A health assessment method and system based on a magnetocardiogram signal

Abstract

This invention discloses a health assessment method and system based on magnetocardiogram (MCC) signals, comprising the following steps: acquiring MCC signal data and preprocessing external physiological parameter data to generate a standardized MCC sequence; decomposing the standardized MCC sequence into multi-scale components to obtain low-frequency, mid-frequency, and high-frequency MCC components; inputting each component into three branches of an improved NARX network to generate multi-scale prediction results; gating and modulating the external physiological parameter data to update the multi-scale prediction results; constructing a cross-scale residual feedback path and coupling correction to generate a fusion prediction result; adjusting the delay order and feedback weights of the three branches based on the fusion prediction result and historical prediction errors through reinforcement learning to generate an optimized fusion prediction result; and performing health status mapping on the optimized fusion prediction result to generate a health score and risk level result. This invention achieves accurate assessment of MCC signals while improving the accuracy and stability of health identification results.

Description

Technical Field

[0001] This invention relates to the field of intelligent health assessment technology, and in particular to a health assessment method and system based on magnetic heart signals. Background Technology

[0002] Magnetocardiogram (MCC) signals, as important bioelectromagnetic signals reflecting the spatial distribution of cardiac electrical activity, possess characteristics such as non-contact and high sensitivity, making them valuable for cardiovascular function assessment and health monitoring. Current technologies primarily rely on electrocardiogram (ECG) signals or single physiological signals for health assessment, extracting features and determining status through time-domain analysis, frequency-domain analysis, or machine learning methods. Some studies have introduced multi-channel MCC acquisition devices to obtain richer spatial information and combined them with traditional filtering, feature extraction, and classification models to achieve health status identification. Other approaches employ deep learning models to model MCC signals, aiming to improve recognition accuracy and automation.

[0003] Existing technologies have significant shortcomings in practical applications. Magnetocardiogram (MCC) signals have weak amplitudes, making them susceptible to environmental magnetic interference and equipment noise. Traditional preprocessing methods struggle to suppress noise interference while preserving effective features, leading to unstable signal quality. In modeling, most methods use a single model to uniformly process the entire signal, failing to differentiate modeling for different frequency bands. This makes it difficult to simultaneously consider low-frequency trend information, mid-frequency structural information, and high-frequency disturbance information, impacting the accuracy of health assessments. While some multi-scale methods decompose the signal, they lack effective information exchange mechanisms between scales, failing to form a cross-scale collaborative modeling structure. External physiological parameters are often directly spliced ​​into the model, lacking modulation mechanisms tailored to the input process, resulting in irrelevant information interfering with model learning. Model parameters are often fixed or trained offline, lacking dynamic adjustment capabilities based on prediction errors, making it difficult to adapt to individual differences and state changes. Health assessment results often rely on simple mapping relationships, lacking stable, quantifiable scoring and risk grading mechanisms. These problems limit the effectiveness and stability of MCC signals in health assessment.

[0004] Therefore, how to provide a health assessment method and system based on magnetic heart signals is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a health assessment method and system based on magnetic field signals. This invention achieves fine modeling and state assessment of magnetic field signals through multi-scale decomposition, cross-scale coupling modeling, and gated modulation processing, and has the advantages of high assessment accuracy, strong noise resistance, and strong adaptability.

[0006] A health assessment method based on magnetic resonance imaging (MRI) signals according to an embodiment of the present invention includes the following steps:

[0007] Multi-channel magnetocardiogram (MCC) signal data and external physiological parameter data were acquired, preprocessed, and a standardized MCC sequence was generated. Multi-scale decomposition was performed on the standardized MCC sequence to obtain low-frequency, mid-frequency, and high-frequency MCC components. These components were then input into three branches of an improved NARX network to generate multi-scale prediction results. Gated modulation was performed on the external physiological parameter data to generate input modulation results, which were then applied to the input processes of the low-frequency, mid-frequency, and high-frequency modeling branches to update the multi-scale prediction results. The updated results were then analyzed. The multi-scale prediction results are used to construct a cross-scale residual feedback path, perform cross-scale coupling correction operations, and generate fused prediction results. The fused prediction results and historical prediction errors are used to construct state information, which is then input into the reinforcement learning policy network to perform policy decision operations, generate parameter adjustment actions, and apply the parameter adjustment actions to the delay order and feedback weights of the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch. The reward value is calculated based on the health status label and the fused prediction results, and the parameters of the reinforcement learning policy network are updated to generate optimized fused prediction results. Finally, a health status mapping operation is performed to generate health score results and risk level results.

[0008] Optionally, the preprocessing includes: performing bandpass filtering on the multi-channel magnetocardiogram (MCC) signal data to filter out low-frequency baseline drift and high-frequency environmental noise; calculating the local mean and subtracting the mean using a sliding window method to eliminate baseline shift; constructing a multi-channel signal matrix on the multi-channel MCC signal data after eliminating baseline shift; performing singular value decomposition to retain the principal singular components; reconstructing the multi-channel MCC signal; linearly scaling the reconstructed multi-channel MCC signal data according to the channel amplitude range; performing time alignment between channels based on key points of the cardiac cycle; and finally segmenting the time-aligned multi-channel MCC signal data according to a fixed-length sliding window to generate a standardized MCC sequence.

[0009] Optionally, the multi-scale decomposition operation on the standardized magnetic resonance imaging (MRI) sequence includes: detecting key points of the cardiac cycle along the time axis of the standardized MRI sequence, and dividing the periodic constraint decomposition interval by the sampling interval between adjacent key points of the cardiac cycle; setting long time windows, medium time windows, and short time windows in parallel within each periodic constraint decomposition interval, and performing multi-time window sliding decomposition on the standardized MRI sequence; performing local smoothing operation on the standardized MRI sequence using the long time window to extract low-frequency MRI components, and subtracting the low-frequency MRI components from the standardized MRI sequence to obtain a first residual sequence; and performing local structure extraction operation on the first residual sequence using the medium time window to obtain a medium-frequency MRI sequence. The magnetic component is obtained by subtracting the intermediate frequency (IF) magnetic component from the first residual sequence to obtain the second residual sequence. A local perturbation extraction operation is performed on the second residual sequence using a short time window to obtain the high-frequency magnetic component. The energy proportion within the periodic constraint decomposition interval is calculated for the low-frequency, intermediate-frequency, and high-frequency magnetic components, respectively. Based on the energy proportion, the window length and sliding step size of the long, medium, and short time windows are adaptively adjusted. The adaptively adjusted low-frequency, intermediate-frequency, and high-frequency magnetic components within each periodic constraint decomposition interval are then concatenated in chronological order to obtain the low-frequency, intermediate-frequency, and high-frequency magnetic components.

[0010] Optionally, the improved NARX network has three branches: a low-frequency modeling branch, a mid-frequency modeling branch, and a high-frequency modeling branch. The modeling process of the three branches includes: inputting the low-frequency magnetocardiogram (MCG) component into the low-frequency modeling branch, extracting the MCG component and the historical output value of the corresponding time step before the current time step according to a preset low-frequency delay order, constructing a low-frequency feedback input sequence in chronological order, inputting the low-frequency feedback input sequence into the low-frequency modeling branch to perform long-term trend modeling, and outputting the low-frequency trend result; inputting the low-frequency trend result as a trend constraint into the mid-frequency modeling branch, extracting the mid-frequency MCG component and the historical output value of the corresponding time step before the current time step according to a preset mid-frequency delay order, calculating the amplitude difference between the mid-frequency MCG components of adjacent time steps, constructing the mid-frequency feedback input sequence of the mid-frequency MCG component, historical output value, amplitude difference, and low-frequency trend result in chronological order, inputting the mid-frequency feedback input sequence into the mid-frequency modeling branch to perform waveform structure modeling, and outputting the mid-frequency structure result; inputting the mid-frequency structure result as a structure filter into the high-frequency modeling branch, extracting the low-frequency MCG component and the historical output value of the corresponding time step before the current time step according to a preset high-frequency delay order. The high-frequency magnetic field components at each time step and the corresponding historical output values ​​are used to calculate the rate of change between high-frequency magnetic field components at adjacent time steps. A high-frequency feedback input sequence is constructed by sequentially combining the high-frequency magnetic field components, historical output values, rates of change, and intermediate-frequency structure results. This high-frequency feedback input sequence is then input into the high-frequency modeling branch to perform local perturbation modeling, outputting high-frequency perturbation results. In the low-frequency modeling branch, a full-time historical output value writing method is used, sequentially writing the historical output values ​​of each time step into the low-frequency feedback input sequence. In the intermediate-frequency modeling branch, a historical output value difference-aligned writing method is used, writing the historical output values ​​of each time step into the intermediate-frequency feedback input sequence according to the position corresponding to the amplitude difference. In the high-frequency modeling branch, a historical output value mutation-triggered writing method is used, writing only the historical output values ​​corresponding to time steps with rates of change exceeding a preset threshold into the high-frequency feedback input sequence. The low-frequency trend results are used as the baseline results. The portion of the intermediate-frequency structure results that deviates from the low-frequency trend results by more than a preset deviation threshold is written into the baseline results. The portion of the high-frequency perturbation results with rates of change exceeding a preset threshold is written into the baseline results updated by the intermediate-frequency structure results, generating multi-scale prediction results.

[0011] Optionally, the process of generating the input modulation result includes:

[0012] External physiological parameter data are grouped according to parameter category to obtain rhythm parameter group, amplitude parameter group, and state parameter group. Normalization is performed on each of these groups, and they are aligned with the standardized magnetocardiogram sequence according to time order. The time-aligned rhythm parameter group is input into the first gating channel to calculate the first modulation coefficient corresponding to the low-frequency modeling branch. The time-aligned amplitude parameter group is input into the second gating channel to calculate the second modulation coefficient corresponding to the mid-frequency modeling branch. The time-aligned state parameter group is input into the third gating channel to calculate the third modulation coefficient corresponding to the high-frequency modeling branch. The first, second, and third modulation coefficients are applied to the current and historical input values ​​of the low-frequency, mid-frequency, and high-frequency modeling branches, respectively, to perform weighted modulation on the input at each time step, generating the input modulation result.

[0013] Optionally, the updating of the multi-scale prediction results includes: aligning the low-frequency, mid-frequency, and high-frequency prediction results according to time indices, using the low-frequency prediction results as the baseline sequence; calculating the difference sequence between the mid-frequency and low-frequency prediction results at each time step, and using the portion of the difference sequence whose absolute value is greater than a preset deviation threshold as the mid-frequency correction amount; superimposing the mid-frequency correction amount onto the low-frequency prediction results according to the corresponding time steps to obtain the first update sequence; calculating the rate of change of the high-frequency prediction results at adjacent time steps, and using the high-frequency prediction results at time steps with a rate of change greater than a preset change threshold as the high-frequency correction amount; superimposing the high-frequency correction amount onto the first update sequence according to the corresponding time steps to obtain the second update sequence; performing time continuity constraint processing on the second update sequence, limiting the abrupt change amplitude according to the amplitude difference between adjacent time steps, generating a smooth update sequence, and using the smooth update sequence as the updated multi-scale prediction result.

[0014] Optionally, the process of constructing a cross-scale residual feedback path and performing cross-scale coupling correction operations includes: aligning the low-frequency, mid-frequency, and high-frequency prediction results according to time indices, using the low-frequency prediction results as the reference sequence; calculating the difference between the mid-frequency and low-frequency prediction results at each time step to obtain the mid-frequency residual sequence, and writing the mid-frequency residual sequence as a feedback quantity into the corresponding time step position of the feedback input sequence of the mid-frequency modeling branch to correct the input of the mid-frequency modeling branch in subsequent time steps; calculating the difference between the high-frequency and mid-frequency prediction results at each time step to obtain the high-frequency residual sequence, and writing the high-frequency residual sequence as a feedback quantity into the feedback input sequence of the mid-frequency modeling branch at the corresponding time step position to correct the input of the mid-frequency modeling branch in subsequent time steps; and calculating the difference between the high-frequency and mid-frequency prediction results at each time step to obtain the high-frequency residual sequence. The column is written as a feedback quantity into the corresponding time step position in the feedback input sequence of the high-frequency modeling branch to correct the input of the high-frequency modeling branch in subsequent time steps; the residual amplitude and residual change rate are calculated for the intermediate frequency residual sequence and the high-frequency residual sequence respectively, and the residual sequence is weighted according to the residual amplitude and residual change rate; the weighted intermediate frequency residual sequence is superimposed on the low-frequency prediction result according to the time step to obtain the first coupled sequence, and the weighted high-frequency residual sequence is superimposed on the first coupled sequence according to the time step to obtain the second coupled sequence. The amplitude constraint processing is performed on the second coupled sequence, and the amplitude of each time step is within the preset range to generate the fusion prediction result.

[0015] Optionally, the process of generating the optimized fusion prediction result includes:

[0016] The difference between the fused prediction result and the health status label at each time step is calculated to obtain the current prediction error sequence. The current prediction error sequence is concatenated with the historical prediction error sequence in chronological order to construct an error sequence. The fused prediction result and the error sequence are concatenated at the same time step to construct state information. The state information is normalized and then input into the reinforcement learning policy network. Forward computation is performed on the state information in the reinforcement learning policy network to output a parameter adjustment action vector. The parameter adjustment action vector includes the delay order adjustment amount and feedback weight adjustment amount corresponding to the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch. The delay order adjustment amount in the parameter adjustment action vector is applied to the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch. The updated delay order is generated by increasing or decreasing the delay order of each branch. The feedback weight adjustment amount in the parameter adjustment action vector is applied to the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch. The updated feedback weight is generated by increasing or decreasing the historical output weight in the feedback input sequence of each branch.

[0017] The reward value is calculated based on the health status label and the fusion prediction result. The reward value is obtained by performing an inverse proportional mapping on the absolute value of the prediction error. The reward value, the current state information and the parameter adjustment action vector are input into the reinforcement learning policy network to perform parameter update operation and iteratively update the parameters of the reinforcement learning policy network. The updated delay order and feedback weights are reapplied to the low-frequency modeling branch, the mid-frequency modeling branch and the high-frequency modeling branch to recalculate the fusion prediction result and generate the optimized fusion prediction result.

[0018] Optionally, the process of performing the health status mapping operation includes: extracting the mean amplitude, variance amplitude, average amplitude difference between adjacent time steps, dominant frequency and spectral energy distribution from the optimized fusion prediction results; constructing a health assessment feature vector from the extracted results in a preset order; inputting the health assessment feature vector into a preset scoring function to perform weighted calculation and normalization processing to generate a health score result; and then comparing the health score result with a preset risk threshold range to generate a risk level result.

[0019] A health assessment system based on magnetic resonance imaging (MRI) signals according to an embodiment of the present invention includes the following modules:

[0020] The data acquisition module is used to acquire multi-channel magnetic resonance imaging (MRI) signal data and external physiological parameter data.

[0021] The preprocessing module is used to perform preprocessing on multi-channel magnetocardiogram (MCC) signal data to generate standardized MCC sequences;

[0022] The multi-scale decomposition module is used to perform multi-scale decomposition operations on standardized magnetic cardiomyocyte sequences to obtain low-frequency, mid-frequency, and high-frequency magnetic cardiomyocyte components.

[0023] The modeling and prediction module is used to input the low-frequency, mid-frequency, and high-frequency magnetic cardiomyocyte components into the three branches of the improved NARX network to generate multi-scale prediction results.

[0024] The gated modulation module is used to perform gated modulation operations on external physiological parameter data, generate input modulation results, and apply them to the input processes of the low-frequency modeling branch, the mid-frequency modeling branch, and the high-frequency modeling branch to update the multi-scale prediction results.

[0025] The residual feedback module is used to construct cross-scale residual feedback paths from the high-frequency modeling branch to the mid-frequency modeling branch and from the mid-frequency modeling branch to the low-frequency modeling branch, perform cross-scale coupling correction operations, and generate fusion prediction results.

[0026] The reinforcement learning optimization module is used to construct state information by combining the fused prediction results with historical prediction errors, inputting it into the reinforcement learning policy network to perform policy decision operations, generating parameter adjustment actions and updating the parameters of the reinforcement learning policy network, and generating optimized fused prediction results.

[0027] The health assessment module is used to perform health status mapping operations on the optimized fusion prediction results to generate health score results and risk level results.

[0028] The beneficial effects of this invention are:

[0029] (1) By performing preprocessing and multi-scale decomposition operations on multi-channel magnetic heart signals, and combining low-frequency modeling branch, mid-frequency modeling branch and high-frequency modeling branch to construct an improved NARX network, the differentiated modeling processing of different frequency band features can be realized, which can simultaneously characterize the trend information, structural information and disturbance information in the magnetic heart signals, and improve the signal representation ability and health assessment accuracy.

[0030] (2) By performing gating modulation operations on external physiological parameter data, the input modulation results are applied to the input process of each modeling branch. Combined with cross-scale residual feedback path and coupling correction operation, the linkage update and collaborative modeling between multi-scale information are realized, enhancing the model's adaptability to complex physiological state changes and improving the stability and anti-interference ability of the evaluation results.

[0031] (3) By constructing a reinforcement learning policy network, the state model of the fusion prediction results and historical prediction errors is performed, the delay order and feedback weight of each branch are dynamically adjusted, and the health score results and risk level results are generated by combining health status mapping operation, so as to realize the adaptive optimization of model parameters and the quantitative output of health status, thereby improving the accuracy and interpretability of the evaluation results. Attached Figure Description

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0033] Figure 1 This is an overall flowchart of a health assessment method based on magnetic cardiac signals proposed in this invention;

[0034] Figure 2 This is a schematic diagram of the network optimization process of the reinforcement learning strategy proposed in this invention, as well as the process for generating health score results and risk level results.

[0035] Figure 3 This is a module connection diagram of a health assessment system based on magnetic heart signals proposed in this invention. Detailed Implementation

[0036] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0037] refer to Figures 1-3 A health assessment method based on magnetic heart signals includes the following steps:

[0038] Multi-channel magnetocardiogram (MCC) signal data and external physiological parameter data were acquired, preprocessed, and a standardized MCC sequence was generated. Multi-scale decomposition was performed on the standardized MCC sequence to obtain low-frequency, mid-frequency, and high-frequency MCC components. These components were then input into three branches of an improved NARX network: a low-frequency modeling branch, a mid-frequency modeling branch, and a high-frequency modeling branch. These branches employed different delay orders and, combined with their historical outputs, performed nonlinear autoregressive modeling to generate multi-scale prediction results. Gated modulation was performed on the external physiological parameter data to generate input modulation results, which were then applied to the outputs of the low-frequency, mid-frequency, and high-frequency modeling branches. The process involves updating the multi-scale prediction results; constructing cross-scale residual feedback paths from the high-frequency modeling branch to the mid-frequency modeling branch and from the mid-frequency modeling branch to the low-frequency modeling branch for the updated multi-scale prediction results; performing cross-scale coupling correction operations to generate fused prediction results; constructing state information from the fused prediction results and historical prediction errors; inputting this information into the reinforcement learning policy network to perform policy decision operations; generating parameter adjustment actions; applying these actions to the delay order and feedback weights of the low-frequency, mid-frequency, and high-frequency modeling branches; calculating reward values ​​based on health status labels and fused prediction results; updating the reinforcement learning policy network parameters; and generating optimized fused prediction results. Finally, performing health status mapping operations on the optimized fused prediction results to generate health score results and risk level results.

[0039] In this embodiment, the preprocessing includes: performing bandpass filtering on the multi-channel magnetocardiogram (MCC) signal data to filter out low-frequency baseline drift and high-frequency environmental noise; calculating the local mean and subtracting the mean using a sliding window method to eliminate baseline shift; constructing a multi-channel signal matrix from the baseline-shift-eliminated MCC signal data; performing singular value decomposition to retain the principal singular components; reconstructing the multi-channel MCC signal; linearly scaling the reconstructed multi-channel MCC signal data according to the channel amplitude range; performing time alignment between channels based on key points of the cardiac cycle; and finally segmenting the time-aligned multi-channel MCC signal data according to a fixed-length sliding window to generate a standardized MCC sequence.

[0040] In this embodiment, the preprocessing of multi-channel magnetocardiogram (MCC) signal data is performed in a fixed order. First, bandpass filtering is applied to the acquired MCC signals of each channel. The lower cutoff frequency of the bandpass filter is used to suppress sensor drift, slow respiratory variation disturbances, and slow baseline fluctuations, while the upper cutoff frequency is used to suppress high-frequency electromagnetic interference from the environment and spike noise introduced by the acquisition circuit. A digital finite impulse response (FIR) filter is used for the bandpass filtering. To avoid phase shift after processing different channels, a zero-phase filtering method is preferably used to perform bidirectional filtering on the MCC signals of each channel to keep the timing position of the waveforms of each channel consistent.

[0041] After bandpass filtering, baseline offset elimination is performed on the magnetocardiogram (MCC) signals of each channel. The specific processing method is as follows: a sliding window is set for each channel of the MCC signal along the time axis. The average value of the sampling point amplitude is calculated within each sliding window. The average value is used as the local baseline estimate at the current position. Then, the original sampling value at the current position is subtracted from the corresponding local baseline estimate to obtain the baseline-corrected MCC signal. The length of the sliding window is set according to the cardiac cycle range so that the sliding window can cover low-frequency drift changes without weakening the main fluctuation characteristics of the effective MCC waveform.

[0042] After eliminating baseline offset, multi-channel collaborative noise reduction is performed on the multi-channel magnetocardiogram (MCC) signals. Specifically, the amplitudes of each channel at the same sampling time are arranged in channel order to form a multi-channel signal matrix, where the rows of the matrix correspond to the sampling time and the columns correspond to the acquisition channels. Singular value decomposition (SVD) is performed on the multi-channel signal matrix, decomposing it into principal component matrix, singular value matrix, and orthogonal basis matrix. The decomposition results are sorted according to the magnitude of the singular values, retaining the first few singular components that represent the main MCC energy distribution, and removing singular components with significantly smaller amplitudes and discrete distributions. The multi-channel MCC signals are reconstructed based on the retained singular components. The selection of the first few singular components can be determined based on the cumulative energy ratio, so that the reconstructed signal can retain the main MCC waveform structure while suppressing random noise between channels and local burst noise.

[0043] After collaborative noise reduction, linear scaling is performed on the magnetocardiogram (MCG) signals of each channel. Specifically, the maximum and minimum amplitudes of the MCG signals of each channel within the current sampling segment are calculated, and the amplitude of each sampling point is mapped to a preset numerical range to ensure that the numerical scales of different channels remain consistent. This processing method can avoid significant deviations in amplitude range due to differences in sensor sensitivity and amplification factor in individual channels, thereby reducing bias caused by inconsistencies in dimensions during multi-scale decomposition and improved NARX network modeling.

[0044] After linear scaling, time alignment is performed on the multi-channel magnetic resonance imaging (MRI) signals. Specifically, key time points with stable representativeness are located within each cardiac cycle. Using the key time points as reference times, the MRI signals of the other channels are shifted and corrected to keep the main feature positions of the waveforms of different channels within the same cardiac cycle synchronized. Key time points can be determined by main peak location, local extreme value search, or energy mutation point detection. After time alignment is completed, the waveform changes of different channels reflecting the same cardiac activity can be compared and jointly modeled under a unified time reference.

[0045] After time alignment, sliding window segmentation is performed on the multi-channel magnetocardiogram (MCC) signals to generate standardized MCC sequences. Specifically, a fixed-length window and a fixed step size are set along the time axis to sequentially extract continuous MCC signals. Each window contains a complete continuous sampling segment. The window length is set to cover at least one complete cardiac cycle, and the step size is set to be smaller than the window length to preserve overlapping areas between adjacent segments.

[0046] To ensure modeling stability, abnormal sampling segments are removed before generating standardized magnetocardiogram (MCC) sequences. The criteria for determining abnormal sampling segments include: continuous saturation of sampling points, constant amplitude across the entire segment, local mutations exceeding the preset amplitude limit, and insufficient number of effective sampling points. Sampling segments identified as abnormal are discarded and not included in subsequent standardized MCC sequences. This approach avoids significantly distorted sampling segments from entering the improved NARX network, which could affect the stability of health assessment results.

[0047] In this embodiment, performing multi-scale decomposition on the standardized magnetic resonance imaging (MRI) sequence includes: detecting key points of the cardiac cycle along the time axis of the standardized MRI sequence; dividing the periodic constraint decomposition interval by the sampling interval between adjacent key points of the cardiac cycle; setting long-term windows, medium-term windows, and short-term windows in parallel within each periodic constraint decomposition interval; performing multi-time-window sliding decomposition on the standardized MRI sequence using the long-term window; and performing local smoothing on the standardized MRI sequence using the long-term window to extract low-frequency MRI components. The local smoothing operation involves performing a weighted moving average on the standardized MRI sequence within the long-term window, assigning different weights based on the time interval between the sampling points and the center position within the window, assigning smaller weights to sampling points far from the center position and larger weights to sampling points close to the center position. The process involves suppressing high-frequency fluctuations, preserving gradual trend information, generating low-frequency magnetocardiogram (MCG) components, and subtracting these low-frequency MCG components from the standardized MCG sequence to obtain the first residual sequence. A local structure extraction operation is then performed on the first residual sequence using a mid-time window to obtain the mid-frequency MCG components. This local structure extraction operation involves performing local peak and valley detection, waveform inflection point location, and morphological matching of adjacent sampling segments within the mid-time window. First, the amplitude difference and slope change between each sampling point and its adjacent sampling points within the mid-time window are calculated. Positions where the amplitude change is continuous and the slope change direction switches are identified as candidate inflection points. Local structural segments are constructed using sampling segments between adjacent candidate inflection points. The duration, amplitude span, and morphological similarity of each local structural segment are calculated to preserve... Local structural segments whose duration is within a preset mid-frequency range and whose amplitude span is higher than the low-frequency residual fluctuation threshold are retained. These retained local structural segments are reconstructed chronologically to generate mid-frequency magnetocardiogram (MCG) components. The first residual sequence is subtracted from the mid-frequency MCG component to obtain a second residual sequence. A local perturbation extraction operation is performed on the second residual sequence using a short time window to obtain the high-frequency MCG component. The local perturbation extraction operation calculates the amplitude change rate between adjacent sampling points within the short time window, enhancing sampling intervals with change rates exceeding a preset threshold and suppressing sampling intervals with change rates below the preset threshold, thereby highlighting sudden fluctuations and fine-grained perturbation characteristics to generate the high-frequency MCG component. The low-frequency, mid-frequency, and high-frequency MCG components are then processed separately... The energy proportion within the periodic constraint decomposition interval is calculated, and the window lengths and sliding step sizes of the long, medium, and short time windows are adaptively adjusted based on the energy proportion. The adaptive adjustment compares the change amplitude of the energy proportion at the same scale between the current periodic constraint decomposition interval and adjacent periodic constraint decomposition intervals. When the change amplitude exceeds a preset threshold, the corresponding time window length is increased and the sliding step size is decreased. When the change amplitude is lower than the preset threshold, the time window length is decreased and the sliding step size is increased, so that the energy distribution at the same scale within each periodic constraint decomposition interval tends to be stable. The low-frequency, medium-frequency, and high-frequency magnetic cardiopulmonary components (MCP) in each periodic constraint decomposition interval are spliced ​​in chronological order to obtain the low-frequency, medium-frequency, and high-frequency MCP components.

[0048] In this embodiment, the improved NARX network has three branches: a low-frequency modeling branch, a mid-frequency modeling branch, and a high-frequency modeling branch. The modeling process of the three branches includes: inputting low-frequency magnetocardiogram (MCC) components into the low-frequency modeling branch; extracting the MCC components and historical output values ​​of several consecutive time steps prior to the current time step according to a preset low-frequency delay order; constructing a low-frequency feedback input sequence in chronological order; inputting the low-frequency feedback input sequence into the low-frequency modeling branch to perform long-term trend modeling; and outputting the low-frequency trend result. Specifically, long-term trend modeling involves performing multi-layer nonlinear mapping operations on the low-frequency feedback input sequence in the low-frequency modeling branch, mapping the inputs of each time step to hidden layer nodes through weighted connections, and performing cumulative processing on the cross-time step inputs in the hidden layer. The process employs a weighted multiplication method to ensure that low-frequency information from historical time steps continuously participates in the current output calculation during the mapping process. This extracts the overall trend of low-frequency magnetocardiogram (MCG) components and outputs the low-frequency trend result. The low-frequency trend result is then used as a trend constraint input to the intermediate frequency (IF) modeling branch. Following a preset IF delay order, the branch extracts IFG components from several consecutive time steps prior to the current moment and the corresponding historical output values. It calculates the amplitude difference between IFG components from adjacent time steps and constructs an IF feedback input sequence from the IFG components, historical output values, amplitude differences, and low-frequency trend result in chronological order. This IF feedback input sequence is then input to the IF modeling branch to perform waveform structure modeling, outputting the IF structure result. Specifically, waveform structure modeling involves analyzing the IFG feedback input sequence from adjacent time steps... The amplitude difference of each step undergoes a weighted mapping process. This enhances the contribution of continuous amplitude variations during the nonlinear mapping process while suppressing discontinuous amplitude variations, thus highlighting the periodic structure and local morphological features of the waveform and yielding the mid-frequency structure result. This mid-frequency structure result is then used as the structure screening input to the high-frequency modeling branch. Following a preset high-frequency delay order, the high-frequency magnetocardiogram (MCC) components and corresponding historical output values ​​from several consecutive time steps prior to the current moment are extracted. The rate of change between the MCC components of adjacent time steps is calculated. The MCC components, historical output values, rate of change, and mid-frequency structure result are then combined in chronological order to construct a high-frequency feedback input sequence. This sequence is then input into the high-frequency modeling branch to perform local perturbation modeling, outputting the high-frequency perturbation result. Specifically, local perturbation modeling involves thresholding and weighting the rate of change in the high-frequency feedback input sequence. Time steps with a rate of change exceeding a preset threshold are assigned higher weights for nonlinear mapping, while time steps with a rate of change below the preset threshold are assigned lower weights. This makes the mapping results highlight sudden perturbations and fine-grained changes, thus obtaining high-frequency perturbation results. In the low-frequency modeling branch, a full-time writing method for historical output values ​​is adopted, sequentially writing the historical output values ​​of each time step into the low-frequency feedback input sequence. The full-time writing method involves writing the historical output values ​​corresponding to each time step into the low-frequency feedback input sequence one by one in chronological order, ensuring that all historical outputs participate in the modeling calculation at the current moment, thus guaranteeing the integrity of time-dependent information during the low-frequency trend modeling process.In the mid-frequency modeling branch, a historical output value difference-aligned writing method is adopted. Historical output values ​​at each time step are written into the mid-frequency feedback input sequence according to the position corresponding to the amplitude difference. The difference-aligned writing method determines the writing position of historical output values ​​based on the amplitude difference of the mid-frequency magnetocardiogram components of adjacent time steps, prioritizing the writing of historical output values ​​corresponding to time steps with larger amplitude differences into the mid-frequency feedback input sequence, thus giving higher weight to regions with significant waveform changes during modeling. In the high-frequency modeling branch, a historical output value mutation-triggered writing method is adopted. Only historical output values ​​corresponding to time steps with a change rate exceeding a preset change threshold are written into the high-frequency feedback input sequence. The mutation-triggered writing method performs a threshold judgment on the change rate of high-frequency magnetocardiogram components. Only when the change rate exceeds the preset change threshold is the historical output value of the corresponding time step written into the high-frequency feedback input sequence, allowing historical information from local mutation regions to participate in modeling calculations and suppressing interference from non-mutation regions. The low-frequency trend results are used as the baseline results. The portion of the mid-frequency structure results that deviates from the low-frequency trend results by more than a preset deviation threshold is written into the baseline results. The portion of the high-frequency disturbance results with a change rate exceeding the preset change threshold is written into the baseline results updated by the mid-frequency structure results, generating multi-scale prediction results.

[0049] In this embodiment, the process of generating the input modulation result includes:

[0050] External physiological parameter data are grouped according to parameter category to obtain rhythm parameter group, amplitude parameter group, and state parameter group. Normalization is performed on each of these groups, and they are aligned with the standardized magnetocardiogram sequence according to time order. The time-aligned rhythm parameter group is input into the first gating channel to calculate the first modulation coefficient corresponding to the low-frequency modeling branch. The first modulation coefficient is obtained by performing a linear weighted summation of the parameter values ​​of the rhythm parameter group at the current time and then inputting it into a nonlinear activation function. Specifically, this involves multiplying each parameter in the rhythm parameter group by a preset weight, summing the results, and then... The Sigmoid function is used for mapping, so that the first modulation coefficient takes a value between 0 and 1, which is used to characterize the modulation intensity of the current rhythm change on the low-frequency modeling branch input. The time-aligned amplitude parameter set is input into the second gated channel to calculate the second modulation coefficient corresponding to the intermediate frequency modeling branch. The second modulation coefficient is obtained by performing difference enhancement processing on the parameter values ​​of the amplitude parameter set at the current time, followed by weighted summation, and then inputting it into a nonlinear activation function. Specifically, the deviation of each parameter in the amplitude parameter set from the historical mean is first calculated, then the deviation is multiplied by a preset weight and summed, and then applied through the Sigmoid function. The mapping generates a second modulation coefficient, giving higher modulation weights to moments with significant changes in the intermediate frequency waveform structure. The time-aligned state parameter set is input into the third gated channel to calculate the third modulation coefficient corresponding to the high-frequency modeling branch. The third modulation coefficient is obtained by performing threshold judgment and piecewise mapping on the parameter values ​​of the state parameter set at the current moment. When a state parameter exceeds a preset state threshold, the corresponding parameter value is multiplied by an enhancement coefficient and then weighted and summed. The third modulation coefficient is then generated through mapping using the Sigmoid function, giving enhanced weights to high-frequency disturbance inputs under abnormal conditions. The first, second, and third modulation coefficients are applied to the current and historical input values ​​of the low-frequency, intermediate-frequency, and high-frequency modeling branches, respectively. Weighted modulation is performed on the input at each time step. This weighted modulation is achieved by multiplying the input value at each time step by the modulation coefficient of the corresponding branch point by point. The current input value is directly multiplied by the modulation coefficient, while the historical input value is multiplied by the modulation coefficient after introducing an attenuation factor based on its distance from the current time. This ensures that inputs closer to the current time maintain higher weights, while inputs farther from the current time have progressively lower weights, generating the input modulation result.

[0051] In this embodiment, updating the multi-scale prediction results includes: aligning the low-frequency, mid-frequency, and high-frequency prediction results according to time indices, using the low-frequency prediction results as the baseline sequence; calculating the difference sequence between the mid-frequency and low-frequency prediction results at each time step, and using the portion of the difference sequence whose absolute value is greater than a preset deviation threshold as the mid-frequency correction amount; superimposing the mid-frequency correction amount onto the low-frequency prediction results according to the corresponding time steps to obtain the first update sequence; calculating the rate of change of the high-frequency prediction results at adjacent time steps, and using the high-frequency prediction results at time steps with a rate of change greater than a preset change threshold as the high-frequency correction amount; superimposing the high-frequency correction amount onto the first update sequence according to the corresponding time steps to obtain the second update sequence; performing time continuity constraint processing on the second update sequence, limiting the abrupt change amplitude according to the amplitude difference between adjacent time steps, generating a smooth update sequence, and using the smooth update sequence as the updated multi-scale prediction result;

[0052] Specifically, the second update sequence is traversed along the time axis for each time step. The difference between the amplitude of the current time step and the amplitude of the previous time step is calculated to obtain the amplitude difference between adjacent time steps. The amplitude difference is compared with a preset mutation threshold. When the amplitude difference is greater than the preset mutation threshold, the amplitude of the current time step is truncated in the manner that the amplitude of the current time step is equal to the amplitude of the previous time step plus the preset mutation threshold. When the amplitude difference is less than the preset mutation threshold, the amplitude of the current time step remains unchanged. The same processing is performed on the sequence after truncation along the time axis until the traversal is completed, resulting in a preliminary constraint sequence. Local mean smoothing is performed on the preliminary constraint sequence using a fixed-length sliding window. The amplitude of each time step within each sliding window is averaged, and the average value replaces the amplitude at the center of the window to eliminate discontinuous changes caused by local mutations, ultimately generating a smooth update sequence.

[0053] In this embodiment, the process of constructing a cross-scale residual feedback path and performing cross-scale coupling correction operations includes: aligning the low-frequency, mid-frequency, and high-frequency prediction results according to time indices, using the low-frequency prediction results as the baseline sequence; calculating the difference between the mid-frequency and low-frequency prediction results at each time step to obtain the mid-frequency residual sequence, and writing the mid-frequency residual sequence as a feedback quantity into the corresponding time step position of the feedback input sequence of the mid-frequency modeling branch to correct the input of the mid-frequency modeling branch in subsequent time steps; calculating the difference between the high-frequency and mid-frequency prediction results at each time step to obtain the high-frequency residual sequence, and writing the high-frequency residual sequence as a feedback quantity into the high-frequency modeling branch. The corresponding time step position in the feedback input sequence of the branch is used to correct the input of subsequent time steps of the high-frequency modeling branch; the residual amplitude and residual change rate are calculated for the intermediate frequency residual sequence and the high-frequency residual sequence respectively, and the residual sequence is weighted according to the residual amplitude and residual change rate so that the time steps with larger residual amplitude and higher change rate are given higher weights; the weighted intermediate frequency residual sequence is superimposed on the low-frequency prediction result according to the time step to obtain the first coupled sequence, and the weighted high-frequency residual sequence is superimposed on the first coupled sequence according to the time step to obtain the second coupled sequence. The amplitude constraint processing is performed on the second coupled sequence, and the amplitude of each time step is within the preset range to generate the fusion prediction result.

[0054] In this embodiment, the process of generating the optimized fusion prediction result includes:

[0055] The difference between the fused prediction result and the health status label at each time step is calculated to obtain the current prediction error sequence. The current prediction error sequence is then concatenated with the historical prediction error sequences in chronological order to construct an error sequence. The fused prediction result and the error sequence are concatenated at the same time step to construct state information. This state information is normalized and then input into the reinforcement learning policy network. Forward computation is performed on the state information within the reinforcement learning policy network, outputting a parameter adjustment action vector. Specifically, this includes: writing the state information as an input vector into the input layer of the policy network; performing a linear combination operation on the input vector according to preset connection weights to obtain the first layer intermediate result; and then inputting the first layer intermediate result element-wise into a nonlinear activation function for mapping processing to obtain... The first layer outputs the result; the first layer output is used as the input for the next layer, and the linear combination operation and nonlinear activation function mapping are repeatedly performed, passing layer by layer until the output layer; at the output layer, the final mapping result is subjected to linear transformation or normalization to generate a parameter adjustment action vector. The linear combination operation is a weighted sum of the input vector and connection weights, and the nonlinear activation function is either the Sigmoid function or the hyperbolic tangent function. Through layer-by-layer mapping, a functional approximation relationship is achieved between state information and the parameter adjustment action vector, thus completing the policy decision. The parameter adjustment action vector includes the delay order adjustment and feedback weight adjustment for the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch; the delay order adjustment in the parameter adjustment action vector is... This method is used for low-frequency, mid-frequency, and high-frequency modeling branches. Updated delay orders are generated by increasing or decreasing the delay order of each branch. Specifically, the delay order adjustment value output by the reinforcement learning policy network is mapped to a discrete adjustment instruction. The discrete adjustment instruction can be increased by one order, decreased by one order, or remain unchanged. The delay order of the current branch is superimposed with the discrete adjustment instruction. When the adjustment instruction is to increase by one order, the current delay order is incremented by one; when the instruction is to decrease by one order, the current delay order is decremented by one; when the instruction is to remain unchanged, the original value is maintained. During the update process, upper and lower limits are set for the delay order. When the updated delay order exceeds the upper limit, it is truncated to the upper limit; when it is below the lower limit, it is truncated. The lower limit is set as follows: After obtaining the updated delay order, the historical input sequence of the corresponding branch is reconstructed according to the new delay order, that is, the corresponding number of time steps of data before the current moment are reselected as input to complete the update of the delay structure; The feedback weight adjustment amount in the parameter adjustment action vector is applied to the low-frequency modeling branch, the mid-frequency modeling branch, and the high-frequency modeling branch, and the updated feedback weight is generated by increasing or decreasing the historical output weight in the feedback input sequence of each branch; Specifically, the feedback weight adjustment amount of the reinforcement learning policy network output is mapped to a weight adjustment vector of length corresponding to the delay order; The historical output value corresponding to each time step in the feedback input sequence is multiplied by the weight adjustment amount at the corresponding position to obtain the updated historical output weight;The weight adjustment amount is obtained by linear scaling within a preset range, ensuring that the weight adjustment value lies within a set interval. During the update process, upper and lower limits are set for the weight values; when the updated weight exceeds the upper limit, it is truncated, and when it falls below the lower limit, it is increased. The updated weights are used to recalculate the feedback input sequence, i.e., the historical output value at each time step is multiplied by the corresponding updated weight and then written into the feedback input sequence, changing the influence of historical information on the current modeling. A reward value is calculated based on the health status label and the fusion prediction result. The reward value is obtained by performing an inverse proportional mapping on the absolute value of the prediction error, increasing the reward value when the prediction error decreases. The reward value, current state information, and parameter adjustment action vector are input into the reinforcement learning policy network to perform parameter update operations, iteratively updating the parameters of the reinforcement learning policy network. Specifically, after completing one fusion prediction... After the results are calculated, the prediction error is calculated based on the current prediction result and the health status label, and the reward value is calculated accordingly. The state information, parameter adjustment actions, and reward value are input into the reinforcement learning policy network to perform a parameter update. After the parameter update is completed, the input sequences of each branch are reconstructed using the updated delay order and feedback weights, and the prediction operations of the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch are re-executed to generate a new fusion prediction result. The new fusion prediction result is used again to construct the state information, and the above process is repeated until the preset termination conditions are met. The termination conditions include the prediction error being less than the preset error threshold or the number of iterations reaching the preset maximum number of iterations. The updated delay order and feedback weights are then applied back to the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch, and the fusion prediction result is recalculated to generate an optimized fusion prediction result.

[0056] In this embodiment, the process of performing the health status mapping operation includes: extracting the mean amplitude, variance amplitude, average amplitude difference between adjacent time steps, dominant frequency and spectral energy distribution from the optimized fusion prediction results; constructing a health assessment feature vector from the extracted results in a preset order; inputting the health assessment feature vector into a preset scoring function to perform weighted calculation and normalization to generate a health score result; and then comparing the health score result with a preset risk threshold range to generate a risk level result.

[0057] In this embodiment, the preset scoring function is constructed using a weighted summation method. The mean amplitude, variance amplitude, average amplitude difference between adjacent time steps, dominant frequency, and spectral energy distribution are extracted from the optimized fusion prediction results as health assessment features. Then, a corresponding weight coefficient is set for each feature, and the sum of the weight coefficients is set to 1. Each feature is multiplied by its corresponding weight and then summed to obtain a health score. The health score is linearly mapped to the minimum and maximum scores in historical samples to convert the health score into a normalized score within a standard interval. Finally, the normalized score is compared with a preset risk threshold interval, and the corresponding risk level is determined based on the interval in which the score falls, thereby generating a health score and a risk level result.

[0058] In this embodiment, the health score result is used to characterize the overall health status level of the tested subject within the current time period, specifically including: standardized score value, score change trend, and score stability index; wherein, the standardized score value is a value after normalization processing, used to reflect the absolute level of the current health status, the score change trend is used to describe the direction and magnitude of the change of the score value within a continuous time window, and the score stability index is obtained by calculating the degree of fluctuation of the score value within adjacent time windows, used to characterize the stability of the health status.

[0059] The risk level results are used to classify and describe health status, specifically including risk level labels, risk range, and risk trigger indicators. Among them, the risk level label indicates the risk category corresponding to the current status, the risk range limits the numerical range to which the score value belongs, and the risk trigger indicator indicates whether the score value has reached the preset risk threshold and triggers the warning condition.

[0060] During output, health score results and risk level results are recorded in chronological order and correlated with the fusion prediction results of the corresponding time step for health status tracking and change analysis.

[0061] A health assessment system based on magnetic resonance imaging (MRI) signals according to an embodiment of the present invention includes the following modules:

[0062] The data acquisition module is used to acquire multi-channel magnetic resonance imaging (MRI) signal data and external physiological parameter data.

[0063] The preprocessing module is used to perform preprocessing on multi-channel magnetocardiogram (MCC) signal data to generate standardized MCC sequences;

[0064] The multi-scale decomposition module is used to perform multi-scale decomposition operations on standardized magnetic cardiomyocyte sequences to obtain low-frequency, mid-frequency, and high-frequency magnetic cardiomyocyte components.

[0065] The modeling and prediction module is used to input the low-frequency, mid-frequency, and high-frequency magnetic cardiomyocyte components into the three branches of the improved NARX network to generate multi-scale prediction results.

[0066] The gated modulation module is used to perform gated modulation operations on external physiological parameter data, generate input modulation results, and apply them to the input processes of the low-frequency modeling branch, the mid-frequency modeling branch, and the high-frequency modeling branch to update the multi-scale prediction results.

[0067] The residual feedback module is used to construct cross-scale residual feedback paths from the high-frequency modeling branch to the mid-frequency modeling branch and from the mid-frequency modeling branch to the low-frequency modeling branch, perform cross-scale coupling correction operations, and generate fusion prediction results.

[0068] The reinforcement learning optimization module is used to construct state information by combining the fused prediction results with historical prediction errors, inputting it into the reinforcement learning policy network to perform policy decision operations, generating parameter adjustment actions and updating the parameters of the reinforcement learning policy network, and generating optimized fused prediction results.

[0069] The health assessment module is used to perform health status mapping operations on the optimized fusion prediction results to generate health score results and risk level results.

[0070] Example 1: To verify the feasibility of this invention in practice, it is applied to a continuous health status assessment scenario based on magnetocardiogram (MCC) signals. In this scenario, a non-contact, multi-channel MCC acquisition device is used to acquire the MCC signals of the subject at rest and during mild activity, while simultaneously acquiring heart rate, blood pressure changes, and other status parameters. Traditional methods in this scenario suffer from significant signal-noise interference, aliasing of information across different frequency bands, and significant fluctuations in assessment results, leading to unstable health status determination and difficulty in achieving continuous tracking and analysis. This implementation addresses these issues by introducing multi-scale decomposition and an improved NARX modeling structure to perform hierarchical modeling of the MCC signals, and by combining gated modulation and reinforcement learning optimization mechanisms.

[0071] In practical applications, the acquired multi-channel magnetocardiogram (MCC) signals are first subjected to bandpass filtering, baseline correction, and multi-channel collaborative noise reduction to obtain a stable, standardized MCC sequence. Subsequently, this sequence is decomposed into low-frequency, mid-frequency, and high-frequency MCC components. The low-frequency component reflects overall cardiac rhythm changes, the mid-frequency component reflects waveform structure changes, and the high-frequency component reflects local perturbation characteristics. By inputting these three types of components into the low-frequency, mid-frequency, and high-frequency modeling branches respectively for modeling, independent modeling and collaborative fusion of information at different scales are achieved.

[0072] During the modeling process, external physiological parameters participate in the input process through a gating modulation mechanism. This results in rhythm-related parameters having a higher weighting influence on the low-frequency modeling branch, amplitude variation parameters modulating the mid-frequency modeling branch, and state variation parameters enhancing the high-frequency modeling branch. This allows the model to dynamically adjust the input feature weights according to different physiological states. After obtaining preliminary multi-scale prediction results, a cross-scale residual feedback path is constructed to feed high-frequency perturbation information back to the mid-frequency modeling branch and mid-frequency structural information back to the low-frequency modeling branch. This establishes an information linkage between the various scales, resulting in a fused prediction result.

[0073] In further processing, the fused prediction results and historical prediction errors are used to construct state information, which is then input into a reinforcement learning policy network. The policy decision output parameters are adjusted to update the delay order and feedback weights of each branch. With multiple iterations, the model gradually converges to a parameter configuration that adapts to the current state of the tested object, significantly improving the stability of the fused prediction results. Finally, a health status mapping operation is performed on the optimized fused prediction results to generate health scores and risk levels, achieving a quantitative assessment of health status.

[0074] In actual testing, multiple sets of magnetocardiogram (MCC) signal data under different conditions were selected for comparative verification. The method of this invention was compared with traditional single-scale modeling methods and multi-scale methods without reinforcement learning optimization. The results show that the method of this invention has significant advantages in signal reconstruction stability, prediction error control, and risk identification accuracy. Specifically, after adopting the method of this invention, the average fluctuation amplitude of the fused prediction results is reduced by about 32.6%, the average prediction error is reduced by about 28.4%, the accuracy of risk level determination is improved to over 94.7%, and the trend of score change over continuous time periods is smoother, effectively reflecting the process of health status changes.

[0075] Further analysis revealed that under high noise interference conditions, this invention effectively suppressed the impact of noise on the modeling results through multi-scale decomposition and gating modulation mechanisms; under conditions of drastic state changes, the reinforcement learning optimization mechanism can quickly adjust the model parameters, enabling the evaluation results to recover stability in a short time; during long-term monitoring, the cross-scale residual feedback mechanism enables information from different frequency bands to work synergistically, improving the overall modeling accuracy. The experimental data are shown in Table 1 below.

[0076] Table 1: Comparative Analysis of Cardiac Health Assessment Methods

[0077]

[0078] As shown in Table 1, the performance differences of different methods in magnetic resonance imaging (MRI) health assessment are quite significant. When using the single-scale method, the average prediction errors were 0.182 and 0.176, respectively, with fluctuation ranges of 0.265 and 0.251. The risk identification accuracy was only 78.3% and 80.5%, respectively, and the health score stability index was at the levels of 0.61 and 0.63. This indicates that the single-scale method is insufficient in expressing multi-frequency information when processing MRI signals, and the assessment results are significantly affected by noise and local fluctuations, resulting in weak output stability.

[0079] After adopting the multi-scale, non-optimized method, the average prediction error decreased to 0.142 and 0.138, the fluctuation amplitude decreased to 0.198 and 0.191, the risk identification accuracy increased to 86.7% and 88.2%, and the health score stability index improved to 0.72 and 0.74. This indicates that multi-scale decomposition can effectively separate low-frequency trend information, mid-frequency structural information, and high-frequency perturbation information in the magnetic resonance imaging (MRI) signal. Compared with the single-scale method, it has already improved the assessment accuracy and result stability to a certain extent. The data also show that although relying solely on multi-scale modeling can improve overall performance, it still has shortcomings in terms of parameter adaptive adjustment and response to complex state changes.

[0080] After adopting the method of this invention, various indicators were further improved. The average prediction errors of samples 5 to 8 were 0.102, 0.097, 0.094, and 0.099, respectively, significantly lower than the previous two methods; the fluctuation ranges were 0.135, 0.128, 0.121, and 0.130, respectively, indicating that the output results were smoother and more stable; the risk identification accuracy reached 93.6%, 94.7%, 95.1%, and 94.2%, respectively; and the health score stability index reached 0.86, 0.88, 0.89, and 0.87, respectively. It can be seen that this invention, through the joint processing of multi-scale decomposition, improved NARX network three-branch modeling, gated modulation, cross-scale residual feedback, and reinforcement learning optimization, enables the model to perform differentiated modeling for information at different scales and dynamically adjust parameters according to prediction errors, thereby reducing prediction errors, minimizing result fluctuations, and improving risk identification accuracy and score stability.

[0081] Analysis of the entire table shows that the method of this invention performs relatively consistently across the four samples, without significant fluctuations in accuracy or a marked decrease in stability. This indicates that the invention not only performs well in single evaluations but also exhibits strong consistency in continuous evaluation scenarios. The data in the table verify that this invention can solve the problems of high noise interference in the central magnetic signal, difficulty in coordinating the modeling of information from different frequency bands, and insufficient stability of evaluation results in existing technologies, demonstrating high practical value.

[0082] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A health assessment method based on a magnetocardiographic signal, characterized by, Includes the following steps: Multi-channel magnetic resonance imaging (MRCI) signal data and external physiological parameter data are collected, preprocessed, and standardized MRCI sequences are generated. Multi-scale decomposition is performed on the standardized MRCI sequences to obtain low-frequency, mid-frequency, and high-frequency MRCI components. The low-frequency, mid-frequency, and high-frequency MRCI components are then input into the three branches of the improved NARX network to generate multi-scale prediction results. Gated modulation operations are performed on external physiological parameter data to generate input modulation results. These input modulation results are then applied to the input processes of the low-frequency modeling branch, the mid-frequency modeling branch, and the high-frequency modeling branch to update the multi-scale prediction results. A cross-scale residual feedback path is constructed for the updated multi-scale prediction results, and a cross-scale coupling correction operation is performed to generate a fusion prediction result. The fusion prediction result and historical prediction error are used to construct state information, which is then input into the reinforcement learning policy network to perform policy decision operation and generate parameter adjustment actions. The parameter adjustment actions are applied to the delay order and feedback weights of the low-frequency modeling branch, the mid-frequency modeling branch and the high-frequency modeling branch. The reward value is calculated based on the health status label and the fusion prediction result, and the parameters of the reinforcement learning policy network are updated to generate an optimized fusion prediction result. Perform a health status mapping operation to generate health score results and risk level results.

2. The health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The preprocessing includes: performing bandpass filtering on the multi-channel magnetocardiogram (MCC) signal data to filter out low-frequency baseline drift and high-frequency environmental noise; calculating the local mean and subtracting the mean using a sliding window method to eliminate baseline shift; constructing a multi-channel signal matrix from the baseline-shift-eliminated MCC signal data; performing singular value decomposition to retain the principal singular components; reconstructing the multi-channel MCC signal; linearly scaling the reconstructed multi-channel MCC signal data according to the channel amplitude range; performing time alignment between channels based on key points of the cardiac cycle; and finally segmenting the time-aligned multi-channel MCC signal data according to a fixed-length sliding window to generate a standardized MCC sequence.

3. The health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The multi-scale decomposition operation on the standardized magnetocardiogram (MCC) sequence includes: detecting key points of the cardiac cycle along the time axis of the standardized MCC sequence, and dividing the periodic constraint decomposition interval by the sampling interval between adjacent key points of the cardiac cycle; setting long time windows, medium time windows, and short time windows in parallel within each periodic constraint decomposition interval, and performing multi-time window sliding decomposition on the standardized MCC sequence; performing local smoothing operation on the standardized MCC sequence using the long time window to extract low-frequency MCC components, and subtracting the low-frequency MCC components from the standardized MCC sequence to obtain the first residual sequence; and performing local structure extraction operation on the first residual sequence using the medium time window to obtain the medium-frequency MCC components. The first residual sequence is subtracted from the intermediate frequency magnetocardiogram (MCG) component to obtain the second residual sequence. A local perturbation extraction operation is performed on the second residual sequence using a short time window to obtain the high-frequency MCG component. The energy proportion within the periodic constraint decomposition interval is calculated for the low-frequency, intermediate-frequency, and high-frequency MCG components, respectively. Based on the energy proportion, the window length and sliding step size of the long, medium, and short time windows are adaptively adjusted. The adaptively adjusted low-frequency, intermediate-frequency, and high-frequency MCG components within each periodic constraint decomposition interval are concatenated in chronological order to obtain the low-frequency, intermediate-frequency, and high-frequency MCG components.

4. The health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The improved NARX network comprises three branches: a low-frequency modeling branch, a mid-frequency modeling branch, and a high-frequency modeling branch. The modeling process for each branch includes: inputting low-frequency magnetocardiogram (MCG) components into the low-frequency modeling branch; extracting the MCG components and historical output values ​​from several consecutive time steps prior to the current moment according to a preset low-frequency delay order; constructing a low-frequency feedback input sequence in chronological order; inputting this sequence into the low-frequency modeling branch to perform long-term trend modeling and outputting the low-frequency trend result; inputting the low-frequency trend result as a trend constraint into the mid-frequency modeling branch; extracting the mid-frequency MCG components and historical output values ​​from several consecutive time steps prior to the current moment according to a preset mid-frequency delay order; calculating the amplitude difference between adjacent time steps of the MCG components; constructing a mid-frequency feedback input sequence from the mid-frequency MCG components, historical output values, amplitude difference, and low-frequency trend result in chronological order; inputting this sequence into the mid-frequency modeling branch to perform waveform structure modeling and outputting the mid-frequency structure result; and inputting the mid-frequency structure result as a structure filter into the high-frequency modeling branch; extracting the low-frequency MCG components and historical output values ​​from several consecutive time steps prior to the current moment according to a preset high-frequency delay order. The high-frequency magnetic field components at each time step and the historical output values ​​at the corresponding time steps are used to calculate the rate of change between the high-frequency magnetic field components at adjacent time steps. A high-frequency feedback input sequence is constructed by sequentially combining the high-frequency magnetic field components, historical output values, rate of change, and mid-frequency structure results. This high-frequency feedback input sequence is then input into the high-frequency modeling branch to perform local perturbation modeling, outputting the high-frequency perturbation results. In the low-frequency modeling branch, a full-time writing method for historical output values ​​is used, sequentially writing the historical output values ​​of each time step into the low-frequency feedback input sequence. In the mid-frequency modeling branch, a historical output value difference-aligned writing method is used, writing the historical output values ​​of each time step into the mid-frequency feedback input sequence according to the position corresponding to the amplitude difference. In the high-frequency modeling branch, a historical output value mutation-triggered writing method is used, writing only the historical output values ​​corresponding to time steps with a rate of change exceeding a preset threshold into the high-frequency feedback input sequence. The low-frequency trend results are used as the baseline results. The portion of the mid-frequency structure results that deviates from the low-frequency trend results by more than a preset deviation threshold is written into the baseline results. The portion of the high-frequency perturbation results with a rate of change exceeding a preset threshold is written into the baseline results updated by the mid-frequency structure results, generating multi-scale prediction results.

5. A health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The process of generating the input modulation result includes: External physiological parameter data are grouped according to parameter category to obtain rhythm parameter group, amplitude parameter group, and state parameter group. Normalization is performed on each of these groups, and they are aligned with the standardized magnetocardiogram sequence according to time order. The time-aligned rhythm parameter group is input into the first gating channel to calculate the first modulation coefficient corresponding to the low-frequency modeling branch. The time-aligned amplitude parameter group is input into the second gating channel to calculate the second modulation coefficient corresponding to the mid-frequency modeling branch. The time-aligned state parameter group is input into the third gating channel to calculate the third modulation coefficient corresponding to the high-frequency modeling branch. The first, second, and third modulation coefficients are applied to the current and historical input values ​​of the low-frequency, mid-frequency, and high-frequency modeling branches, respectively, to perform weighted modulation on the input at each time step, generating the input modulation result.

6. The health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The process of updating the multi-scale prediction results includes: aligning the low-frequency, mid-frequency, and high-frequency prediction results according to time indices, using the low-frequency prediction results as the baseline sequence; calculating the difference sequence between the mid-frequency and low-frequency prediction results at each time step, and using the portion of the difference sequence whose absolute value is greater than a preset deviation threshold as the mid-frequency correction amount; superimposing the mid-frequency correction amount onto the low-frequency prediction results according to the corresponding time steps to obtain the first update sequence; calculating the rate of change of the high-frequency prediction results at adjacent time steps, and using the high-frequency prediction results at time steps with a rate of change greater than a preset change threshold as the high-frequency correction amount; superimposing the high-frequency correction amount onto the first update sequence according to the corresponding time steps to obtain the second update sequence; performing time continuity constraint processing on the second update sequence, limiting the abrupt change amplitude based on the amplitude difference between adjacent time steps, generating a smooth update sequence, and using the smooth update sequence as the updated multi-scale prediction result.

7. A health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The process of constructing a cross-scale residual feedback path and performing cross-scale coupled correction operations includes: aligning the low-frequency, mid-frequency, and high-frequency prediction results according to time indices, using the low-frequency prediction results as the baseline sequence; calculating the difference between the mid-frequency and low-frequency prediction results at each time step to obtain the mid-frequency residual sequence, and writing the mid-frequency residual sequence as a feedback quantity into the corresponding time step position of the feedback input sequence of the mid-frequency modeling branch to correct the input of the mid-frequency modeling branch in subsequent time steps; calculating the difference between the high-frequency and mid-frequency prediction results at each time step to obtain the high-frequency residual sequence, and using the high-frequency residual sequence as... The feedback quantity is written into the corresponding time step position in the feedback input sequence of the high-frequency modeling branch to correct the input of the high-frequency modeling branch in subsequent time steps; the residual amplitude and residual change rate are calculated for the intermediate frequency residual sequence and the high-frequency residual sequence respectively, and the residual sequence is weighted according to the residual amplitude and residual change rate; the weighted intermediate frequency residual sequence is superimposed on the low-frequency prediction result according to the time step to obtain the first coupled sequence, and the weighted high-frequency residual sequence is superimposed on the first coupled sequence according to the time step to obtain the second coupled sequence. The amplitude constraint processing is performed on the second coupled sequence, and the amplitude of each time step is within the preset range to generate the fusion prediction result.

8. A health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The process of generating the optimized fusion prediction result includes: The difference between the fused prediction result and the health status label at each time step is calculated to obtain the current prediction error sequence. The current prediction error sequence is concatenated with the historical prediction error sequence in chronological order to construct an error sequence. The fused prediction result and the error sequence are concatenated at the same time step to construct state information. The state information is normalized and then input into the reinforcement learning policy network. Forward computation is performed on the state information in the reinforcement learning policy network to output a parameter adjustment action vector. The parameter adjustment action vector includes the delay order adjustment amount and feedback weight adjustment amount corresponding to the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch. The delay order adjustment amount in the parameter adjustment action vector is applied to the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch. The updated delay order is generated by increasing or decreasing the delay order of each branch. The feedback weight adjustment amount in the parameter adjustment action vector is applied to the low-frequency modeling branch, mid-frequency modeling branch, and high-frequency modeling branch. The updated feedback weight is generated by increasing or decreasing the historical output weight in the feedback input sequence of each branch. The reward value is calculated based on the health status label and the fusion prediction result. The reward value is obtained by performing an inverse proportional mapping on the absolute value of the prediction error. The reward value, the current state information and the parameter adjustment action vector are input into the reinforcement learning policy network to perform parameter update operation and iteratively update the parameters of the reinforcement learning policy network. The updated delay order and feedback weights are reapplied to the low-frequency modeling branch, the mid-frequency modeling branch and the high-frequency modeling branch to recalculate the fusion prediction result and generate the optimized fusion prediction result.

9. A health assessment method based on magnetic cardiomyography signals according to claim 1, characterized in that, The process of performing the health status mapping operation includes: extracting the mean amplitude, variance amplitude, average amplitude difference between adjacent time steps, dominant frequency and spectral energy distribution from the optimized fusion prediction results; constructing a health assessment feature vector from the extracted results in a preset order; inputting the health assessment feature vector into a preset scoring function to perform weighted calculation and normalization to generate a health score result; and then comparing the health score result with a preset risk threshold range to generate a risk level result.

10. A health assessment system based on magnetic cardiac signals, applied to the health assessment method based on magnetic cardiac signals according to any one of claims 1 to 9, characterized in that, Includes the following modules: The data acquisition module is used to acquire multi-channel magnetic resonance imaging (MRI) signal data and external physiological parameter data. The preprocessing module is used to perform preprocessing on multi-channel magnetocardiogram (MCC) signal data to generate standardized MCC sequences; The multi-scale decomposition module is used to perform multi-scale decomposition operations on standardized magnetic cardiomyocyte sequences to obtain low-frequency, mid-frequency, and high-frequency magnetic cardiomyocyte components. The modeling and prediction module is used to input the low-frequency, mid-frequency, and high-frequency magnetic cardiomyocyte components into the three branches of the improved NARX network to generate multi-scale prediction results. The gated modulation module is used to perform gated modulation operations on external physiological parameter data, generate input modulation results, and apply them to the input processes of the low-frequency modeling branch, the mid-frequency modeling branch, and the high-frequency modeling branch to update the multi-scale prediction results. The residual feedback module is used to construct cross-scale residual feedback paths from the high-frequency modeling branch to the mid-frequency modeling branch and from the mid-frequency modeling branch to the low-frequency modeling branch, perform cross-scale coupling correction operations, and generate fusion prediction results. The reinforcement learning optimization module is used to construct state information by combining the fused prediction results with historical prediction errors, inputting it into the reinforcement learning policy network to perform policy decision operations, generating parameter adjustment actions and updating the parameters of the reinforcement learning policy network, and generating optimized fused prediction results. The health assessment module is used to perform health status mapping operations on the optimized fusion prediction results to generate health score results and risk level results.

Images