Non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals

Abstract

The invention discloses a non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals, which relates to the technical field of blood pressure monitoring and comprises an original signal acquisition unit, an extreme value analysis unit, a sample generation unit, a model parameter determination unit, a relation model construction unit and a blood pressure calibration unit. According to the method, the hyper-parameters are randomly valued through the model parameter determining unit, multiple parameter combinations are generated, each parameter combination is finely adjusted in two different modes, in the first mode, the hyper-parameters in the combinations are randomly selected for replacement, the real effect of replacement is reflected according to the accuracy, and the replacement efficiency is improved; the mode can avoid that all parameter combinations approach to the same numerical value in the fine tuning process, the second mode is to further refine the numerical value and improve the quality of the parameter combinations, the two modes are mutually fused to enable all hyper-parameters to quickly determine the optimal numerical value, and the prediction precision of the model is further improved.

Description

Technical Field

[0001] This invention relates to the field of blood pressure monitoring technology, specifically to a non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals. Background Technology

[0002] A non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals utilizes pulse wave signals, electrocardiogram (ECG) signals, heart sound signals, and machine learning algorithms to achieve continuous and dynamic blood pressure monitoring. It improves measurement accuracy through adaptive calibration, ensuring that the patient's true blood pressure can be continuously and clearly indicated during surgical procedures. Patent application number 201910305930.0 discloses "A cuffless, long-term continuous non-invasive blood pressure monitoring method, belonging to the field of biomedical engineering." This method includes: acquiring physiological and pathological characteristic parameters through an ECG signal and photoplethysmography (PPG) pulse wave signal monitoring subsystem connected to a computer; and non-invasive dynamic blood pressure monitoring... The model cluster and the long-term dynamic blood pressure monitoring software subsystem seamlessly integrate and process parameters to obtain blood pressure values ​​for each heartbeat and achieve continuous, dynamic blood pressure monitoring over a long period. The non-invasive dynamic blood pressure monitoring model cluster, starting from cuffless blood pressure measurement, uses the aforementioned signal measurement subsystem to monitor the patient's physiological and pathological characteristic parameters. Based on patient individual differences, the average influence value of characteristic parameters, and a machine learning method, the system possesses a self-correcting model capability for long-term continuous blood pressure measurement. This invention improves the stability and monitoring accuracy of long-term blood pressure measurement; continuous beat or long-term blood pressure measurement can lay a more effective foundation for the prevention and treatment of hypertension.

[0003] The aforementioned existing technologies have solved the problem of the inability to monitor blood pressure non-invasively for a long period of time and in a dynamic manner. However, during system operation, the signal attribute representation cannot be filtered, resulting in excessive redundancy in the overall dimensions, which reduces model efficiency. Furthermore, the system does not optimize the various hyperparameters in the blood pressure prediction model, leading to a decline in parameter quality and affecting the model's prediction accuracy. At the same time, the lack of analysis in conjunction with heart sound signals means that a single prediction result may have a large error, thus affecting measurement accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals, comprising an extreme value analysis unit and a relational model construction unit;

[0006] The original signal acquisition unit acquires the original signal, performs timestamp alignment, filtering, and noise reduction on it to obtain the processed pulse wave signal, electrocardiogram signal, and heart sound signal.

[0007] The sample generation unit analyzes the attribute representations of pulse wave signals and electrocardiogram signals, constructs multiple samples based on the attribute representations, stores the samples in the training set and the test set, uses the samples in the training set to complete the XGBoost model training, and analyzes the model accuracy through the samples in the test set.

[0008] The model parameter determination unit determines the hyperparameters of the XGBoost model. After randomly selecting values ​​for each hyperparameter, multiple parameter combinations are obtained. The accuracy of the XGBoost model corresponding to each parameter combination is calculated. A hyperparameter is randomly selected from each parameter combination, and the new value corresponding to the hyperparameter is analyzed using a parameter update algorithm. The corresponding accuracy is calculated. All parameter combinations are sorted according to their accuracy, and the best parameter combination is selected. The new value of the hyperparameter in each parameter combination is analyzed using a coefficient analysis algorithm, and the corresponding accuracy is calculated again. The operation is repeated until the loop ends, and the best parameter combination is output.

[0009] The blood pressure calibration unit acquires the current pulse wave signal, electrocardiogram signal, and heart sound signal, analyzes the corresponding blood pressure prediction value using the XGBoost model and relational model, sets weight coefficients, calculates the actual blood pressure prediction value based on the prediction value and the corresponding weight coefficients, and optimizes the corresponding parameters of the current XGBoost model and relational model using the actual blood pressure prediction value.

[0010] Preferably, the original signal acquisition unit includes a signal filtering module and a noise reduction module. The signal filtering module acquires the original pulse wave signal P, electrocardiogram signal E, and heart sound signal Q. After aligning them with timestamps, it analyzes P, E, and Q using a flattest amplitude filter to obtain the pulse wave output signal P′, electrocardiogram output signal E′, and heart sound output signal Q′. The filtered pulse wave signal P″, electrocardiogram signal E″, and heart sound signal Q″ are calculated using P, P′, E, E′, Q, and Q′, where P″ = PP′, E″ = EE′, and Q″ = QQ′. The noise reduction module uses wavelet contraction to perform noise reduction on the pulse wave signal P″, electrocardiogram signal E″, and heart sound signal Q″ to obtain the denoised pulse wave signal. electrocardiogram signal and heart sound signals

[0011] Preferably, the extreme value analysis unit includes a judgment value determination module, an extreme value filtering module, and a signal point traversal module. The judgment value determination module reads the pulse wave signal frequency, sets the length of the sliding window and the movement step size according to the signal frequency magnitude, and statistically analyzes all local minimum value sequences s corresponding to the pulse wave signal within the current sliding window. min (t1),smin (t2),...,s min (t i ),...,s min (t m ) and local maximum sequence s max (t1),s max (t2),...,s max (t i ),...,s max (t n After that, according to s min (t1),s min (t2),...,s min (t m Calculate the minimum decision value α of the current sliding window. min ,in m represents the number of local minima within the current sliding window, i represents the index, and s... max (t1),s max (t2),...,s max (t n Calculate the maximum judgment value α of the current sliding window. max ,in n represents the number of local minima within the current sliding window, and k represents the floating coefficient. If the extreme value filtering module has s... min (t i )≥α min If s exists, then remove it from the local minimum sequence. max (t i )≤α max If the signal point traversal module traverses all signal points according to the step size, it counts the local minimum and local maximum values ​​corresponding to the sliding window at different positions. Similarly, all local minimum and local maximum values ​​of the ECG signal can be obtained.

[0012] Preferably, the sample generation unit includes an attribute characterization analysis module, a blood pressure annotation module, and a set partitioning module. The attribute characterization analysis module analyzes all attribute characterizations of the pulse wave signal, electrocardiogram (ECG) signal, and their corresponding local maxima and local minima, calculates the importance of these attribute characterizations, and deletes those with importance below a threshold. The blood pressure annotation module partitions the pulse wave signal and ECG signal according to the length of a sliding window, stores the attribute characterizations of the pulse wave signal and ECG signal within the same sliding window in the same sample, and stores each sample in a set after annotating its blood pressure value. The set partitioning module... The block divides the dataset into training and testing sets according to a preset ratio. Samples from the training set are input into the XGBoost model to complete model training. The accuracy and performance coefficients of the model are analyzed using samples from the testing set. The pulse wave signal attributes include pulse wave amplitude, systolic time, diastolic time, conduction time, enhancement index, heart rate, variability, dominant frequency energy, respiratory frequency energy, energy ratio, harmonic component ratio, and approximate entropy. The electrocardiogram signal attributes include RR interval, QT interval, PR interval, QRS width, ST segment elevation, T wave amplitude, P wave amplitude, high-frequency power, low-frequency power, very low-frequency power, total power, energy ratio, and spectral entropy.

[0013] Preferably, the model parameter determination unit includes a parameter combination determination module, a coefficient calculation module, and an accuracy comparison module. The parameter combination determination module determines the hyperparameters of the XGBoost model as c1, c2, ..., c ε Where ε represents the number of hyperparameters, and c1, c2, ..., c ε The corresponding search range, according to the search range, is c1, c2, ..., c ε Randomly select values ​​and combine them to obtain multiple parameter combinations C1, C2, ..., C j ,...,C φ ,in φ represents the number of parameter combinations. C represents the j-th parameter combination. j The value of the first hyperparameter c1 in the equation. C represents the j-th parameter combination. j The ε-th hyperparameter c ε The numerical value is determined by the coefficient calculation module, which trains the XGBoost model according to each parameter combination, calculates the corresponding accuracy, and then... j Randomly select hyperparameter c i Statistical c i upper limit of search range and lower limit value according to and Calculate the first adjustment coefficient in λ1 represents the first random number, γ represents the reciprocal of the loop count, i,j represent the sequence numbers, and C is determined. j accuracy τ j According to τ j and c i The corresponding current value Calculate the second adjustment coefficient in λ2 represents the second random number, which is updated using a parameter update algorithm. and Analysis yields c i Corresponding to the new value The accuracy comparison module calculates c. i Corresponding to the new value Accuracy; if the accuracy is greater than the original accuracy, then directly... As C j The hyperparameter c in i The actual value, and vice versa. As C j The hyperparameter c in i The actual value of the parameter update algorithm is as follows:

[0014]

[0015] in, This represents the current value of the i-th hyperparameter in the j-th parameter combination. This represents the new value of the i-th hyperparameter in the j-th parameter combination. This represents the upper limit of the search range for the i-th hyperparameter. w represents the upper limit of the search range for the i-th hyperparameter. max w represents the maximum weight. min represents the minimum weight, w represents the current weight, L represents the maximum number of iterations, and l represents the current number of iterations.

[0016] Preferably, the model parameter determination unit further includes a parameter optimization module and a parameter output module. The parameter optimization module repeats the operation until all parameter combinations have been randomly selected for analysis of a hyperparameter. Then, it sorts all parameter combinations according to their accuracy, selects the best parameter combination and the top ten parameter combinations, calculates the corresponding hyperparameter mean based on the hyperparameter values ​​of each of the current ten parameter combinations, and calculates a third adjustment coefficient using the hyperparameter values ​​and the hyperparameter mean in the best parameter combination through a coefficient analysis algorithm. Based on the third adjustment coefficient, it calculates the new values ​​of the hyperparameters in each parameter combination and analyzes the accuracy corresponding to the new values. If the accuracy is higher than the original value, the new value is retained; otherwise, the new value is not retained. After the parameter output module ends the current loop, it counts the current loop count. If the loop count is lower than a set value, it enters the next loop; otherwise, it outputs the best parameter combination and uses the values ​​in the current parameter combination as the final hyperparameter values ​​of the XGBoost model.

[0017] Preferably, the relational model construction unit includes a transmission time analysis module, a velocity calculation module, an independent segment grouping module, and a parameter fitting module. The transmission time analysis module reads the heart sound signal and pulse wave signal, performs a Hilbert transform on the heart sound signal to determine its corresponding maximum peak point S1, analyzes the pulse wave origin based on the attribute characterization of the pulse wave signal, and calculates the pulse wave transmission time using the maximum peak point and the pulse wave origin. The velocity calculation module receives the pulse wave transmission length, calculates the corresponding pulse wave velocity based on the pulse wave transmission length and time, and constructs a relational model between pulse wave velocity and blood pressure value. The independent segment grouping module evenly divides the heart sound signal and pulse wave signal into multiple independent segments, sorts the independent segments according to the time axis, combines the heart sound independent segments and pulse independent segments at the same moment, and divides the independent segments into groups of five, with each group being an independent unit. The parameter fitting module adds a corresponding blood pressure value to each independent unit, randomly selects two values ​​in each independent unit for relational model parameter fitting, and uses the remaining independent segments to determine the performance coefficients of the current relational model.

[0018] Preferably, the blood pressure calibration unit includes a blood pressure prediction module and a parameter update module. The blood pressure prediction module acquires the current pulse wave signal and electrocardiogram signal, aligns the timestamps, extracts the attribute representations of the current pulse wave signal and electrocardiogram signal, and transmits them to the XGBoost model for analysis to obtain the corresponding blood pressure prediction value. After acquiring the current heart sound signal and pulse wave signal, aligning the timestamps and calculating the corresponding pulse wave velocity, it transmits them to the relational model for analysis to obtain the corresponding blood pressure prediction value. The parameter update module sets weight coefficients according to the performance coefficients of the XGBoost model and the relational model, calculates the actual blood pressure prediction value based on the prediction value and the corresponding weight coefficients, and uses the actual blood pressure prediction value to optimize the corresponding parameters of the current XGBoost model and the relational model.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. This invention uses an extreme value analysis unit to filter all local maxima and local minima corresponding to pulse wave signals and electrocardiogram signals. This reduces interference, improves the accuracy of attribute representation, enhances the reliability of signal analysis, and optimizes blood pressure prediction. Furthermore, the sample generation unit calculates the importance of all attribute representations of pulse wave signals and electrocardiogram signals and deletes attribute representations with importance below the threshold. This simplifies the overall dimension of attribute representation, improves model efficiency, avoids overfitting, and enhances generalization ability.

[0021] 2. This invention uses a model parameter determination unit to randomly select hyperparameters, generating multiple parameter combinations. Each parameter combination is fine-tuned using two different methods. The first method randomly selects and replaces hyperparameters within the combination, reflecting the true effect of the replacement based on accuracy. This method avoids all parameter combinations approaching the same value during fine-tuning. The second method further refines the values, improving the quality of the parameter combinations. The fusion of these two methods allows all hyperparameters to quickly determine their optimal values, further improving the model's prediction accuracy. The relational model construction unit analyzes pulse wave and heart sound signals to obtain pulse wave propagation velocity and constructs a relationship model between pulse wave propagation velocity and blood pressure. This relational model is used to predict blood pressure, which is then combined with the predicted blood pressure value generated by the XGBoost model. The resulting blood pressure value is more closely aligned with the actual value, demonstrating superior performance in subsequent calibration processes. Attached Figure Description

[0022] Figure 1 A schematic diagram of the overall system flow is provided for embodiments of the present invention;

[0023] Figure 2 A detailed flowchart is provided for embodiments of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1-2 The present invention provides a technical solution: a non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals, including an extreme value analysis unit and a relation model construction unit;

[0026] The raw signal acquisition unit acquires the raw signal, performs timestamp alignment, filtering, and noise reduction on it to obtain the processed pulse wave signal, electrocardiogram signal, and heart sound signal.

[0027] The sample generation unit analyzes the attribute representations of pulse wave signals and electrocardiogram signals, constructs multiple samples based on the attribute representations, stores the samples in the training set and the test set, completes the XGBoost model training using the samples in the training set, and analyzes the accuracy of the model using the samples in the test set.

[0028] The model parameter determination unit determines the hyperparameters of the XGBoost model. After randomly selecting values ​​for each hyperparameter, multiple parameter combinations are obtained. The accuracy of the XGBoost model corresponding to each parameter combination is calculated. A hyperparameter is randomly selected from each parameter combination, and the new value corresponding to the hyperparameter is analyzed using a parameter update algorithm. The corresponding accuracy is calculated. All parameter combinations are sorted according to their accuracy, and the best parameter combination is selected. The new value of the hyperparameter in each parameter combination is analyzed using a coefficient analysis algorithm, and the corresponding accuracy is calculated again. This process is repeated until the loop ends, and the best parameter combination is output.

[0029] The blood pressure calibration unit acquires the current pulse wave signal, electrocardiogram signal, and heart sound signal. It then uses the XGBoost model and relational model to analyze the corresponding predicted blood pressure value, sets weight coefficients, calculates the actual predicted blood pressure value based on the predicted value and the corresponding weight coefficients, and optimizes the corresponding parameters of the current XGBoost model and relational model using the actual predicted blood pressure value.

[0030] The raw signal acquisition unit includes a signal filtering module and a noise reduction module. The signal filtering module acquires the raw pulse wave signal P, electrocardiogram (ECG) signal E, and heart sound signal Q. After aligning them with timestamps, it analyzes P, E, and Q using a flattest amplitude filter to obtain the pulse wave output signal P′, ECG output signal E′, and heart sound output signal Q′. Using P, P′, E, E′, Q, and Q′, it calculates the filtered pulse wave signal P″, ECG signal E″, and heart sound signal Q″, where P″ = PP′, E″ = EE′, and Q″ = QQ′. The noise reduction module uses wavelet contraction to denoise the pulse wave signal P″, ECG signal E″, and heart sound signal Q″, obtaining the denoised pulse wave signal. electrocardiogram signal and heart sound signals

[0031] The extreme value analysis unit includes a judgment value determination module, an extreme value filtering module, and a signal point traversal module. After reading the pulse wave signal frequency, the judgment value determination module sets the length of the sliding window and the movement step size according to the signal frequency magnitude, and counts all local minimum value sequences s corresponding to the pulse wave signal within the current sliding window. min (t1),s min (t2),...,s min (t i ),...,s min (t m ) and local maximum sequence s max (t1),s max (t2),...,s max (t i ),...,s max (t n After that, according to s min (t1),s min (t2),...,s min (t m Calculate the minimum decision value α of the current sliding window. min ,in m represents the number of local minima within the current sliding window, i represents the index, and s... max (t1),s max (t2),...,s max (t n Calculate the maximum judgment value α of the current sliding window. max ,in n represents the number of local minima within the current sliding window, and k represents the floating coefficient. If the extreme value filtering module has s... min (t i )≥α min If s exists, then remove it from the local minimum sequence.max (t i )≤α max If the signal point traversal module traverses all signal points according to the step size, it counts the local minimum and local maximum values ​​corresponding to the sliding window at different positions. Similarly, all local minimum and local maximum values ​​of the ECG signal can be obtained.

[0032] The sample generation unit includes an attribute characterization analysis module, a blood pressure annotation module, and a set partitioning module. The attribute characterization analysis module analyzes all attribute characterizations of the pulse wave signal, electrocardiogram (ECG) signal, and their corresponding local maxima and local minima, calculates the importance of these attribute characterizations, and deletes those with importance below a threshold. The blood pressure annotation module partitions the pulse wave signal and ECG signal according to the sliding window length, storing the attribute characterizations of the pulse wave signal and ECG signal within the same sliding window as the same sample. After annotating the blood pressure value for each sample, it stores it in a set. The set partitioning module then partitions the set... The model is divided into training and testing sets according to a preset ratio. The samples in the training set are input into the XGBoost model to complete the model training. The accuracy and performance coefficients of the model are analyzed through the samples in the testing set. The pulse wave signal attributes include pulse wave amplitude, systolic time, diastolic time, conduction time, enhancement index, heart rate, variability, dominant frequency energy, respiratory frequency energy, energy ratio, harmonic component ratio, and approximate entropy. The electrocardiogram signal attributes include RR interval, QT interval, PR interval, QRS width, ST segment elevation, T wave amplitude, P wave amplitude, high frequency power, low frequency power, very low frequency power, total power, energy ratio, and spectral entropy.

[0033] The model parameter determination unit includes a parameter combination determination module, a coefficient calculation module, and an accuracy comparison module. The parameter combination determination module determines the hyperparameters of the XGBoost model as c1, c2, ..., c ε Where ε represents the number of hyperparameters, and c1, c2, ..., c ε The corresponding search range, according to the search range, is c1, c2, ..., c ε Randomly select values ​​and combine them to obtain multiple parameter combinations C1, C2, ..., C j ,...,C φ ,in φ represents the number of parameter combinations. C represents the j-th parameter combination. j The value of the first hyperparameter c1 in the equation. C represents the j-th parameter combination. j The ε-th hyperparameter c εThe numerical value is determined, and the coefficient calculation module trains the XGBoost model separately according to each parameter combination, calculates the corresponding accuracy, and then performs the calculation in C. j Randomly select hyperparameter c i Statistical c i upper limit of search range and lower limit value according to and Calculate the first adjustment coefficient in λ1 represents the first random number, γ represents the reciprocal of the loop count, i,j represent the sequence numbers, and C is determined. j accuracy τ j According to τ j and c i The corresponding current value Calculate the second adjustment coefficient in λ2 represents the second random number, which is updated using a parameter update algorithm. and Analysis yields c i Corresponding to the new value The accuracy comparison module calculates c. i Corresponding to the new value Accuracy; if the accuracy is greater than the original accuracy, then directly... As C j The hyperparameter c in i The actual value, and vice versa. As C j The hyperparameter c in i The actual value, and the specific parameter update algorithm is as follows:

[0034]

[0035] in, This represents the current value of the i-th hyperparameter in the j-th parameter combination. This represents the new value of the i-th hyperparameter in the j-th parameter combination. This represents the upper limit of the search range for the i-th hyperparameter. w represents the upper limit of the search range for the i-th hyperparameter. max w represents the maximum weight. min represents the minimum weight, w represents the current weight, L represents the maximum number of iterations, and l represents the current number of iterations;

[0036] The model parameter determination unit also includes a parameter optimization module and a parameter output module. The parameter optimization module repeats its operation until all parameter combinations have had a hyperparameter randomly selected for analysis. Then, it sorts all parameter combinations according to their accuracy, selecting the optimal parameter combination and the top ten. It calculates the mean hyperparameter value for each of the current ten parameter combinations. Using a coefficient analysis algorithm, it calculates a third adjustment coefficient based on the hyperparameter values ​​and the mean hyperparameter value within the optimal combination. Based on the third adjustment coefficient, it calculates new values ​​for the hyperparameters within each parameter combination and analyzes the accuracy corresponding to the new values. If the accuracy is higher than the original value, the new value is retained; otherwise, it is not. After the parameter output module finishes its current loop, it counts the number of iterations. If the number of iterations is lower than a set value, it proceeds to the next loop; otherwise, it outputs the optimal parameter combination. The values ​​in the current parameter combination are used as the final hyperparameter values ​​for the XGBoost model. The coefficient analysis algorithm is as follows:

[0037]

[0038] in, Let C represent the j-th new parameter combination. j Let C represent the j-th parameter combination. P This indicates the optimal combination of parameters. A1 and A2 represent the parameter combination constructed from the mean of all hyperparameters, and A1 and A2 represent two learning factors.

[0039] The relational model construction unit includes a transmission time analysis module, a velocity calculation module, an independent segment grouping module, and a parameter fitting module. The transmission time analysis module reads the heart sound signal and pulse wave signal, performs Hilbert transform on the heart sound signal to determine the corresponding maximum peak point of heart sound S1, analyzes the pulse wave origin based on the attribute characterization of the pulse wave signal, and calculates the pulse wave transmission time using the maximum peak point and the pulse wave origin. The velocity calculation module receives the pulse wave transmission length, calculates the corresponding pulse wave velocity based on the pulse wave transmission length and time, and constructs a relational model between pulse wave velocity and blood pressure value. The independent segment grouping module evenly divides the heart sound signal and pulse wave signal into multiple independent segments, sorts the independent segments according to the time axis, combines the heart sound independent segments and pulse independent segments at the same moment, and divides the independent segments into groups of five, with each group being an independent unit. The parameter fitting module adds the corresponding blood pressure value to each independent unit, randomly selects two in each independent unit for relational model parameter fitting, and uses the remaining independent segments to determine the performance coefficient of the current relational model.

[0040] The blood pressure calibration unit includes a blood pressure prediction module and a parameter update module. The blood pressure prediction module acquires the current pulse wave signal and ECG signal, aligns the timestamps, extracts the attribute representations of the current pulse wave signal and ECG signal, and transmits them to the XGBoost model for analysis to obtain the corresponding blood pressure prediction value. After acquiring the current heart sound signal and pulse wave signal, it aligns the timestamps, calculates the corresponding pulse wave velocity, and transmits it to the relational model for analysis to obtain the corresponding blood pressure prediction value. The parameter update module sets weight coefficients according to the performance coefficients of the XGBoost model and the relational model, calculates the actual blood pressure prediction value based on the predicted value and the corresponding weight coefficients, and uses the actual blood pressure prediction value to optimize the corresponding parameters of the current XGBoost model and the relational model. The relational model is as follows:

[0041]

[0042] Where z represents the artery material parameters, ρ represents the density, and d represents the artery thickness. O represents the arterial radius, F1 and F2 represent the arterial combination parameters, V represents the pulse wave velocity, and H represents the blood pressure value.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals, comprising an extreme value analysis unit and a relational model construction unit, characterized in that: The original signal acquisition unit acquires the original signal, performs timestamp alignment, filtering, and noise reduction on it to obtain the processed pulse wave signal, electrocardiogram signal, and heart sound signal. The sample generation unit analyzes the attribute representations of pulse wave signals and electrocardiogram signals, constructs multiple samples based on the attribute representations, stores the samples in the training set and the test set, uses the samples in the training set to complete the XGBoost model training, and analyzes the model accuracy through the samples in the test set. The model parameter determination unit determines the hyperparameters of the XGBoost model. After randomly selecting values ​​for each hyperparameter, multiple parameter combinations are obtained. The accuracy of the XGBoost model corresponding to each parameter combination is calculated. A hyperparameter is randomly selected from each parameter combination, and the new value corresponding to the hyperparameter is analyzed using a parameter update algorithm. The corresponding accuracy is calculated. All parameter combinations are sorted according to their accuracy, and the best parameter combination is selected. The new value of the hyperparameter in each parameter combination is analyzed using a coefficient analysis algorithm, and the corresponding accuracy is calculated again. The operation is repeated until the loop ends, and the best parameter combination is output. The blood pressure calibration unit acquires the current pulse wave signal, electrocardiogram signal, and heart sound signal, analyzes the corresponding blood pressure prediction value using the XGBoost model and relational model, sets weight coefficients, calculates the actual blood pressure prediction value based on the prediction value and the corresponding weight coefficients, and optimizes the corresponding parameters of the current XGBoost model and relational model using the actual blood pressure prediction value.

2. The non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals according to claim 1, characterized in that: The original signal acquisition unit includes a signal filtering module and a noise reduction module. The signal filtering module acquires the original pulse wave signal P, electrocardiogram signal E, and heart sound signal Q. After aligning them with timestamps, it analyzes P, E, and Q using a flattest amplitude filter to obtain the pulse wave output signal P′, electrocardiogram output signal E′, and heart sound output signal Q′. The filtered pulse wave signal P″, electrocardiogram signal E″, and heart sound signal Q″ are calculated using P, P′, E, E′, Q, and Q′, where P″ = PP′, E″ = EE′, and Q″ = QQ′. The noise reduction module uses wavelet contraction to perform noise reduction on the pulse wave signal P″, electrocardiogram signal E″, and heart sound signal Q″ to obtain the denoised pulse wave signal. electrocardiogram signal and heart sound signals 3. The non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals according to claim 1, characterized in that: The extreme value analysis unit includes a judgment value determination module, an extreme value filtering module, and a signal point traversal module. The judgment value determination module reads the pulse wave signal frequency, sets the length of the sliding window and the movement step size according to the signal frequency magnitude, and counts all local minimum value sequences s corresponding to the pulse wave signal within the current sliding window. min (t1),s min (t2),...,s min (t i ),...,s min (t m ) and local maximum sequence s max (t1),s max (t2),...,s max (t i ),...,s max (t n After that, according to s min (t1),s min (t2),...,s min (t m Calculate the minimum decision value α for the current sliding window. min ,in m represents the number of local minima within the current sliding window, i represents the index, and s... max (t1),s max (t2),...,s max (t n Calculate the maximum judgment value α of the current sliding window. max ,in n represents the number of local minima within the current sliding window, and k represents the floating coefficient. If the extreme value filtering module has s... min (t i )≥α min If s exists, then remove it from the local minimum sequence. max (t i )≤α max If the signal point traversal module traverses all signal points according to the step size, it counts the local minimum and local maximum values ​​corresponding to the sliding window at different positions. Similarly, all local minimum and local maximum values ​​of the ECG signal can be obtained.

4. The non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals according to claim 1, characterized in that: The sample generation unit includes an attribute characterization analysis module, a blood pressure annotation module, and a set partitioning module. The attribute characterization analysis module analyzes all attribute characterizations of the pulse wave signal, electrocardiogram signal, and corresponding local maxima and local minima based on the pulse wave signal, electrocardiogram signal, and their corresponding local maxima and local minima. It calculates the importance of the attribute characterizations of the pulse wave signal and electrocardiogram signal and deletes attribute characterizations with importance below a threshold. The blood pressure annotation module partitions the pulse wave signal and electrocardiogram signal according to the length of the sliding window. It stores the attribute characterizations of the pulse wave signal and electrocardiogram signal within the same sliding window in the same sample. After annotating the blood pressure value of each sample, it stores it in the set. The set partitioning module divides the set into a training set and a test set according to a preset ratio. The samples in the training set are input into the XGBoost model to complete the model training. The accuracy and performance coefficients of the model are analyzed through the samples in the test set.

5. The non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals according to claim 1, characterized in that: The model parameter determination unit includes a parameter combination determination module, a coefficient calculation module, and an accuracy comparison module. The parameter combination determination module determines the hyperparameters of the XGBoost model as c1, c2, ..., c ε Where ε represents the number of hyperparameters, and c1, c2, ..., c are set. ε The corresponding search range, according to the search range, for c1, c2, ..., c ε Randomly select values ​​and combine them to obtain multiple parameter combinations C1, C2, ..., C j ,...,C φ The coefficient calculation module trains the XGBoost model according to each parameter combination, calculates the corresponding accuracy, and then... j Randomly select hyperparameter c i Statistical c i upper limit of search range and lower limit value according to and Calculate the first adjustment coefficient in λ1 represents the first random number, γ represents the reciprocal of the loop count, i,j represent the sequence numbers, and C is determined. j accuracy τ j According to τ j and c i The corresponding current value Calculate the second adjustment coefficient in λ2 represents the second random number, which is updated using a parameter update algorithm. and Analysis yields c i Corresponding to the new value The accuracy comparison module calculates c. i Corresponding to the new value Accuracy; if the accuracy is greater than the original accuracy, then directly... As C j The hyperparameter c in i The actual value, and vice versa. As C j The hyperparameter c in i The actual value.

6. The non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals according to claim 5, characterized in that: The model parameter determination unit also includes a parameter optimization module and a parameter output module. The parameter optimization module repeats the operation until all parameter combinations have been randomly selected for analysis of a hyperparameter. Then, it sorts all parameter combinations according to their accuracy, selects the best parameter combination and the top ten parameter combinations, calculates the corresponding hyperparameter mean based on the hyperparameter values ​​of each of the current ten parameter combinations, and calculates the third adjustment coefficient using the hyperparameter values ​​and the hyperparameter mean in the best parameter combination through a coefficient analysis algorithm. Based on the third adjustment coefficient, it calculates the new values ​​of the hyperparameters in each parameter combination and analyzes the accuracy corresponding to the new values. If the accuracy is higher than the original value, the new value is retained; otherwise, the new value is not retained. After the parameter output module ends the current loop, it counts the current loop count. If the loop count is lower than a set value, it enters the next loop; otherwise, it outputs the best parameter combination and uses the values ​​in the current parameter combination as the final hyperparameter values ​​of the XGBoost model.

7. The non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals according to claim 1, characterized in that: The relational model construction unit includes a transmission time analysis module, a velocity calculation module, an independent segment grouping module, and a parameter fitting module. The transmission time analysis module reads heart sound signals and pulse wave signals, performs Hilbert transform on the heart sound signals to determine the corresponding maximum peak point of heart sound S1, analyzes the pulse wave origin based on the attribute characterization of the pulse wave signal, and calculates the pulse wave transmission time using the maximum peak point and the pulse wave origin. The velocity calculation module receives the pulse wave transmission length, calculates the corresponding pulse wave velocity based on the pulse wave transmission length and time, and constructs a relational model between pulse wave velocity and blood pressure value. The independent segment grouping module evenly divides the heart sound signals and pulse wave signals into multiple independent segments, sorts the independent segments according to the time axis, combines the heart sound independent segments and pulse independent segments at the same moment, and divides the independent segments into groups of five, with each group constituting an independent unit. The parameter fitting module adds a corresponding blood pressure value to each independent unit, randomly selects two values ​​in each independent unit for relational model parameter fitting, and uses the remaining independent segments to determine the performance coefficients of the current relational model.

8. The non-invasive blood pressure monitoring dynamic calibration system based on pulse wave signals according to claim 1, characterized in that: The blood pressure calibration unit includes a blood pressure prediction module and a parameter update module. The blood pressure prediction module acquires the current pulse wave signal and electrocardiogram signal, aligns the timestamps, extracts the attribute representations of the current pulse wave signal and electrocardiogram signal, and transmits them to the XGBoost model for analysis to obtain the corresponding blood pressure prediction value. After acquiring the current heart sound signal and pulse wave signal, aligning the timestamps, and calculating the corresponding pulse wave velocity, it transmits them to the relational model for analysis to obtain the corresponding blood pressure prediction value. The parameter update module sets weight coefficients according to the performance coefficients of the XGBoost model and the relational model, calculates the actual blood pressure prediction value based on the prediction value and the corresponding weight coefficients, and uses the actual blood pressure prediction value to optimize the corresponding parameters of the current XGBoost model and the relational model.

Images