Health status assessment method based on real-time physiological signal feature extraction

Abstract

The application discloses a health state evaluation method based on real-time physiological signal feature extraction, comprising the following steps: synchronously collecting multi-source physiological signals by using a multi-channel cooperative acquisition unit; performing real-time quality evaluation on the collected signals, calculating an optimized sampling frequency according to the evaluation result, and dynamically adjusting the sampling parameters of each signal channel according to the calculation result; performing three-level serial processing on the multi-source physiological signals to identify abnormal signals; performing multi-domain feature extraction on the multi-source physiological signals, and adopting a dynamic weight network to fuse and extract features to form a feature dataset; analyzing the physiological correlation between each feature in the feature dataset based on a physiological system knowledge graph, extracting the topological relationship between each feature through a graph neural network, and optimizing feature selection to screen out a feature subset; and inputting the feature subset into a multi-system health evaluation model for evaluation and calculation, and outputting a health state evaluation result. The application improves the accuracy, reliability and practicability of health evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical monitoring, and in particular to a health state evaluation method based on real-time physiological signal feature extraction. BACKGROUND

[0002] In recent years, with the rapid development of sensor technology, signal processing technology and artificial intelligence technology, research institutions have developed a variety of health systems based on wearable devices, which can realize continuous monitoring and collection of physiological signals and basic analysis, providing hardware foundation and data support for health state evaluation. The traditional health evaluation method relies on periodic check data and single physiological indicator monitoring, and the multi-source physiological signal based method has poor timeliness and single evaluation dimension limitation.

[0003] At present, in the signal collection channel, the traditional method adopts independent signal collection channels, lacks effective time synchronization mechanism, and makes it difficult to realize the mutual relationship between multi-source signals; in the signal quality evaluation and parameter optimization aspect, the existing technology generally adopts fixed sampling parameters and simple signal quality capture method, which cannot dynamically adjust the collection strategy according to the signal characteristics and collection environment; in the processing aspect, the existing method often separates the denoising and serious correction processing links, and lacks systematic serial processing mechanism; in the feature extraction and selection interval, the traditional method only focuses on the features of a single domain, ignores the physiological correlation between different features, and the feature selection process lacks the guidance of physiological theory.

[0004] In the aspect of health state evaluation model construction, the existing technology mostly adopts a single evaluation model, without fully considering the response and influence between physiological systems. SUMMARY

[0005] This section aims to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, abstract and title. Such simplifications or omissions cannot be used to limit the scope of the present application.

[0006] In view of the above existing problems, the present application is proposed.

[0007] To solve the above technical problems, the present application provides the following technical solutions: a multi-channel cooperative collection unit is used to synchronously collect multi-source physiological signals, the multi-source physiological signals at least including electrocardiogram signals, respiratory signals, blood oxygen signals and heart rate signals;

[0008] The collected multi-source physiological signals are subjected to real-time quality evaluation, and the optimized sampling frequency is calculated according to the evaluation results, and the sampling parameters of each signal channel are dynamically adjusted according to the calculation results;

[0009] performing three-level serial processing on the multi-source physiological signals after quality evaluation, to identify abnormal signals;

[0010] performing multi-domain feature extraction on the multi-source physiological signals, the multi-domain features including time domain features, frequency domain features and local features, and using a dynamic weight network to fuse and extract features to form a feature dataset;

[0011] analyzing the physiological correlation between each feature in the feature dataset based on a physiological system knowledge graph, extracting the topological relationship between each feature through a graph neural network, and optimizing feature selection to screen out a feature subset with significant physiological significance;

[0012] inputting the screened feature subset into a multi-system health assessment model for evaluation and calculation, and outputting a quantitative health status assessment result.

[0013] As a preferred scheme of the health status assessment method based on real-time physiological signal feature extraction, the multi-source physiological signals synchronously collected by the multi-channel cooperative acquisition unit further include body temperature signals, electromyography signals, electroencephalogram signals and electrodermal signals.

[0014] As a preferred scheme of the health status assessment method based on real-time physiological signal feature extraction, the multi-channel cooperative acquisition unit includes a signal acquisition module, a signal conditioning module and a data conversion module, wherein:

[0015] The signal acquisition module includes independent signal acquisition channels, each channel is equipped with a dedicated sensor, the signal conditioning module is equipped with independent signal conditioning circuits for each acquisition channel, and the data conversion module includes a main controller, an ADC converter and a GPS time server for time synchronization processing.

[0016] As a preferred scheme of the health status assessment method based on real-time physiological signal feature extraction, the signal acquisition channels include an electrocardiogram signal acquisition channel, a respiration signal acquisition channel, a blood oxygen signal acquisition channel, a heart rate signal acquisition channel, a body temperature signal acquisition channel, an electromyography signal acquisition channel, an electroencephalogram signal acquisition channel and an electrodermal signal acquisition channel.

[0017] The signal conditioning circuits include a preamplifier, a filter bank, a power frequency trap and a gain amplifier.

[0018] As a preferred scheme of the health status assessment method based on real-time physiological signal feature extraction, the sampling parameters of each signal channel are dynamically adjusted according to the calculation results, including:

[0019] When V score >0.1, completely accept adjustment;

[0020] When 0.05≤V score ≤0.1, part of the adjustment is accepted, indicating that the adjustment effect is relatively good, but not completely as expected;

[0021] When 0≤V score <0.05, the current parameters are maintained, indicating that the adjustment effect is not obvious, and the system can maintain the status quo;

[0022] When V score <0, the rollback mechanism is triggered, indicating that the adjustment effect is not ideal, and the system needs to roll back to the parameter state before adjustment;

[0023] Wherein, V score is the adjustment result verification score, indicating the effect of parameter adjustment.

[0024] As a preferred scheme of the health state evaluation method based on real-time physiological signal feature extraction according to the present application, the three-stage serial processing includes using an improved adaptive Kalman filter to process signal noise, using an improved wavelet transform method to correct signal baseline, and using an improved bidirectional LSTM network to identify abnormal signals.

[0025] As a preferred scheme of the health state evaluation method based on real-time physiological signal feature extraction according to the present application, the identification rule of the abnormal signal is set as:

[0026] |s(t)-μ|>k1·σ

[0027]

[0028] D pattern >k3·D threshold

[0029] Wherein, k1, k2, k3 are adaptive threshold coefficients, σ is the standard deviation of the signal, σ rate is the standard deviation of the change rate, D pattern is the morphological abnormality determination standard;

[0030] If the change of the signal exceeds its normal fluctuation range, it is a sudden abnormality;

[0031] If the speed of signal change is significantly higher or lower, it is an outlet abnormality;

[0032] If the change trend of the signal exceeds the normal change range, i.e. the trend change speed of the signal is too fast, it is a trend abnormality;

[0033] If the morphology of the signal changes significantly and exceeds the threshold D pattern , it is a morphological abnormality.

[0034] As a preferred scheme of the health state evaluation method based on real-time physiological signal feature extraction, the forming feature data set comprises:

[0035] F dataset time freq local fused

[0036] Wherein, F time is a time domain feature, F freq is a frequency domain feature, F local is a local feature, and F fused is a fusion feature.

[0037] As a preferred scheme of the health state evaluation method based on real-time physiological signal feature extraction, the screening feature subset with significant physiological significance comprises:

[0038] According to the evaluation score Q validation , a preset threshold is judged.

[0039] If Q validation > Q threshold , the current feature subset is considered effective and directly accepted.

[0040] If Q threshold > Q validation > Q min , the current feature subset is continued to be optimized.

[0041] If Q validation <Q min , the current feature subset is unqualified and needs to be reselected.

[0042] Wherein, Q threshold is the minimum required score of the feature subset after evaluation, indicating whether the feature subset meets the minimum standard, and Q min is a lower evaluation threshold, when the score is lower than the threshold, it indicates that the feature subset cannot meet any expected standard and the feature needs to be reselected.

[0043] The beneficial effects of the present application are:

[0044] 1. By using a multi-channel synchronous acquisition unit to synchronously acquire multi-source physiological signals, the accurate synchronization of multiple physiological signals is realized, the time synchronization between multi-source signals is ensured, a high-quality data basis is provided for subsequent multi-dimensional analysis, and the accuracy and reliability of health state evaluation are improved.

[0045] ​​​​​2、By introducing real-time quality evaluation and dynamic sampling parameter adjustment mechanism, the system can optimize the sampling frequency in real time according to the signal quality, and ensure the quality of signal acquisition for different signal channels dynamic sampling parameters, avoid the waste of resources caused by oversampling and information loss caused by undersampling, and through the real-time optimization of parameters, improve the adaptability and stability of the system in complex environment;

[0046] 3、By adopting three-level serial processing architecture, the system realizes systematic processing of signal denoising, key correction and anomaly detection, avoids mutual interference of processing results, and improves the accuracy and reliability of subsequent analysis through accurate identification and processing of abnormal signals;

[0047] 4、Through multi-dimensional feature extraction and dynamic weight network fusion, the system realizes comprehensive capture and effective integration of time domain, frequency domain and local features, fully excavates various feature information contained in physiological signals, and realizes adaptive fusion of different features through dynamic weight adjustment mechanism, so as to provide more comprehensive and reliable feature basis for health status evaluation;

[0048] 5、By introducing physiological system knowledge graph and graph neural network, the system realizes deep mining and optimal selection of physiological correlation between features, ensures that the selected features have clear physiological significance, and reveals the internal relationship between features through analysis of topological relationship, improves the rationality of feature selection;

[0049] 6、By constructing multi-system health evaluation model, comprehensive analysis of selected features and quantitative evaluation of health status are realized, the independent state of each physiological system is considered, and through overall optimization of the model, comprehensive evaluation of human health status is realized, which provides reliable decision basis for clinical diagnosis and health management. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor. Among them:

[0051] Figure 1 The flowchart of the health status evaluation method based on real-time physiological signal feature extraction shown in the present application. DETAILED DESCRIPTION

[0052] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments.

[0053] All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative labor should belong to the protection scope of the present application.

[0054] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in other manners different from those described herein, and those skilled in the art can make similar generalizations without departing from the spirit and scope of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0055] According to the embodiments of the present application, combined with the flow chart shown in the figure, a health state evaluation method based on real-time physiological signal feature extraction includes: Figure 1

[0056] S1, a multi-channel cooperative acquisition unit is used to synchronously acquire multi-source physiological signals, and the multi-source physiological signals at least include electrocardiogram signals, respiratory signals, blood oxygen signals and heart rate signals. It should be noted that in this step:

[0057] The multi-channel cooperative acquisition unit includes a signal acquisition module, a signal conditioning module and a data conversion module. The signal acquisition module includes independent signal acquisition channels, each channel is equipped with a dedicated sensor. The signal conditioning module is equipped with an independent signal conditioning circuit for each acquisition channel. The data conversion module includes a main controller, an ADC converter and a GPS time server, which is used for time synchronization processing;

[0058] As an example, the signal acquisition channels include an electrocardiogram signal acquisition channel (such as an Ag / AgCl dry electrode), a respiratory signal acquisition channel (such as a piezoresistive breathing belt sensor), a blood oxygen signal acquisition channel (such as a dual-wavelength reflective photoelectric sensor), a heart rate signal acquisition channel (such as a photoelectric pulse wave sensor), a body temperature signal acquisition channel (such as a medical NTC thermistor), an electromyography signal acquisition channel (such as a surface electromyography bipolar electrode), an electroencephalogram signal acquisition channel (such as a dry brain electrode) and a skin electricity signal acquisition channel (such as an Ag / AgCl electrode);

[0059] As an example, the equipped signal conditioning circuit includes a preamplifier (such as a low-noise instrument amplifier), a filter bank, a power frequency trap (such as a double-T trap circuit) and a gain amplifier;

[0060] As an example, the time synchronization processing is realized by using multiple time synchronization algorithms, as shown in the following formula:

[0061] ​

[0062] wherein, T sync (i) is the synchronization time of the i-th channel, T base is the system reference time, ΔT offset (i) is the i-th channel deviation correction value, K adjust is the adjustment coefficient, the value range is 0.8-1.2, is the phase compensation factor of the i-th channel;

[0063] Further, the collected multi-source physiological signals also include body temperature signals, electromyography signals, electroencephalogram signals and electrodermal signals.

[0064] S2, real-time quality assessment is performed on the collected multi-source physiological signals, and an optimized sampling frequency is calculated according to the assessment result, and meanwhile, the sampling parameters of each signal channel are dynamically adjusted according to the calculation result. It should be noted that the present step is as follows:

[0065] By comprehensively analyzing the defects, signal-to-noise ratio and stability of the signals, comprehensive assessment of the signal quality is realized. Specifically, the input multi-source physiological signals are processed in segments, and the sliding window method is used to check the data integrity of the signals to ensure the integrity and quality of each small segment of data. According to the type of the signals, different window lengths are set, for example, a 2-second window is used for electrocardiogram signals, a 5-second window is used for respiration signals, and a 3-second window is used for blood oxygen and heart rate signals.

[0066] In each window, a defect index is calculated, which comprehensively considers the data loss rate, signal change rate and signal standard deviation, and adjusts the coefficient according to the type of the signal, that is, the defect index directly affects the subsequent signal-to-noise ratio evaluation weight distribution;

[0067] For example, the mathematical expression formula is as follows:

[0068]

[0069] wherein, CI(t) is the defect index, representing the signal defect degree at a certain time, the value range is [0, 1], N d is the number of detected data points, N is the total sampling point number in the evaluation window, dS / dt represents the signal change rate, which is calculated by the central integral method, reflecting the change degree of the signal over time, σ is the signal standard deviation, which is obtained by the sliding estimation method, reflecting the fluctuation degree of the signal, K type is the signal type adjustment coefficient, which is used to adjust the calculation of the defect index according to different signal types;

[0070] Further, according to the calculated defect index, the weight of the signal-to-noise ratio (SNR) evaluation is dynamically adjusted, the influence of the defect index on the signal quality is quantitatively evaluated, and the weight distribution in the signal-to-noise ratio evaluation is adjusted, wherein:

[0071] When CI(t)>0.95, the signal-to-noise ratio evaluation weight is increased by 20%;

[0072] When 0.85≤CI(t)≤0.95, the original weight is maintained;

[0073] When CI(t)<0.85, the signal-to-noise ratio evaluation weight is reduced by 15%;

[0074] The dynamic evaluation of the signal-to-noise ratio is based on the above defect evaluation results, and an improved wavelet level method is used to calculate the signal-to-noise ratio. The evaluation process includes signal level, energy estimation and dynamic calibration, wherein:

[0075] The dynamic evaluation of the signal-to-noise ratio is based on the wavelet decomposition method to separate the signal and the noise, calculate the signal-to-noise ratio, and optimize the signal quality through dynamic correction. The calculation formula is:

[0076]

[0077] Wherein, SNR(t) is the signal-to-noise ratio, which represents the index of signal quality, the larger the better, P s is the signal power, which is obtained by wavelet reconstruction, representing the energy size of the signal, P n is the noise power, which is estimated by the high-frequency sub-band coefficient, representing the energy size of the noise, W CI (t) is the threat evaluation weight factor, which is positively correlated with the defect index CI(t), representing the influence of the defect index on the signal-to-noise ratio evaluation, K adjust (t) is an adaptive adjustment coefficient, which is dynamically adjusted according to the signal change, used to adjust the accuracy of the signal-to-noise ratio calculation, and its calculation method is:

[0078]

[0079] Wherein, α is the adjustment factor, representing the sensitivity of the adjustment coefficient (initial value is 0.3), represents the change rate of the signal-to-noise ratio, reflecting the change of the signal quality with time, S factor is the signal factor, which is an adjustment factor set according to the signal type (such as electrocardiogram, respiration, etc.), used to adjust the calculation of the signal-to-noise ratio.

[0080] It should be noted that the signal quality to noise ratio evaluation result is used to update the signal state matrix, which records the quality change trend of each signal channel at different time scales. For different types of physiological signals, different signal channels are used for calibration to adapt to the quality change trend of the group signal, providing a reference for subsequent sampling frequency optimization.

[0081] Further, after completing the signal to noise ratio evaluation, a multi-reading cursor analysis method is used to evaluate the stability of the signal. Through a sliding window mechanism, the signal complexity at different time readings is calculated in real time, and the evaluation formula is:

[0082] MSE(t) = -∑ i p(i)·log2p(i)·W(scale)·F adapt (t)

[0083] Wherein, MSE(t) is the signal complexity (mean square error), reflecting the stability of the signal, p(i) is the probability distribution of the signal at different sensitivities, which is obtained by kernel density estimation, indicating the probability of signal distribution, W(scale) is the measurement weight function, which adopts exponential decay form, indicating the signal weight at different scales, F adapt (t) is an adaptive adjustment function, which is dynamically adjusted according to the stability of the signal, and the calculation formula is:

[0084]

[0085] Wherein, β is the stability adjustment factor, indicating the sensitivity of signal stability adjustment (range [0.1, 0.4]), SNR ref is the reference signal to noise ratio, which is determined according to the historical data of the signal, and is used to evaluate the signal quality, τ is the time constant, which is adaptively adjusted according to the signal characteristics, and is used to control the time influence of signal stability, t is the current time, which is used to track the signal change.

[0086] In an optional embodiment, the evaluation results of the signal deviation, signal to noise ratio and signal stability are comprehensively considered, the weight factor is updated in real time through a dynamic adjustment mechanism, and when the system resource occupancy rate exceeds a preset threshold, the weight update rate is reduced. An adaptive weight fusion method is used to calculate the final quality score:

[0087]

[0088] Wherein, Q score (t) is the signal quality score, which comprehensively evaluates the quality of the signal and reflects the quality of each dimension, w1, w2, w3 are dynamic weight coefficients, respectively indicating the importance of defect index, signal to noise ratio and stability in the total score, CI(t) is the defect index, indicating the degree of signal defect, SNR(t) is the signal to noise ratio, indicating the signal quality, SNRmax MSE(t) is the best quality criterion of signal, MSE(t) is the signal stability, which represents the complexity of signal, MSE(t) is the reference mean square error, which is used to standardize the stability evaluation results, G(t) is the global adjustment factor, which considers the factors of time decay and quality change, and the calculation formula is: ref factor

[0089]

[0090] wherein δ is the change rate sensitivity coefficient, which represents the influence degree of quality score change on the global adjustment factor, and δ is the quality score change rate, which represents the change of quality score with time, and μ is the time decay coefficient, which represents the influence of time on the quality score.

[0091] In an optional embodiment, the sampling parameter optimization process is based on the comprehensive quality score, and the dynamic optimization considering the sampling parameters is implemented, including:

[0092] The basic sampling frequency is set according to the characteristic probability of different physiological signals;

[0093] The actual sampling frequency is dynamically adjusted through an adaptive adjustment algorithm, and the adjustment process considers the signal rate, the quality score change trend and the system resource status;

[0094] The change of sampling frequency meets the stability constraint, that is, the frequency change in the adjacent adjustment interval shall not exceed 20% of the basic sampling frequency;

[0095] For example, the basic sampling frequency configuration system sets the basic sampling frequency according to the characteristic probability of different physiological signals, such as 1000 Hz for electrocardiogram signal, 100 Hz for respiratory signal, 100 Hz for blood oxygen signal and 200 Hz for myocardial signal;

[0096] Specifically, the optimization of sampling frequency needs to be dynamically adjusted based on the signal quality score, and the sampling frequency formula is:

[0097] f s,opt (t)=f s,base· F adjust (t)·[1+λ·S complexity (t)]·K stability

[0098] wherein f s,opt (t) is the optimized sampling frequency, which is dynamically adjusted based on the signal quality evaluation and system state, f s,base is the basic sampling frequency, which is set based on the characteristics of signal and Shannon sampling theorem, and F adjust ​​​(t) is a dynamic adjustment factor, considering signal quality, system resources and real-time requirements, λ is a complexity adjustment coefficient (range [0.1, 0.5]), representing the influence of signal complexity in sampling frequency optimization, K stability is a stability guarantee factor, representing the stability of the optimized sampling frequency, preventing the frequency from changing too drastically, S complexity (t) is a signal complexity index, representing the complexity of the signal, and the calculation formula is:

[0099]

[0100] where H entropy is the signal entropy value, representing the uncertainty of the signal, H max is the preset entropy value, used for standardizing the signal complexity, and ρ is the acceleration sensitivity coefficient, representing the influence of signal acceleration change on sampling frequency optimization, is the acceleration of the signal, representing the acceleration of signal change;

[0101] In an optional embodiment, a gradual parameter adjustment method is used for parameter updating, ensuring smooth transition of the system:

[0102] P new (t) = P old (t) + ΔP(t)·η(t)·M smooth (t)

[0103] where P new (t) is the updated parameter value, P old (t) is the current parameter value, ΔP(t) is the parameter adjustment amount, η(t) is the adaptive learning rate, representing the sensitivity of parameter updating, and M smooth (t) is the smoothing factor;

[0104] Further, through sampling frequency change constraint, it is ensured that the sampling frequency changes within a reasonable range, and the formula is:

[0105]

[0106] where α max is the maximum change rate, usually taking a value of 0.2;

[0107] Through quality constraint, it is ensured that the quality score is not lower than the preset threshold, and the formula is:

[0108] Q score,new ≥ Q threshold · Q score,old

[0109] where Q score,new is the current signal quality score, and Q score,oldQ threshold is a preset quality score threshold value;

[0110] By resource utilization constraints, the system resources are guaranteed not to exceed the maximum resource occupation, and the formula is:

[0111] R total = w1·R cpu + w2·R mem + w3·R buffer ≤ R max

[0112] wherein R total is the total resource occupation rate, R cpu , R mem , R buffer respectively represent the resource occupation of CPU, memory and buffer, R max is the maximum allowed resource occupation, set to 0.8;

[0113] For example, the sliding window verification method is used to verify the effect of the adjustment result, and the mathematical expression formula is

[0114]

[0115] wherein V score represents the adjustment result verification score, represents the effect of parameter adjustment, Q new represents the quality score after adjustment, Q old represents the quality score before adjustment, R cost is the resource consumption increment, representing the increase proportion of system resource usage in the adjustment process, E stability is a system stability evaluation function, representing the stability of the system after parameter adjustment, and the calculation formula is:

[0116]

[0117] wherein σ params is the standard deviation of parameter change, reflecting the amplitude and fluctuation of parameter adjustment, σ ref is a reference threshold value, representing the maximum allowed parameter change amplitude, set as a constant.

[0118] Preferably, the sliding window method is used in this step to check the parameter adjustment effect at different time points in real time, and the change of quality score before and after adjustment is continuously observed to evaluate whether the adjustment is effective. Considering the change of quality score and resource consumption, the sliding window method can effectively avoid the influence of noise in a short time, so as to obtain a smooth evaluation result.

[0119] Preferably, V score is used to determine whether to accept parameter adjustment, and for different Vscore wherein:

[0120] when V score >0.1, fully accept the adjustment;

[0121] when 0.05≤V score ≤0.1, partially accept the adjustment, indicating that the adjustment effect is relatively good, but not completely as expected;

[0122] when 0≤V core <0.05, maintain the current parameters, indicating that the adjustment effect is not obvious, and the system can maintain the status quo;

[0123] when V score <0, trigger the rollback mechanism, indicating that the adjustment effect is not ideal, and the system needs to roll back to the parameter state before adjustment.

[0124] S3, performing three-level serial processing on the multi-source physiological signals after quality evaluation to identify abnormal signals. It should be noted that in this step:

[0125] The three-level serial processing includes using an improved adaptive Kalman filter for signal noise, using an improved wavelet transform method for signal baseline correction, and using an improved bidirectional LSTM (BiLSTM) network to identify abnormal signals.

[0126] For example, in the first level, an improved adaptive Kalman filter is used for signal denoising, and the filter parameters are dynamically adjusted according to the signal characteristics to achieve optimal denoising effect, wherein:

[0127] A state space model is constructed:

[0128] X(k+1)=Φ(k)X(k)+G(k)W(k)

[0129] Z(k)=H(k)X(k)+V(k)

[0130] wherein X(k) is the state vector of the signal at time k, Φ(k) is the state transition vector, G(k) is the system noise vector, W(k) is the system noise, Z(k) is the observation vector, i.e. the actual measured signal, H(k) represents the conversion quantity that describes the state transition to the observation value, and V(k) is the initial noise;

[0131] Then, the statistical characteristics of the signal and noise are estimated using the sliding window method, and the Kalman gain is calculated, and the state update is:

[0132] P(k|k)=[1-K(k)H(k)]P(k|k-1)

[0133] Wherein, P(k|k) is the updated signal state estimation, K(k) is the Kalman gain, indicating how to correct the current predicted signal, P(k|k-1) is the updated prediction covariance matrix.

[0134] Exemplarily, in the second stage, the signal is baseline corrected by using the improved wavelet transform method, and accurate calibration is realized through multi-scale analysis, including:

[0135] Select the optimal wavelet basis:

[0136] Ψ opt = argmin{||s-sΨ i || 2}

[0137] Wherein, Ψ i is the candidate wavelet basis function, S is the original signal, sΨ i is the signal reconstructed by using Ψ i ;

[0138] Multi-dimensional wavelet transform:

[0139]

[0140] Wherein, a is the measurement scale, b is the translation parameter, and J is the maximum number of layers;

[0141] Baseline is identified by using threshold method:

[0142] D(j,k) = |d(j,k)| < T(j)·σ(j)

[0143] Wherein, d(j,k) is the wavelet coefficient of the jth layer and the kth position, T(j) is the adaptive threshold, and σ(j) is the noise standard deviation of the jth layer;

[0144] Intensity correction:

[0145] s corrected (t) = s(t)-B(t)

[0146]

[0147] Wherein, B(t) is the estimated baseline drift, c J,k is the approximation coefficient, d j,k is the detail coefficient, is the scaling function, Ψ j,k (t) is the wavelet function.

[0148] As an example, in the third stage, the signal anomaly is identified and processed by using an improved bidirectional LSTM (BiLSTM) network, and the improved bidirectional LSTM (BiLSTM) network structure includes an input layer, a BiLSTM layer, an attention layer, an output layer, and a feature extraction layer, which is implemented by a multi-dimensional feature extractor and is used to extract time domain, frequency domain and statistical features;

[0149] For example, the mathematical expression formula of the feature extraction layer is

[0150] F(t) = [f time (t), f freq (t), f stat (t)]

[0151] Where f time (t) is the time domain feature, f freq (t) is the frequency domain feature, and f stat (t) is the statistical feature.

[0152] The abnormality discrimination rule is set as:

[0153] |s(t)-μ|>k1·σ

[0154]

[0155] D pattern >k3·D threshold

[0156] Where k1, k2, and k3 are adaptive threshold coefficients, σ is the signal standard deviation, σ rate is the standard deviation of the change rate, and D pattern is the morphological abnormality determination standard.

[0157] The abnormality processing strategy includes:

[0158] Abrupt abnormality: smoothed by using median filtering;

[0159] Export abnormality: reconstructed by using interpolation method;

[0160] Trend abnormality: repaired by using polynomial fitting;

[0161] Morphological abnormality: repaired by using template matching;

[0162] Processing effect verification:

[0163] The abnormality processing effect is evaluated by a multi-index evaluation system:

[0164] Quality Score = w1·SNR + w2·RMSE + w3·Correlation

[0165] wherein SNR is the signal to noise ratio improvement, RMSE is the root mean square error, Correlation is the signal correlation, and Quality is the quality of the signal, when Quality is greater than a set threshold value, the processing result is accepted, otherwise an artificial review mechanism is triggered. Score

[0166] S4, multi-domain feature extraction is performed on the multi-source physiological signals again, the multi-domain features include time domain features, frequency domain features and local features, and a dynamic weight network is used for feature fusion and extraction to form a feature data set. It should be noted that in this step that:

[0167] The time domain feature extraction adopts a hierarchical progressive processing strategy, and the features are extracted from statistical features, waveform features to dynamic features, wherein:

[0168] The statistical features include time-varying mean, standard deviation, skewness and kurtosis, and the adaptive window method is used for calculation, and the window length is dynamically adjusted according to the signal characteristics;

[0169] The waveform features include maximum amplitude, feature interval, waveform slope and waveform ratio;

[0170] The dynamic features include trend features, periodic features and complexity features.

[0171] The domain feature extraction adopts a multi-resolution analysis method, and the accurate frequency extraction of the features is realized in combination with the time-frequency joint representation, wherein:

[0172] The short-time Fourier transform in the frequency domain is improved, and the transform length is adaptively adjusted according to the signal variation rate;

[0173] Then, multi-layer wavelet packet decomposition is performed, and the energy of each frequency band is calculated;

[0174] A noise feature set is extracted, including power spectral density, main frequency feature, spectral width feature and harmonic ratio.

[0175] The local feature extraction analyzes the local morphology and change characteristics of the signal, wherein:

[0176] A multi-scale morphological algorithm is used to extract morphological features, including opening damage features, closing damage features and morphological gradients;

[0177] According to the signal feature structure element size, local singularity analysis is performed, and a local singularity index is calculated;

[0178] Then, a local pattern feature is extracted through an LBP (Local Binary Pattern) algorithm.

[0179] Further, an improved attention mechanism is used to realize feature fusion, including:

[0180] ​The feature evaluation function is constructed by comprehensively considering the correlation, redundancy and stability of the features:

[0181] I(f) = w1·R relevance + w2·R redundancy + w3·R stability

[0182] wherein R relevance is the feature correlation, R redundancy is the feature redundancy, and R stability is the feature stability;

[0183] The attention weight is calculated using the multi-head attention mechanism, and the features are fused to obtain the weighted fusion features F fused

[0184] Finally, the feature dataset is formed:

[0185] F dataset = {F time , F freq , F local , F fused}

[0186] wherein F time is the time domain feature, F freq is the frequency domain feature, F local is the local feature, and F fused is the fusion feature.

[0187] S5, based on the physiological system knowledge graph, the physiological correlation between each feature in the feature dataset is analyzed, the topological relationship between each feature is extracted through the graph neural network, and the feature selection is optimized to screen out a feature subset with significant physiological significance. It should be noted that this step is:

[0188] The knowledge graph is represented by a directed graph, wherein the node represents the feature, the edge represents the correlation between the features, and the structure of the graph is defined by the feature node set, the feature correlation edge set and the correlation weight matrix:

[0189] G = (V, E, W)

[0190] wherein V is the feature node set, representing the node of each feature, E is the feature correlation edge set, representing the correlation between the features, and W is the correlation weight matrix, representing the correlation weight between each pair of features;

[0191] For example, the graph adopts a hierarchical structure and is divided into four layers:

[0192] Basic feature layer: directly extracting original features from signals;

[0193] Combined feature layer: combination of multiple basic features, reflecting more complex signal features;

[0194] Systemic characteristic layer: Characteristics that reflect the state of a specific physiological system (such as cardiovascular, respiratory, etc.);

[0195] Integrated Feature Layer: Comprehensive features across systems, combining relevant features from multiple systems.

[0196] The strength of the association between each layer is calculated using the following formula, reflecting the strength of the relationship between each level:

[0197]

[0198] Where, β base The basic correlation strength is set to 0.8, d ij The distance between levels represents the hierarchical relationship between different levels. θ is the attenuation coefficient, controlling the influence of the inter-level distance on the association strength, and is set to 2.0. mutual Mutual information content measures the degree to which information is shared between features;

[0199] Then, the feature association strength is calculated using a multi-factor fusion method:

[0200] w ij =w1·C pearson +w2·C mutual +w3·C physio

[0201] Among them, C pearson C is the Pearson correlation coefficient, which measures the linear correlation between features. mutual To normalize mutual information and measure the nonlinear correlation between features, C physio Physiological correlation degree reflects the strength of the physiological relationship between different features in a physiological system.

[0202] It should be further explained that this embodiment adopts an improved graph attention network (GAT) structure to extract the topological relationship between features for deep analysis. The graph neural network structure includes a feature transformation layer (used to perform linear transformation on node features), a focus calculation layer (calculating the relative importance between each feature), and a message aggregation layer (aggregating messages from neighboring nodes and updating the node representation). That is, multi-head attention calculation is used to calculate the relationship between nodes through focus coefficients, and the final node representation is obtained through the aggregation of multi-head attention.

[0203] In an optional implementation, feature optimization selection, i.e., feature filtering, is performed based on the analysis results of the graph neural network. This is done through a strategy iteration method, including:

[0204] Set an empty set of selected features and a set of candidate features;

[0205] In each iteration, the top scored feature is selected from the candidate feature set;

[0206] The selected feature is verified, including strength verification, redundancy verification and stability verification;

[0207] The verified feature is added to the selected feature set and removed from the candidate set, and the scoring of the remaining features is continuously updated;

[0208] The final feature set contains time domain features, frequency domain features, local features and fusion features, and is formed by weighted aggregation:

[0209] F=[w t ·F time ;w f ·F freq ;w l ·F local ;w c ·F fusion ]·C norm

[0210] Where w t , w f , w l , w c are the domain weights of time domain features, frequency domain features, local features and fusion features respectively, and G norm is a global normalization factor to balance the overall quality score and complexity;

[0211] The cross-validation is used to evaluate the quality of the feature subset:

[0212] Q validation =ω1·Accuracy+ω2·Stability+ω3·Efficiency

[0213] Where Q validation is the evaluation score obtained by cross-validation, representing the overall quality of the feature subset, Accuracy is the accuracy of the feature subset in cross-validation, Stability represents the consistency of the feature subset on different datasets, and Efficiency represents the efficiency of the feature subset in terms of computational resource consumption or training time;

[0214] According to the evaluation score Q validation and the preset threshold, including:

[0215] Accept the current feature subset: if Q validation >Q threshold , the feature subset is considered effective and can be directly accepted;

[0216] Continue optimization: if Q threshold >Qvalidation >Q min , then continue to optimize the current feature subset;

[0217] Re-select features: if Q validation <Q min , the current feature subset is not qualified and needs to re-select features;

[0218] wherein Q threshold is the minimum required score after the feature subset evaluation, indicating whether the feature subset meets the minimum standard, Q min is a lower evaluation threshold, when the score is lower than this value, it means that the feature subset cannot meet any expected standard and needs to re-select features.

[0219] S6, input the screened feature subset into the multi-system health assessment model for evaluation and calculation, and output the quantitative health status assessment result. It needs to be explained that the multi-system health assessment model includes feature sparsity layer, single system evaluation layer, inter-system good evaluation layer, comprehensive state evaluation layer and risk warning layer, and each layer network framework executes the evaluation task in turn, wherein:

[0220] In the feature sparsity layer, the input feature subset is standardized, including dynamic range calculation, normalization and feature compensation of the feature;

[0221] In the single system evaluation layer, the state of each physiological system is evaluated, such as cardiovascular system evaluation (combining electrocardiogram features, blood pressure features and myocardial features to evaluate the cardiovascular system), respiratory system evaluation (combining respiratory rate, respiratory depth and blood oxygen permeability level to evaluate the health status of the respiratory system), energy consumption and heart rate regulation evaluation (analyzing energy consumption and heart rate regulation features to evaluate the state of the overall physiological system);

[0222] As an example, the basic score calculation formula is:

[0223] Score i =∑ j w ij ·f ij

[0224] wherein w ij is the weight of the feature, and f ij is the feature after standardization.

[0225] In the inter-system good evaluation layer, the mutual influence and synergistic effect between each physiological system are analyzed, wherein:

[0226] The influence degree of different systems is determined by calculating the connection degree between the systems, and the calculation formula is:

[0227]

[0228] wherein C ij represents the connection degree between system i and j, N is the number of systems, and represents the relationship between all systems;

[0229] Cooperation score calculation:

[0230] CooperationScore =∑ i,j C ij ·S i ·S j ·G coupling

[0231] wherein S i and S j are the base scores of system i and j, and G coupling is the system coupling coefficient, reflecting the degree of synergy between systems.

[0232] In the comprehensive state evaluation layer, the evaluation results of each system are fused to generate the overall health state score, and the calculation process is as follows:

[0233] Final Score = (w base ·Base Score + w syn ·Cooperation Scire )·F dynamic

[0234] wherein w base and w syn are the weights of the base score and the cooperation score, and F dynamic is a dynamic adjustment factor that adjusts the timeliness and real-time of the score.

[0235] In the risk warning layer, health risk analysis and warning are performed according to the evaluation results, multi-level warning thresholds are set to identify abnormal health status, and multiple factors (such as health status, system score, and cooperation score) are considered to assess the risk level, analyze the trend of health status, and predict short-term and long-term changes in health status.

[0236] Exemplarily, the final output of the health state evaluation result includes:

[0237] Quantitative score indicators: overall health score, system function score, and system coordination score;

[0238] Health level: excellent, good, general, attention needed, and intervention needed, five levels;

[0239] Health risk level: low risk, medium risk, and high risk, reflecting the potential risk of health status;

[0240] Dynamic change indicators: short-term change trend, long-term change pattern, volatility indicators;

[0241] Suggestions: health maintenance suggestions, risk prevention measures.

[0242] Preferably, the embodiment realizes comprehensive analysis of screening features and quantitative assessment of health status by constructing a multi-system health assessment model, considers the independent state of each physiological system, and realizes comprehensive assessment of the health status of the human body through overall optimization of the model, thereby providing a reliable decision basis for clinical diagnosis and health management.

[0243] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A health state assessment method based on real-time physiological signal feature extraction, characterized in that, Comprise: Synchronous acquisition of multi-source physiological signals by a multi-channel cooperative acquisition unit, the multi-source physiological signals at least including electrocardiogram signals, respiratory signals, blood oxygen signals and heart rate signals; Real-time quality assessment of the collected multi-source physiological signals, and calculation of optimized sampling frequency according to the assessment results, and dynamic adjustment of the sampling parameters of each signal channel according to the calculation results: When V score >0.1, fully accept adjustment; When 0.05≤V score ≤0.1, part of the receiving is adjusted, indicating that the adjustment effect is relatively good, but it does not fully meet the expectations; When 0≤V score When 0.05 < V < 0.1, the current parameters are maintained, indicating that the adjustment effect is not obvious, and the system can maintain the status quo. When V score <0, trigger the fallback mechanism, indicating that the adjustment effect is not ideal, and the system needs to fall back to the parameter state before adjustment; V score is the result validation score, indicating the effect of the parameter adjustment; Three-level serial processing of the multi-source physiological signals after quality assessment to identify abnormal signals; The three-level serial processing includes using an improved adaptive Kalman filter to process signal noise, using an improved wavelet transform method to correct signal baseline, and using an improved bidirectional LSTM network to identify abnormal signals; The identification rule of the abnormal signals is set as: If the change of the signal exceeds its normal fluctuation range, it is a sudden abnormality; D pattern >k3·D threshold Wherein, k1, k2, k3 are adaptive threshold coefficients, σ is signal standard deviation, σ rate is the standard deviation of the rate of change, D pattern is the morphological abnormality determination standard; If the speed of signal change is significantly higher or lower, it is an outlet abnormality; If the change trend of the signal exceeds the normal change range, i.e. the trend change speed of the signal is too fast, it is a trend abnormality; Multi-domain feature extraction is performed on the multi-source physiological signals, the multi-domain features include time domain features, frequency domain features and local features, and dynamic weight network is used to fuse and extract features to form a feature data set; If the shape of the signal changes significantly and exceeds the threshold D pattern then the shape is abnormal; Based on the physiological system knowledge graph, the physiological correlation between each feature in the feature data set is analyzed, the topological relationship between each feature is extracted by a graph neural network, and the feature selection is optimized to screen out a feature subset with significant physiological significance: The screened feature subset is input into a multi-system health assessment model for evaluation and calculation, and a quantitative health status assessment result is output: According to the evaluation score Q validation a determination is made with respect to a preset threshold value; If Q validation >Q threshold , the current feature subset is considered valid and accepted directly; If Q threshold >Q validation >Q min then continue optimizing the current feature subset; If Q validation <Q min , the current feature subset is unqualified and needs to be reselected. where Q threshold is the minimum required score after evaluation, indicating whether the feature subset meets the minimum standard, Q min is a lower evaluation threshold, indicating that the feature subset fails to meet any expected standard and needs to be reselected when the score is lower than this value. Collaborative effect score calculation: In the comprehensive state assessment layer, the evaluation results of each system are fused to generate the overall health status score, and the calculation process is as follows: CooperationScore =∑ i,j C ij ·S i ·S j ·G coupling where S i , S j are the base scores of systems i and j, G coupling is the system coupling coefficient, reflecting the degree of synergy between systems, C ij represents the connection degree between systems i and j; The multi-source physiological signals acquired by the multi-channel cooperative acquisition unit also include body temperature signals, electromyography signals, electroencephalogram signals and electrodermal signals. Final Score = (w base ·Base Score + w syn ·Cooperation Score )·F dynamic wherein w base and w syn are the weights of the base score and the synergy score, F dynamic is a dynamic adjustment factor that adjusts the timeliness and real-time of the score.

2. The health state evaluation method based on real-time physiological signal feature extraction according to claim 1, characterized in that, The multi-channel cooperative acquisition unit includes a signal acquisition module, a signal conditioning module and a data conversion module, wherein:

3. The health state assessment method based on real-time physiological signal feature extraction according to claim 2, characterized in that, The signal acquisition module includes independent signal acquisition channels, each channel is equipped with a dedicated sensor, the signal conditioning module is equipped with independent signal conditioning circuits for each acquisition channel, and the data conversion module includes a main controller, an ADC converter and a GPS time server for time synchronization processing. The signal acquisition channels include electrocardiogram signal acquisition channels, respiratory signal acquisition channels, blood oxygen signal acquisition channels, heart rate signal acquisition channels, body temperature signal acquisition channels, electromyography signal acquisition channels, electroencephalogram signal acquisition channels and electrodermal signal acquisition channels; 4. The health state evaluation method based on real-time physiological signal feature extraction according to claim 3, characterized in that, The signal conditioning circuits include preamplifiers, filter banks, power frequency traps and gain amplifiers. The formation of the feature data set includes:

5. The health state assessment method based on real-time physiological signal feature extraction according to claim 1, characterized in that, ​ F dataset = {F time , F freq , F local , F fused} wherein F time is a time domain feature, F freq is a frequency domain feature, F local is a local feature, and F fused is a fusion feature.

Images