Old people brain health monitoring and cerebral infarction early warning method and system

Abstract

The invention discloses a brain health monitoring and cerebral infarction early warning method and system for old people. The method comprises the steps that electroencephalogram, electrocardio, blood oxygen and cerebral blood flow velocity multi-mode physiological signal data are collected; adopting a characteristic mode decomposition method to decompose the multi-mode physiological signal data; the method comprises the following steps: introducing a KAN network into an MATE structure, carrying out nonlinear mapping and feature reconstruction, and constructing a KAMATE model; an improved Bayesian optimization algorithm is obtained based on the Bayesian optimization algorithm in combination with a dynamic balance strategy; the improved Bayesian optimization algorithm is combined with the characteristic mode decomposition method to achieve characteristic decomposition of the multi-mode physiological signals, and parameters of the KAMATE model are optimized; dividing the brain health state into four grades according to a prediction result output by the KAMATE model; according to the invention, real-time early warning of abnormal cerebral blood supply, cognitive decline and cerebral infarction high-risk states of the elderly can be realized, intelligent support is provided for health management of the elderly, and the method has important application value.

Description

Technical Field

[0001] The present invention relates to an early warning method and system, and in particular to a method and system for monitoring brain health and early warning of cerebral infarction in the elderly. Background Art

[0002] As the global population ages, brain health issues among the elderly have become a major public health challenge. Traditional brain health monitoring methods often rely on single physiological signals, such as brain waves, electrocardiograms, or blood oxygen saturation. While these methods can help monitor some health conditions to a certain extent, they often struggle to comprehensively and accurately assess complex brain health status, especially in providing early warning of sudden illnesses such as cerebral infarction.

[0003] Against this backdrop, multimodal physiological signal fusion technology has gradually become a research hotspot. By integrating multiple physiological signals, such as brain waves, electrocardiograms, blood oxygen saturation, photoplethysmography, galvanic skin response, near-infrared spectroscopy, as well as body temperature and inertial sensor signals, the human body's physiological state can be reflected from multiple dimensions, thereby more comprehensively assessing health risks. However, multimodal signal fusion technology still faces many challenges, such as synchronization issues in signal acquisition, the complexity of feature extraction, and the high requirements for model building. Existing technologies still suffer from low computational efficiency and slow convergence when processing large-scale multimodal data, making it difficult to meet the needs of real-time monitoring and early warning.

[0004] At the same time, Bayesian optimization can dynamically adjust parameters through probabilistic models, improving the model's adaptability and prediction accuracy. However, traditional Bayesian optimization algorithms still suffer from low computational efficiency and slow convergence when processing large-scale multimodal data, and cannot meet the needs of real-time monitoring and early warning. In addition, although deep learning technologies such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) perform well in processing time series data and image data, they still face problems with insufficient feature extraction and poor model interpretability in multimodal signal fusion, limiting their widespread application in the field of brain health monitoring. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a method for brain health monitoring and cerebral infarction warning in the elderly to conduct comprehensive and accurate brain health assessment and individualized monitoring. On the other hand, it provides a brain health monitoring and cerebral infarction warning system for the elderly.

[0006] Technical solution: The early warning method of the present invention comprises the following steps:

[0007] S1. Collect multimodal physiological signal data including EEG, ECG, blood oxygen and cerebral blood flow velocity;

[0008] S2. Decomposing the multimodal physiological signal data using an eigenmode decomposition method, adjusting decomposition parameters, and extracting key features of instantaneous frequency, energy entropy, and kurtosis;

[0009] S3, introduce the KAN network into the MATE structure, perform nonlinear mapping and feature reconstruction, and build the KAMATE model;

[0010] S4. Based on the traditional Bayesian optimization algorithm and combined with the dynamic balance strategy, an improved Bayesian optimization algorithm is obtained;

[0011] S5. Using the improved Bayesian optimization algorithm in combination with the eigenmode decomposition method to achieve eigendecomposition of the multimodal physiological signal and optimize the KAMATE model parameters;

[0012] S6. Construct a cerebral infarction warning mechanism, and divide the brain health status into a healthy level, a mild warning level, a high-risk level, and an acute cerebral infarction level according to the prediction results output by the KAMATE model.

[0013] Preferably, the multimodal physiological signal data collected by S1 can be expressed as:

[0014] X=[X t-T ,X t-T+1 ,…,X t ];

[0015] Among them, X t represents all physiological signal features collected at time step t, and T represents the length of the history window.

[0016] Preferably, the characteristic mode decomposition method in S2 decomposes the input signal into an eigenmode function, and the decomposition expression is as follows:

[0017]

[0018] Among them, the IMF j represents the jth intrinsic mode function decomposed, R represents the residual term, and K represents the number of decomposed modes.

[0019] Preferably, the S3 includes:

[0020] S31. Extract the common structural features between the modalities through the latent representation encoder and output the shared latent variable Z c ;

[0021] S32, through the modality-specific latent representation encoder, each modality is independently modeled to obtain exclusive feature representations and output a specific latent variable set [Z S1 ,Z S2 ,…,Z SM ], where M represents the total number of modalities and S represents specificity;

[0022] S33, inputting the shared latent variables and the specific latent variables into the nonlinear feature reconstruction layer designed based on the KAN structure respectively, and reconstructing the latent vector Z using a combination of multiple sets of nonlinear functions;

[0023] S34, mapping is performed through a nonlinear transformation function to obtain a mapped output Y;

[0024] S44. The mapped shared latent variables and specific latent variables are fused in the time series dimension to form a time series feature tensor, and the multimodal physiological signal reconstruction and trend prediction values ​​are output.

[0025] Preferably, the output Y formula in S34 is as follows:

[0026]

[0027] Among them, z j represents the jth component of the latent variable generated by MATE, w ij and b i Represent the weight and bias of the KAN network, φ i (·) represents a learnable nonlinear activation function, N represents the KAN output dimension, and d represents the input latent variable dimension.

[0028] Preferably, the S4 includes:

[0029] S41. Use Gaussian process regression as a proxy model to approximate the objective function and model the objective function;

[0030] S42. Performing posterior inference on the Gaussian process based on the historical evaluation data set to obtain a predicted mean value of the objective function in the entire hyperparameter space;

[0031] S43, using the expected improvement as an acquisition function, and using a dynamic balance strategy to regulate the search weight of the acquisition function, so that the dynamic weight changes adaptively with the number of iterations;

[0032] S44. Select the next evaluation point by maximizing the acquisition function. If the maximum number of iterations is reached or the objective function converges, terminate the optimization process; otherwise, enter the next round of sampling iteration.

[0033] Preferably, the objective function in S41 is modeled as:

[0034] f(θ)~GP(m(θ),k(θ,θ′));

[0035] Among them, m(θ) represents the mean function, which is usually set to 0, and k(θ,θ') represents the kernel function;

[0036] The dynamic weight formula in S43 is:

[0037]

[0038] Among them, w t represents the exploration weight of the t-th iteration, α represents the smoothing factor (0<α<1), β represents the adjustment parameter, t represents the current iteration number, and T represents the total number of iterations;

[0039] The acquisition function is:

[0040] Acq(θ)=w t ·μ(θ)+(1-w t )·EI(θ);

[0041] EIf(θ)=E[max(f(θ)-f(θ + ),0)];

[0042] Where EI(θ) represents the acquisition function, f(θ + ) represents the optimal objective function value found so far.

[0043] Preferably, the S5 includes:

[0044] S51. An improved Bayesian optimization algorithm with a dynamic weight adjustment mechanism is introduced to optimize the parameters in the eigenmode decomposition module and KAMATE modeling.

[0045] S52, the optimized eigenmode decomposition module provides time-frequency features, which are input into the KAMATE model;

[0046] S53. The optimized KAMATE model monitors the dynamic change trends of EEG and ECG signals in real time.

[0047] Preferably, the health level indicates that the physiological signal is stable and no abnormalities are found; the mild warning level indicates that some key indicators have slight abnormal fluctuations; the high-risk level indicates that the monitoring signal is obviously abnormal and there is a possibility of cerebral infarction in the short term; the acute cerebral infarction level indicates that the physiological signal characteristics mutate or are severely disordered, and an alarm signal must be issued immediately and the medical intervention module must be linked to achieve real-time response.

[0048] The early warning system of the present invention comprises:

[0049] Data acquisition module, used to collect multimodal physiological signal data such as EEG, ECG, blood oxygen and cerebral blood flow velocity;

[0050] Algorithm optimization module, which is used to obtain an improved Bayesian optimization algorithm by combining the traditional Bayesian optimization algorithm with a dynamic balance strategy to optimize the parameter search process;

[0051] The nonlinear reconstruction module is designed based on the MATE and KAN structures to build the KAMATE model and enhance the collaborative expression and time series modeling capabilities between multimodal signals.

[0052] The eigenmode decomposition module is used to achieve eigendecomposition of multimodal physiological signals and optimize KAMATE model parameters through an improved Bayesian optimization algorithm;

[0053] The cerebral infarction warning module is used to divide brain health status into four levels: healthy, mild warning, high risk and acute cerebral infarction based on the multimodal feature prediction results, to achieve real-time warning and intervention.

[0054] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: 1. By integrating multimodal physiological signals of EEG, ECG, blood oxygen and cerebral blood flow velocity, combined with personalized health records, a long-term health trend model is constructed to achieve comprehensive and accurate brain health assessment and individualized monitoring, effectively improving the ability to capture latent features such as abnormal cerebral blood supply and cognitive decline, thereby enhancing the recognition and prediction stability of the system under complex physiological conditions; 2. Through the characteristic mode decomposition method, the stability and accuracy of the intrinsic mode function IMF in the signal decomposition can be effectively improved, and the instantaneous frequency, energy entropy, kurtosis and other key features, significantly improving the feature quality and system response speed; 3. The KAMATE structure can effectively extract the shared features and specific modal features of multimodal signals through the variational inference mechanism, enhance the decoupling ability between signals and the temporal dependency modeling ability, thereby improving the discrimination ability and recognition accuracy of early warning of cerebral infarction; 4. Based on large-scale multi-source data training and construction of personalized health trend models, it has the ability to dynamically learn and adapt to individual differences, can be widely applied to different elderly groups, provide differentiated predictions and personalized interventions, and improve the intelligence and precision of brain health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the process of the present invention;

[0056] Figure 2 Schematic diagram of characteristic mode decomposition of the present invention;

[0057] Figure 3 Schematic diagram of the KAMATE model structure of the present invention;

[0058] Figure 4 Schematic diagram of the improved Bayesian optimization algorithm of the present invention. DETAILED DESCRIPTION

[0059] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0060] S1. Collect and store historical data of multimodal physiological signals including EEG, ECG, blood oxygenation (SpO2), and cerebral blood flow velocity. The collected multimodal physiological signal historical data can be expressed as a time series matrix:

[0061] X=[X t-T ,X t-T+1 ,…,X t ];

[0062] Among them, X t represents all physiological signal features collected at time step t, and T represents the length of the history window.

[0063] S2. In the feature decomposition stage, the feature mode decomposition (FMD) method is used to decompose the multimodal physiological signals to extract stable and predictive intrinsic mode functions (IMFs), thereby improving the reliability and prediction performance of signal feature extraction, adjusting the decomposition parameters and extracting key features such as instantaneous frequency, energy entropy and kurtosis.

[0064] The FMD method is used to decompose the input signal into intrinsic mode functions (IMFs). The decomposition expression is as follows:

[0065]

[0066] Among them, the IMF j represents the jth intrinsic mode function decomposed, aiming to maximize the second-order cyclostationarity index (ICS2), thereby improving the identifiability of fault features. R represents the residual term, and K represents the number of decomposed modes. The FMD method achieves signal decomposition by iteratively updating the filter coefficients. The update process of the filter coefficients involves an optimization problem, the goal of which is to maximize the second-order cyclostationarity index (ICS2):

[0067]

[0068] Restrictions:

[0069]

[0070] Among them, h q Represents the coefficient of the qth filter, IMF q represents the qth intrinsic mode function, R XWX and R XX It indicates that the correlation matrix W is the intermediate variable matrix that controls the weighted correlation matrix.

[0071] In the eigendecomposition stage, the eigenmode decomposition (FMD) method is used to process multimodal physiological signals. By adjusting key hyperparameters such as the number of modes K and the filter length L, the stability and accuracy of the decomposition results are optimized. In addition, the improved Bayesian optimization (IBO) strategy is introduced to perform adaptive parameter optimization on the FMD decomposition process to further improve the feature extraction performance. The optimized modal components are obtained through FMD decomposition, and the optimal hyperparameter θ is finally obtained. FMD , used for FMD decomposition to obtain the optimized modal components:

[0072]

[0073] S3. Based on the MATE structure, the KAN network is introduced for nonlinear mapping and feature reconstruction to construct the KAMATE model.

[0074] During the model building phase, the system first receives multimodal physiological signal inputs, including but not limited to electroencephalogram (EEG), electrocardiogram (ECG), blood oxygenation (SpO2), and cerebral blood flow velocity multimodal physiological signals.

[0075] In order to fully explore the potential time-frequency characteristic structure in each modal signal, the characteristic mode decomposition (FMD) processing is first applied to each type of original signal to decompose it into several physically meaningful intrinsic mode functions (IMFs) to extract its time-frequency joint features, providing a structurally rich feature foundation for subsequent modeling.

[0076] For the input multi-source modal data, the system designs two types of encoding modules:

[0077] One is the potential representation encoding module used to extract the common structural features between all modalities, and the encoding result is Z c .

[0078] The second is the modality-specific latent representation encoder, which models each modality independently, obtains its own unique feature representation, and outputs a set of specific latent variables [Z S1 ,Z S2 ,…,Z SM ], where M represents the total number of modalities. The shared latent variables and the modality-specific latent variables are input into the nonlinear feature reconstruction layer, which is based on the Kolmogorov-Arnold Network (KAN) structure and has a strong nonlinear mapping capability. The KAN module reconstructs the latent vector Z using a combination of multiple nonlinear functions, and the output is:

[0079]

[0080] Among them, z j represents the jth component of the latent variable generated by MATE, wij and b i Represent the weight and bias of the KAN network, φ i (·) represents a learnable nonlinear activation function, N represents the KAN output dimension, and d represents the input latent variable dimension. This formula takes the weighted sum of each modality's latent variable and the network weights, then maps them through a nonlinear transformation function. The resulting mapped output, Y, is used for subsequent feature reconstruction and trend prediction.

[0081] The shared and specific potential representations after KAN mapping are fused in the time series dimension to form the time series feature tensor Z fused (t), which fuses global modal correlation and local modal dynamics while preserving the temporal order characteristics of the original sequence, serving as a unified input for subsequent prediction and reconstruction tasks.

[0082] In the output stage, the system designs two decoding branches:

[0083] The first branch is the latent representation decoder, which mainly realizes the reconstruction of the original modal signal and optimizes the following reconstruction objective function through variational inference:

[0084] L recon =E q(z|x) [log p(x|z)]-KL[q(z|x)||p(z)];

[0085] Among them, the first term is the reconstruction error, and the second term is the KL divergence, which is used to constrain the distribution consistency of the latent space.

[0086] The second branch is the trend prediction subnetwork, which performs time series trend prediction on key physiological indicators based on the KAN structure, and the output is:

[0087]

[0088] Prediction results and true value S t+1 ,…,S t+T The error between them is optimized using mean square error (MSE):

[0089]

[0090] in, Indicates the actual value, represents the predicted value, and N is the number of samples.

[0091] Finally, the entire model is trained end-to-end through a joint loss function, taking into account reconstruction quality, modality consistency, and prediction accuracy. The joint optimization objective function is defined as:

[0092] L total =L recon+λ1L share +λ2L pred ;

[0093] Among them, L share represents the modal shared consistency regular term, and its expression is:

[0094]

[0095] The above formula constrains the structural consistency between the shared latent variables and the modality-specific latent variables, enhancing the collaborative learning ability between multiple modalities. The coefficients λ1 and λ2 represent hyperparameters that adjust the weights of different loss terms.

[0096] Ultimately, by minimizing this joint loss function, the model achieves high-precision multimodal physiological signal reconstruction and trend prediction, enhancing the system's ability to identify brain health status and provide early warning of cerebral infarction. Furthermore, to achieve optimal configuration in the parameter space, the structural parameters and training hyperparameters of the KAMATE model are automatically tuned using the IBO algorithm, further improving system performance and predictive robustness.

[0097] S4. By combining the traditional Bayesian optimization algorithm with a dynamic balancing strategy, we developed an improved Bayesian optimization (IBO) algorithm to optimize the parameter search process. This algorithm optimizes the multimodal physiological signal feature decomposition (FMD) and KAMATE model parameters, thereby improving overall prediction performance and system robustness.

[0098] Gaussian Process Regression (GPR) is used as a proxy model to approximate the objective function and model the objective function. Let the optimization variable be θ, then the objective function is modeled as:

[0099]

[0100] Among them, m(θ) represents the mean function, which is usually set to 0; k(θ,θ') represents the kernel function, which is used to describe the similarity between input θ and θ'.

[0101] Based on historical evaluation datasets:

[0102]

[0103] Perform posterior inference on the Gaussian process to obtain the predicted mean μ(θ) of the objective function in the entire hyperparameter space.

[0104] In order to achieve adaptive adjustment between global exploration and local exploitation, a dynamic balance strategy is adopted to regulate the search weight of the acquisition function.

[0105] Let the dynamic weight w t Adaptively changes with the number of iterations t, the formula is:

[0106]

[0107] Among them, w t represents the exploration weight of the t-th iteration, α represents the smoothing factor (0<α<1), which controls the update speed of the weight, β represents the adjustment parameter, which determines the rate of exploration and exploitation conversion; t represents the current iteration number, and T represents the total number of iterations.

[0108] Expected Improvement (EI) is used as the acquisition function to guide the selection of the next sampling point:

[0109] EI(θ)=E[max(f(θ)-f(θ) + ),0)];

[0110] Among them, f(θ + ) represents the optimal objective function value found so far; after combining the dynamic weights, the acquisition function can be further expanded to:

[0111] Acq(θ)=w t ·μ(θ)+(1-w t )·EI(θ);

[0112] Among them, w t is the exploration weight of the tth iteration, which is a parameter between 0 and 1 and is used to control the balance between exploration and exploitation. μ(θ) is the mean prediction of f(θ), EI(θ) represents the acquisition function, and f(θ + ) represents the optimal objective function value found so far.

[0113] The next evaluation point is selected by maximizing the acquisition function:

[0114]

[0115] In θ t+1 The objective function is evaluated and the surrogate model and observation set are updated.

[0116] If the maximum number of iterations is reached or the objective function converges, the optimization process is terminated; otherwise, the next round of sampling iteration is entered.

[0117] During the system optimization phase, an improved Bayesian optimization algorithm with a dynamic weight adjustment mechanism is introduced to jointly adaptively optimize key parameters in the eigenmode decomposition (FMD) module and the KAMATE modeling structure, thereby improving the overall system's modeling capabilities and trend prediction performance for multimodal physiological signals. Specifically, IBO optimizes the parameter set θ = {θ1, θ2} of the FMD and KAMATE models as a whole, where θ1 represents key parameters of the FMD module, such as the number of modes K, penalty factor α, bandwidth threshold τ, and other parameters that control signal decomposition quality; θ2 represents key hyperparameters in the KAMATE structure, such as the number of layers and hidden dimensions of the latent representation network, the number of nodes and activation function form of the KAN mapping layer, and the learning rate, connection method, and regularization strategy of the prediction subnetwork.

[0118] During the optimization process, the IBO algorithm utilizes a dynamic balancing weight mechanism to adaptively adjust the search strategy between global exploration and local exploitation, jointly optimizing the optimal parameter configuration of the FMD and KAMATE architectures. The optimized FMD module ultimately provides more discriminative time-frequency features, serving as high-quality input for the KAMATE model. The optimized KAMATE model not only enhances its adaptability to complex physiological signals but also enables real-time monitoring of dynamic trends in electroencephalogram (EEG) and electrocardiogram (ECG) signals, improving the stability and accuracy of cerebral infarction risk prediction.

[0119] S6. Build a cerebral infarction warning mechanism, and conduct a graded assessment of individual health status based on the multimodal feature prediction results output by the model, dividing it into four levels: "healthy", "mild warning", "high risk", and "acute cerebral infarction".

[0120] Among them, the health level means that the physiological signals are stable and there are no abnormalities; the mild warning indicates that some key indicators such as blood pressure, cerebral blood flow or EEG signals have slight abnormal fluctuations; the high-risk level means that the monitoring signals are obviously abnormal and there is a possibility of cerebral infarction in the short term; the acute cerebral infarction level means that the physiological signal characteristics mutate or are seriously disordered, and an alarm signal must be issued immediately and the medical intervention module must be linked to achieve real-time response.

Claims

1. A method for monitoring brain health and early warning of cerebral infarction in the elderly, characterized in that: The following steps are involved: S1. Collect multimodal physiological signal data including EEG, ECG, blood oxygen and cerebral blood flow velocity; S2. Decomposing the multimodal physiological signal data using an eigenmode decomposition method, adjusting decomposition parameters, and extracting key features of instantaneous frequency, energy entropy, and kurtosis; S3, introduce the KAN network into the MATE structure, perform nonlinear mapping and feature reconstruction, and build the KAMATE model; S4. Based on the traditional Bayesian optimization algorithm and combined with the dynamic balance strategy, an improved Bayesian optimization algorithm is obtained; S5. Using the improved Bayesian optimization algorithm in combination with the eigenmode decomposition method to achieve eigendecomposition of the multimodal physiological signal and optimize the KAMATE model parameters; S6. Construct a cerebral infarction warning mechanism, and divide the brain health status into a healthy level, a mild warning level, a high-risk level, and an acute cerebral infarction level according to the prediction results output by the KAMATE model.

2. The early warning method according to claim 1, characterized in that: The multimodal physiological signal data collected by S1 can be expressed as: X=[X t-T ,X t-T+1 ,…,X t ]; Among them, X t represents all physiological signal features collected at time step t, and T represents the length of the history window.

3. The early warning method according to claim 1, characterized in that: The characteristic mode decomposition method described in S2 decomposes the input signal into the eigenmode function, and the decomposition expression is as follows: Among them, the IMF j represents the jth intrinsic mode function decomposed, R represents the residual term, and K represents the number of decomposed modes.

4. The early warning method according to claim 1, characterized in that: The S3 includes: S31. Extract the common structural features between the modalities through the latent representation encoder and output the shared latent variable Z c ; S32, through the modality-specific latent representation encoder, each modality is independently modeled to obtain exclusive feature representations and output a specific latent variable set [Z S1 ,Z S2 ,…,Z SM ], where M represents the total number of modalities and S represents specificity; S33, inputting the shared latent variables and the specific latent variables into the nonlinear feature reconstruction layer designed based on the KAN structure respectively, and reconstructing the latent vector Z using a combination of multiple sets of nonlinear functions; S34, mapping is performed through a nonlinear transformation function to obtain a mapped output Y; S44. The mapped shared latent variables and specific latent variables are fused in the time series dimension to form a time series feature tensor, and the multimodal physiological signal reconstruction and trend prediction values ​​are output.

5. The early warning method according to claim 4, characterized in that: The output Y formula of S34 is as follows: Among them, z j represents the jth component of the latent variable generated by MATE, w ij and b i Represent the weight and bias of the KAN network, φ i (·) represents a learnable nonlinear activation function, N represents the KAN output dimension, and d represents the input latent variable dimension.

6. The early warning method according to claim 1, characterized in that: The S4 includes: S41. Use Gaussian process regression as a proxy model to approximate the objective function and model the objective function; S42. Performing posterior inference on the Gaussian process based on the historical evaluation data set to obtain a predicted mean value of the objective function in the entire hyperparameter space; S43, using the expected improvement as an acquisition function, and using a dynamic balance strategy to regulate the search weight of the acquisition function, so that the dynamic weight changes adaptively with the number of iterations; S44. Select the next evaluation point by maximizing the acquisition function. If the maximum number of iterations is reached or the objective function converges, terminate the optimization process; otherwise, enter the next round of sampling iteration.

7. The early warning method according to claim 6, characterized in that: The objective function described in S41 is modeled as: f(θ)~GP(m(θ),k(θ,θ′)); Among them, f(θ) represents the target function, m(θ) represents the mean function, which is usually set to 0, and k(θ,θ ' ) represents the kernel function; The dynamic weight formula in S43 is: Among them, w t represents the exploration weight of the t-th iteration, α represents the smoothing factor (0<α<1), β represents the adjustment parameter, t represents the current iteration number, and T represents the total number of iterations; The acquisition function is: Acq(θ)=w t ·μ(θ)+(1-w t )·EI(θ); EI(θ)=E[max(f(θ)-f(θ + ),0)]; Among them, w t is the exploration weight of the tth iteration, between 0 and 1, μ(θ) is the mean prediction of f(θ), EI(θ) represents the acquisition function, f(θ + ) represents the optimal objective function value found so far.

8. The early warning method according to claim 1, characterized in that: The S5 includes: S51. An improved Bayesian optimization algorithm with a dynamic weight adjustment mechanism is introduced to optimize the parameters in the eigenmode decomposition module and KAMATE modeling. S52, the optimized eigenmode decomposition module provides time-frequency features, which are input into the KAMATE model; S53. The optimized KAMATE model monitors the dynamic change trends of EEG and ECG signals in real time.

9. The early warning method according to claim 1, characterized in that: The health level indicates that the physiological signal is stable and no abnormalities are found; the mild warning level indicates that some key indicators have slight abnormal fluctuations; the high-risk level indicates that the monitoring signal is obviously abnormal and there is a possibility of cerebral infarction in the short term; the acute cerebral infarction level indicates that the physiological signal characteristics have mutated or are seriously disordered, and an alarm signal must be issued immediately and the medical intervention module must be linked to achieve real-time response.

10. A brain health monitoring and cerebral infarction early warning system for the elderly, characterized by: include: Data acquisition module, used to collect multimodal physiological signal data such as EEG, ECG, blood oxygen and cerebral blood flow velocity; Algorithm optimization module, which is used to obtain an improved Bayesian optimization algorithm by combining the traditional Bayesian optimization algorithm with a dynamic balance strategy to optimize the parameter search process; The nonlinear reconstruction module is designed based on the MATE and KAN structures to build the KAMATE model and enhance the collaborative expression and time series modeling capabilities between multimodal signals. The eigenmode decomposition module is used to achieve eigendecomposition of multimodal physiological signals and optimize KAMATE model parameters through an improved Bayesian optimization algorithm; The cerebral infarction warning module is used to divide brain health status into four levels: healthy, mild warning, high risk and acute cerebral infarction based on the multimodal feature prediction results, to achieve real-time warning and intervention.

Images