as follows:

[0123] Experimental preparation:

[0124] Experimental subjects and equipment: Two types of experimental subjects were selected: healthy group (120, no history of cardiovascular disease, age 20-55 years old), high-risk group (30, diagnosed with arrhythmia or family history of sudden death, age 35-65 years old); The collection equipment uses a multi-channel synchronous collection device containing a bioelectric signal electrode array 110, an optical sensing unit 120, and an impedance detection circuit 139, with a sampling frequency of 250 Hz to ensure signal time synchronization.

[0125] Parameter preset: In step one, the signal quality evaluation, the signal-to-noise ratio threshold is set to 30 dB, and the motion artifact recognition threshold is calibrated to 0.5g based on the acceleration signal characteristics; In step two, the step size of the Euclidean distance matrix calculation of the dynamic time warping algorithm is 0.1s; In step three, the initial value of the exponential expansion factor of the time convolution network is 2, and the multi-scale pooling window size is set to 3x3; In step four, the K value of the K nearest neighbor non-parametric estimation is 25, and the Jackknife resampling number is 5; In step five, the adaptive bandwidth of the kernel density estimation is determined by the Silverman rule, and the iteration number of the expectation maximization algorithm is 100; In step six, the warning protocol is set to three levels: risk index < 2.0 is no warning, 2.0 ≤ risk index < 4.0 is low-risk warning, and risk index ≥ 4.0 is high-risk warning.

[0126] Experimental steps:

[0127] Signal collection and quality evaluation: Collect 12-hour signals from two types of objects, calculate the signal-to-noise ratio using the power spectral ratio method, and identify motion artifacts using acceleration characteristics. Retain qualified signals (1152 hours for the healthy group and 324 hours for the high-risk group).

[0128] Noise reduction and time alignment: Use adaptive signal decomposition algorithm for noise reduction (ECG signal-to-noise ratio for the healthy group is improved to 42dB), and use ECG as the reference to achieve multi-modal signal time alignment through dynamic time warping and cubic spline interpolation.

[0129] High-dimensional feature extraction: Time convolution network extracts local time-frequency domain features, and gated attention recurrent unit extracts long-term nonlinear features, resulting in a 64-dimensional high-dimensional feature vector (feature variance for the high-risk group is 1.56, which is larger than that for the healthy group 0.82).

[0130] Feature fusion: Use ECG as the reference to improve the multi-head self-attention algorithm for modality correlation, use K nearest neighbor non-parametric method to estimate mutual information entropy (error after correction is 6%), assign weights according to entropy values (ECG 0.35, skin electricity 0.28, etc.), and use point multiplication and weighting to generate a fusion tensor.

[0131] Risk index calculation: Kernel density estimation calculates abnormal scores (mean value for the high-risk group is 4.8), combines historical states, and uses state space model + expectation maximization algorithm to output dynamic risk index (mean value for the healthy group is 1.5, and for the high-risk group is 4.2).

[0132] Early warning and feedback: Early warning triggered by risk index (86.7% of high-risk group with high early warning), real-time display of curve, early warning level and suggestion by human-computer interaction interface, feedback delay <1s.

[0133] Experimental results and analysis:

[0134] Table 1: Key indicator data of each step

[0135]

[0136] Method performance comparison:

[0137] Compared with traditional single electrocardio signal evaluation method: The sudden death risk identification accuracy of the method is 92.3%, and that of the traditional method is 68.5%; the early warning delay is <1s, and that of the traditional method is 3-5s; the multi-modal fusion reduces the false positive rate from 22% to 8% and the false negative rate from 18% to 3%, fully demonstrating the advantages of the method in evaluation accuracy, timeliness and reliability.

[0138] Experimental conclusion:

[0139] Through the control verification of the healthy group and the high-risk group, it is shown that the sudden death risk real-time evaluation method based on multi-modal physiological signal fusion is feasible in each step logic and sufficient in data support: from signal acquisition to early warning feedback, the whole process can effectively screen high-quality signals, extract high-dimensional features, realize multi-modal intelligent fusion, accurately calculate the dynamic risk index and trigger graded early warning, and the evaluation accuracy, real-time performance and clinical applicability are better than those of the traditional single modal method, which can provide reliable technical support for human physiological state monitoring and sudden death risk early warning.

[0140] Example 8:

[0141] In the fourth step, the inter-modal mutual information entropy I(X;Y) is calculated by the following method:

[0142] Discretize the continuous feature vectors X and Y based on K-nearest neighbors, and divide the feature space into multiple local regions;

[0143] Calculate the joint probability distribution based on K-nearest neighbor non-parametric kernel density estimation method and the marginal probability distribution The estimation formula is:

[0144]

[0145] In formula (1), n represents the total number of samples for estimation, K(·) is a Gaussian kernel function, h x and h y are the adaptive bandwidths of features X and Y, respectively, x represents the target estimation point of feature X, and xi represents the actual observation value of feature X in the i-th sample, ||x-x i represents the Euclidean distance between the target estimated point of feature X and the observation value of the i-th sample, represents the joint bandwidth normalization coefficient;

[0146] The mutual information entropy is finally calculated by the following bias correction formula:

[0147]

[0148] In formula (2), is a double summation operator, represents the joint probability distribution estimate when feature X takes x i , and feature Y takes y i represents the marginal probability distribution estimate when feature X takes x i represents the marginal probability distribution estimate when feature Y takes y i , and ∈ is a small positive value, log is a logarithmic operator, λ represents a bias correction coefficient, MI corr represents the bias correction term of mutual information entropy;

[0149] The calculated mutual information entropy value is used for weight distribution in subsequent feature weighting fusion, and the weight calculation satisfies:

[0150] To verify the effectiveness of the calculation process and weight distribution of inter-modal mutual information entropy in step four, experiments are designed to take multi-modal physiological signal features as the object, and the complete process of discretization processing, probability distribution estimation, bias correction and weight calculation is presented through real data. The experiment is as follows:

[0151] Experiment preparation:

[0152] Sample and feature selection: 100 healthy adults' multi-modal physiological signals are collected, and two types of core modal features are extracted after preprocessing-ECG feature X (RR interval coefficient of variation) and GSR feature Y (skin conductance activity fluctuation amplitude), a total of 800 valid samples (8 groups of synchronous feature data per person); In addition, respiratory feature Z (respiratory frequency variation) and blood pulse feature W (pulse peak intensity) are selected to form another 800 samples for multi-modal cross-validation.

[0153] Parameter setting: the K value of K-nearest neighbor discretization is set to 20 (balance local area partitioning accuracy and data representativeness); The adaptive bandwidth h x , h γ ​​h_z, h_w are calibrated by Silverman's rule combined with sample distribution, which are 0.8, 0.75, 0.82, 0.78 respectively; the tiny positive value ε is 10 -8 (avoiding zero value in logarithm calculation); the bias correction coefficient λ is determined as 0.12 through cross-validation, and the bias correction term MI_corr is calculated through Jackknife resampling method (based on the error mean of 5 repeated sampling).

[0154] Experimental steps:

[0155] Feature discretization processing: taking each sample as the center, 20 adjacent samples are selected to form a local area, and the continuous features are divided into 15 discrete intervals according to the feature value distribution density (such as X discrete intervals [0.02, 0.04), [0.04, 0.06)…, Y discrete intervals [1.2, 1.4), [1.4, 1.6)…), to realize the discretization conversion of continuous features, and Z and W are processed in the same way.

[0156] Probability distribution estimation: for the discrete intervals of X and Y, the joint probability distribution and marginal probability distribution are calculated by non-parametric kernel density estimation. Taking x = 0.05 as the target estimation point of X and y = 1.5 as the target estimation point of Y as an example, 800 samples are traversed, the similarity (the closer the distance, the higher the similarity) between each sample and the target point is calculated by Gaussian kernel function, and the joint bandwidth normalization coefficient is combined to accumulate the joint probability distribution estimation value The discrete intervals of X and Y are traversed separately, and the marginal probability distribution estimation value is calculated respectively.

[0157] Mutual information entropy calculation: the is substituted into the bias correction formula, the logarithmic value of each discrete interval is calculated first, and then multiplied by the corresponding After accumulation, λ·MI corr (correction of estimation bias) is added to get the mutual information entropy I(X; Y) of X and Y; repeat the above process to calculate I(X; Z), I(X; W), I(Y; Z), I(Y; W), I(Z; W).

[0158] Weight distribution verification: taking the electrocardiogram feature X as the reference, the I(X; Y), I(X; Z), I(X; W) are substituted into the weight formula to calculate the fusion weight of Y, Z, W relative to X, and verify the positive correlation between the weight and the mutual information entropy (the higher the mutual information entropy, the greater the weight, the stronger the modal correlation).

[0159] Experimental data and results:

[0160] Mutual information entropy calculation results:

[0161] Table 2: Mutual information entropy values between different modal features

[0162]

[0163] As can be seen from the table, the mutual information entropy between the electrocardiogram feature X and the skin electricity feature Y is the highest, indicating that the two are the most relevant in reflecting physiological state changes, which is consistent with the physiological mechanism that "electrocardiogram and skin electricity activity are regulated by autonomic nerves" in clinical practice. The mutual information entropy between Y and Z is the lowest, reflecting that the correlation between respiration and skin electricity activity is weak, which is consistent with the physiological logic.

[0164] Weight distribution results:

[0165] Table 3: Fusion weight calculation

[0166] Modal feature Mutual information entropy I(X; Y) Weight ω Y 1.82 0.41 Z 1.25 0.28 W 1.53 0.31

[0167] As can be seen from the table, the total weight is 1.0, and the weight of Y with the highest mutual information entropy is the largest (0.41), and the weight of Z with the lowest mutual information entropy is the smallest (0.28), which fully complies with the distribution rule that "mutual information entropy determines weight", proving that the mutual information entropy calculation result can effectively guide the weight setting of multi-modal feature fusion.

[0168] Bias correction effect verification:

[0169] Table 4: Comparison of mutual information entropy values with and without bias correction

[0170] Calculation method Mutual information entropy value Deviation from true correlation Unbiased correction 2.15 18.1％ Biased correction 1.82 4.3％

[0171] As can be seen from the table, after correction by λ·MI corr , the mutual information entropy value is closer to the real modality correlation verified by clinical verification, and the bias is reduced by 13.8%, proving that the bias correction formula can effectively improve the calculation accuracy.

[0172] Experimental conclusion:

[0173] Through the real multi-modal physiological feature data verification, the calculation process of mutual information entropy between modalities in step four can accurately quantify the modality correlation, and the weight distribution rule based on mutual information entropy complies with the multi-modal fusion logic, providing reliable quantitative basis for subsequent feature weighted fusion. At the same time, the bias correction mechanism significantly improves the calculation accuracy of mutual information entropy, ensuring that the result can truly reflect the internal correlation between physiological features.

[0174] Example 9:

[0175] In the fifth step, the abnormal score S(x) is calculated by the following process:

[0176] The feature distribution of normal historical samples is modeled by using a local adaptive bandwidth-based Gaussian kernel density estimation method, and the probability density estimation function is:

[0177]

[0178] In formula (3), represents the probability density estimation value of the normal historical sample features, n is the total number of normal historical samples, is a summation operator, d represents the dimension of the feature vector, H i represents the local adaptive bandwidth matrix corresponding to the i-th normal sample, |H i is the determinant of the bandwidth matrix H i , represents the inverse matrix of the bandwidth matrix H i , x is the target feature vector to be evaluated, x i is the feature vector of the i-th normal historical sample, (x-x i ) T is the transpose of the difference vector (x-x i ), exp(·) is the exponential function, is the normalization coefficient of the Gaussian kernel function;

[0179] The calculation formula of the local adaptive bandwidth matrix H

[0180] H i = h·∑ i (4)

[0181] In formula (4), h is a global bandwidth parameter, ∑ i is a local covariance matrix centered at the sample point x i ;

[0182] The anomaly score is defined as the normalized negative logarithmic probability density combined with the historical risk trend:

[0183]

[0184] In formula (5), α and β are weight coefficients, R t , R t-1 are the current and previous risk indexes, σ R is the standard deviation of the historical risk index.

[0185] To verify the calculation process and effectiveness of the anomaly score in step five, a targeted experiment is designed to visually present the modeling process of the local adaptive bandwidth Gaussian kernel density estimation and the calculation logic of the anomaly score through real physiological signal feature data. The experiment is as follows:

[0186] Experiment preparation:

[0187] Sample and feature determination: 80 healthy adults (without cardiovascular disease history) were recruited to collect multi-modal physiological signals (ECG, skin electricity, respiration, blood volume pulse), and 4 core features (ECG RR interval stability, skin electricity signal fluctuation, respiration frequency variation, blood volume pulse peak intensity) were extracted after preprocessing to form 1200 normal historical samples (15 valid feature data per person); 8 patients with arrhythmia (sudden death high-risk population) were also recruited to collect the same type of feature data, forming 80 target samples to be evaluated.

[0188] Parameter setting: the global bandwidth parameter is calibrated to 0.9 according to clinical data; the local covariance matrix is calculated by the feature distribution of 25 adjacent samples around each normal sample; the weight coefficient is determined to be α = 0.6, β = 0.4 (balance the influence of probability density and historical risk trend) through cross-validation; the historical risk index is obtained from the dynamic evaluation of normal samples, and the standard deviation σ_R is statistically 0.15.

[0189] Experimental steps:

[0190] Normal sample feature distribution modeling: for each normal historical sample, first calculate the local covariance matrix centered on it, then combine the global bandwidth parameter, and get the local adaptive bandwidth matrix corresponding to each normal sample according to the relationship "local adaptive bandwidth matrix = global bandwidth parameter × local covariance matrix". Subsequently, based on the normalization rule of Gaussian kernel function, combined with the feature difference between the target sample to be evaluated and the normal sample, all normal samples are traversed, and the probability density estimation value of the target sample in the normal feature distribution is accumulated and calculated.

[0191] Abnormal score calculation: first, take the logarithm of the probability density estimation value obtained above, take the negative, and multiply by the weight coefficient α; second, obtain the risk index of the target sample at the current time and the risk index at the previous time, calculate the absolute value of the difference between the two, and divide by the standard deviation σ_R of the historical risk index; third, add the results of the first two steps to get the final abnormal score. R

[0192] Control verification: set up a traditional fixed bandwidth 0.9 modeling group, calculate the probability density and abnormal score with the same samples, and compare with this method.

[0193] Experimental data and results:

[0194] Probability density estimation results: select 5 typical normal samples and 5 high-risk target samples to compare the probability density estimation values of the two methods:

[0195] Table 5: Comparison of probability density estimation values

[0196] Sample probability Probability density estimate value Fixed bandwidth method probability density estimate value Normal sample 1 2.15 1.42 Normal sample 2 1.98 1.35 Normal sample 3 2.07 1.40 High-risk sample 1 0.32 0.68 High-risk sample 2 0.28 0.71 ​

[0197] It can be seen that the probability density estimation of the method is closer to the true distribution for normal samples, and the differentiation of high-risk samples is more obvious.

[0198] Abnormal score results:

[0199] Table 6: Abnormal score of two types of samples

[0200]

[0201]

[0202] In the method, the abnormal score mean of the high-risk sample is 3.92 times that of the normal sample, and the overlapping area of the scores of the two types of samples accounts for only 4.8%; in the fixed bandwidth method, the mean of the high-risk sample is only 2.08 times that of the normal sample, and the overlapping area accounts for 19.3%, proving that the abnormal score of the method can more accurately distinguish between normal and high-risk states.

[0203] Experimental conclusion:

[0204] Through the verification of real sample data, the abnormal score calculation process in step five can effectively model the normal feature distribution and quantify the sample abnormality: first, the local adaptive bandwidth Gaussian kernel density estimation is used to accurately describe the feature distribution law of normal samples, and then combined with the standardized change of historical risk trend, the final abnormal score can clearly distinguish between normal and high-risk samples, providing reliable quantitative basis for sudden death risk assessment.

[0205] Although the specific embodiments of the present application are described above, those skilled in the art should understand that these specific embodiments are only illustrative, and those skilled in the art can make various omissions, substitutions and changes to the details of the above method and system without departing from the principles and essence of the present application; for example, combining the above method steps, performing substantially the same function in substantially the same manner to achieve substantially the same result according to the same method belongs to the scope of the present application; therefore, the scope of the present application is only limited by the appended claims.

Claims

1. A sudden death risk real-time evaluation system based on multi-modal physiological signal fusion, applied to the field of human physiological state monitoring and early warning, characterized in that: The system comprises a signal acquisition module (100), a signal preprocessing module (200), a high-dimensional feature extraction module (300), a multi-modal feature fusion module (400), a sudden death risk quantification module (500), a model updating engine (600) and a early warning feedback module (700); The signal acquisition module (100) adopts a multi-channel synchronous acquisition technology architecture to synchronously acquire electrocardio signals, skin electricity signals, respiratory signals and blood volume pulse signals, and maintain the time synchronization and data integrity of the multi-modal physiological signals; the signal acquisition module (100) comprises a bioelectric signal electrode array (110), an optical sensing unit (120) and an impedance detection circuit (130), and completes the synchronous acquisition of multi-source signals through a special interface protocol; The signal preprocessing module (200) integrates an adaptive filter (210) and a quality evaluation timing unit (220), realizes adaptive noise reduction of signals through an adaptive signal decomposition algorithm, and completes the timing alignment of multi-modal signals through a dynamic time warping technology; the signal preprocessing module (200) comprises a signal input interface (230), a variational modal decomposer (240), a dynamic threshold processor (250), a timing calibrator (260), a signal quality evaluator (270) and a preprocessing output interface (280), and constitutes a complete signal preprocessing pipeline; The high-dimensional feature extraction module (300) fuses a timing convolution network (310) and a gated recurrent unit (320), extracts the time domain features, frequency domain features and nonlinear features of signals in parallel, and generates a high-dimensional feature vector space; the high-dimensional feature extraction module (300) comprises a feature extraction controller (330), a timing convolution processor (340), a recurrent neural network unit (350) and a feature vector synthesizer (360); The multi-modal feature fusion module (400) adopts an improved multi-head self-attention mechanism (410), takes electrocardio signal features as a reference, realizes intelligent weighted fusion of multi-modal features through calculation of inter-modal mutual information entropy; the multi-modal feature fusion module (400) comprises an attention weight calculation unit (420), a feature projection transformer (430), a mutual information entropy analyzer (440) and a fused feature synthesizer (450), and constitutes a multi-level feature fusion architecture; The sudden death risk quantification module (500) integrates a timing anomaly detector (510) and a risk regressor (520), calculates an abnormal score based on a kernel density estimation algorithm, and generates a dynamic risk index in combination with a historical risk state; The sudden death risk quantification module (500) comprises an anomaly detection controller (530), a kernel density estimation unit (540), a risk state storage (550) and a regression predictor (560); The model updating engine (600) adopts an online sequence learning strategy, realizes the incremental updating and optimization of model parameters through a regularized least square algorithm; the model updating engine (600) comprises an incremental learning controller (610), a weight updater (620), a parameter optimizer (630) and a model verification unit (640). The early warning feedback module (700) triggers a multi-level early warning protocol according to the risk index classification, and completes real-time feedback and display of the evaluation result. The early warning feedback module (700) comprises a risk level determinator (710), an early warning protocol executor (720), a feedback information generator (730), and a man-machine interaction interface (740).

2. The real-time sudden death risk assessment system based on multi-modal physiological signal fusion according to claim 1, characterized in that, The signal preprocessing module (200) comprises a signal input interface (230), a variational modal decomposer (240), a dynamic threshold processor (250), a timing calibrator (260), a signal quality evaluator (270), and a preprocessing output interface (280). The variational modal decomposer (240) adopts an adaptive bandwidth optimization algorithm architecture, comprising a wave generator (241), a modal decomposer (242), and a reconstruction filter (243). The wave generator (241) calculates signal envelope features through Hilbert transform. The modal decomposer (242) solves variational constraint problems by using an alternating direction multiplier method. The reconstruction filter (243) reconstructs high-quality signals after denoising through inverse transform. The timing calibrator (260) is realized by using a dynamic time warping algorithm, comprising a distance matrix calculation unit (261), a warping path searcher (262), and a time scaler (263). The distance matrix calculation unit (261) measures signal similarity features by using Euclidean distance. The warping path searcher (262) finds the optimal alignment path by using a dynamic programming algorithm. The time scaler (263) realizes signal resampling by using a cubic spline interpolation algorithm. The signal quality evaluator (270) integrates a multi-dimensional quality index evaluation system, comprising a signal-to-noise ratio calculation unit (271), a waveform distortion detection unit (272), and a motion artifact identification unit (273). The signal-to-noise ratio calculation unit (271) calculates signal quality indicators by using a power spectral ratio method. The waveform distortion detection unit (272) calculates waveform correlation coefficients by using a template matching algorithm. The motion artifact identification unit (273) performs artifact compensation processing based on acceleration signal features. 3.The real-time sudden death risk assessment system based on multi-modal physiological signal fusion according to claim 1, wherein, The high-dimensional feature extraction module (300) comprises a feature extraction controller (330), a timing convolution processor (340), a recurrent neural network unit (350), and a feature vector synthesizer (360). The timing convolution processor (340) adopts an expanded causal convolution architecture with an exponential expansion factor, comprising a convolution kernel optimization unit (341), a feature mapping unit (342), and a pooling layer (343). The convolution kernel optimization unit (341) optimizes convolution kernel parameters by using an adaptive gradient descent algorithm. The feature mapping unit (342) realizes feature transformation by using a nonlinear activation function. The pooling layer (343) adopts a multi-scale pooling strategy combining maximum pooling and average pooling. The recurrent neural network unit (350) adopts a gated attention recurrent unit structure, and comprises a time attention mechanism (351), a gated update unit (352), and a state memory unit (353); the time attention mechanism (351) calculates attention weights through historical state similarity; the gated update unit (352) integrates the calculation processes of update gates and reset gates; and the state memory unit (353) adopts a long short-term memory structure to store time sequence feature information.

4. The real-time sudden death risk assessment system based on multi-modal physiological signal fusion according to claim 1, characterized in that, The multi-modal feature fusion module (400) comprises an attention weight calculation unit (420), a feature projection transformer (430), an mutual information entropy analyzer (440), and a fusion feature synthesizer (450); The mutual information entropy analyzer (440) adopts a K-nearest neighbor-based non-parametric mutual information estimation structure, and comprises a probability density estimation unit (441), an entropy value calculation unit (442), and a bias correction unit (443); the probability density estimation unit (441) calculates a joint probability distribution by using a kernel density estimation method; the entropy value calculation unit (442) calculates a mutual information entropy value by using a log likelihood ratio; and the bias correction unit (443) corrects an estimation bias by using a Jackknife resampling technique; The fusion feature synthesizer (450) adopts a weighted feature fusion architecture, and comprises a weight distribution unit (451), a feature weighting unit (452), and a fusion output unit (453); the weight distribution unit (451) distributes weights of each modal feature according to the mutual information entropy value; the feature weighting unit (452) realizes feature weighting by using point multiplication operation; and the fusion output unit (453) generates a final fusion feature tensor by using a splicing operation.

5. The real-time sudden death risk assessment system based on multi-modal physiological signal fusion according to claim 1, characterized in that, The sudden death risk quantification module (500) comprises an anomaly detection controller (530), a kernel density estimation unit (540), a risk state memory (550), and a regression predictor (560); The anomaly detection controller (530) adopts a multi-index anomaly detection architecture, and comprises a threshold self-adaptive unit (531), an anomaly scorer (532), and an alarm triggering unit (533); the threshold self-adaptive unit (531) dynamically adjusts an anomaly threshold according to historical data distribution; the anomaly scorer (532) calculates an anomaly score by using a density estimation method; and the alarm triggering unit (533) triggers a hierarchical early warning mechanism according to the score result; The risk state memory (550) adopts a space-time state storage architecture, and comprises a state encoder (551), a memory matrix (552), and a state retriever (553); the state encoder (551) encodes historical state information by using a recurrent neural network; the memory matrix (552) stores key state features by using an attention mechanism; and the state retriever (553) retrieves relevant historical states by similarity matching. The regression predictor (560) adopts a state space model and a deep learning hybrid architecture, and comprises a state transition unit (561), an observation model unit (562), and a parameter optimization unit (563); the state transition unit (561) models a state evolution process by using a linear dynamic system; The observation model unit (562) establishes a mapping relationship between a state and an observation by using a deep learning network; and the parameter optimization unit (563) optimizes model parameters by using an expectation maximization algorithm.

6. The real-time sudden death risk assessment system based on multi-modal physiological signal fusion according to claim 1, characterized in that, The model updating engine (600) comprises an incremental learning controller (610), a weight updater (620), a parameter optimizer (630), and a model verification unit (640); The incremental learning controller (610) adopts an online sequence extreme learning machine architecture, and comprises a hidden node management unit (611), an output weight calculation unit (612), and a model updating unit (613); the hidden node management unit (611) dynamically adds or deletes a hidden layer node and optimizes a node parameter; the output weight calculation unit (612) calculates an output layer weight by using a recursive least squares method; and the model updating unit (613) realizes incremental updating of a model by using an online learning algorithm; The weight updater (620) adopts a recursive least squares architecture with a forgetting factor, and comprises a forgetting factor adjustment unit (621), a covariance matrix updating unit (622), and a weight correction unit (623); the forgetting factor adjustment unit (621) dynamically adjusts a forgetting factor size according to a prediction error; the covariance matrix updating unit (622) updates a covariance matrix by using a rank-one update algorithm; and the weight correction unit (623) corrects a weight parameter by using a gradient descent method.

7. A sudden death risk real-time evaluation method based on multi-modal physiological signal fusion, applied to the field of human physiological state monitoring and early warning, characterized in that, The method comprises the following steps: Step one, multi-modal physiological signal synchronous acquisition and quality evaluation: Through a multi-channel synchronous acquisition architecture composed of the bioelectric signal electrode array (110), the optical sensing unit (120), and the impedance detection circuit (130), the electrocardiosignal, the skin electricity signal, the respiration signal, and the blood volume pulse signal of the user are synchronously acquired; a power spectrum ratio method and an acceleration signal feature analysis method are used to calculate the signal-to-noise ratio of each signal in real time and identify motion artifacts, and the signal quality is evaluated in multiple dimensions, and only the signal segments with qualified quality are reserved for subsequent processing; Step two, adaptive noise reduction and multi-modal time sequence alignment: An adaptive signal decomposition algorithm is used to perform adaptive noise reduction processing on various physiological signals, a signal envelope feature is calculated by using Hilbert transform, and a variation constraint problem is solved by using an alternating direction multiplier method, and a high-quality signal is reconstructed; a dynamic time warping algorithm is used to calculate an Euclidean distance matrix of the remaining signals and a reference signal with the electrocardiosignal as the time sequence reference, an optimal alignment path is searched by using dynamic programming, and accurate resampling and time sequence alignment of the multi-modal signals are completed by using cubic spline interpolation; Step three, parallel feature extraction in high-dimensional time-frequency domain and nonlinearity: Local time-domain and frequency-domain features of the signal are extracted using the time convolution network (310) which adopts a dilated causal convolution structure with an exponential dilation factor and combines a multi-scale pooling strategy of maximum pooling and average pooling; long-term dependencies and nonlinear features of the signal are extracted in parallel using the gated attention recurrent unit (320) structure, attention weights are calculated through historical state similarity, and state information flow is controlled using update gates and reset gates; and finally the feature vector synthesizer (360) synthesizes the above features into a unified high-dimensional feature vector; Step four, multi-modal feature fusion based on mutual information weighting: Based on the ECG signal features as the benchmark reference, the correlation between different modal features is calculated through the improved multi-head self-attention mechanism; the mutual information entropy between modalities is estimated using a non-parametric method based on K-nearest neighbors, the joint probability distribution is calculated using kernel density estimation, and the estimation bias is corrected using the Jackknife resampling technique; based on the mutual information entropy value, appropriate weights are assigned to each modal feature, feature weighting is achieved using point multiplication, and the final multi-modal fusion feature tensor is generated through splicing operation; Step five, calculation and prediction of dynamic risk index: The abnormal score of the current fusion feature vector is calculated based on the kernel density estimation algorithm, and the adaptive bandwidth optimization technique is used to improve the accuracy of probability density estimation; the historical state information stored in the risk state memory is combined to model the temporal evolution process of risk through a hybrid architecture of state space model and deep learning, a linear dynamic system is used to describe state transition, and the model parameters are optimized using the expectation maximization algorithm, and the dynamic risk index is output; Step six, multi-level early warning triggering and real-time feedback: According to the size of the dynamic risk index, the corresponding warning level is triggered according to the preset multi-level warning protocol; real-time feedback information is generated and visualized through the man-machine interaction interface, including risk index curve, warning level prompt and health suggestion, forming a complete monitoring-evaluation-warning closed loop.

8. The real-time sudden death risk assessment method according to claim 7, characterized in that, In step four, the inter-modal mutual information entropy I(X;Y) is calculated as follows: Discretize the continuous feature vectors X and Y based on K-nearest neighbors, and divide the feature space into multiple local regions; A non-parametric kernel density estimation method based on K-nearest neighbors is used to compute the joint probability distribution with the marginal probability distributions The estimation formula is In formula (1), n represents the total number of samples for estimation, K(·) is a Gaussian kernel function, h x and h y are adaptive bandwidths of features X and Y respectively, x represents the target estimation point of feature X, x i represents the actual observed value of feature X in the i-th sample, ||x-x i || represents the Euclidean distance between the target estimation point of feature X and the observed value of the i-th sample, represents the joint bandwidth normalization coefficient; The mutual information entropy is finally calculated by the following bias correction formula: In Equation (2), is a double summation operator, denotes an estimated joint probability distribution value when the feature X takes x i , the feature Y takes y i , and ∈ is a small positive value. denotes an estimated marginal probability distribution value when the feature X takes x i , and ∈ is a small positive value. denotes an estimated marginal probability distribution value when the feature Y takes y i , and ∈ is a small positive value. corr denotes a bias correction term for mutual information entropy; and λ denotes a bias correction coefficient. The calculated mutual information entropy values are used for weight distribution in subsequent feature weighting fusion, and the weight calculation satisfies:

9. The real-time sudden death risk assessment method according to claim 7, characterized in that, In step five, the abnormal score S(x) is calculated as follows: The feature distribution of normal historical samples is modeled using a Gaussian kernel density estimation method based on local adaptive bandwidth, and the probability density estimation function is: In formula (3), is the probability density estimation value representing the normal historical sample feature, n is the total number of normal historical samples, is the summation operator, d represents the dimension of the feature vector, H i represents the local adaptive bandwidth matrix corresponding to the i-th normal sample, |H i is the determinant of the bandwidth matrix H i , represents the inverse matrix of the bandwidth matrix H i , x is the target feature vector to be evaluated, x i is the feature vector of the i-th normal historical sample, (x-x i ) T is the transpose of the difference vector (x-x i ), exp(·) is the exponential function, is the normalization coefficient of the Gaussian kernel function; The calculation formula of the local adaptive bandwidth matrix HiHi is: H i = h · ∑ i (4) In equation (4), h is a global bandwidth parameter, ∑ i is the local covariance matrix centered at the sample point x i ; The abnormal score is defined as the standardized negative log probability density combined with the historical risk trend: In equation (5), α and β are weight coefficients, R t and R t-1 are the risk indices at the current and previous time, respectively, and σ R is the standard deviation of the historical risk indices.

Images