Dual-mode blood pressure calculation method, device, equipment and storage medium
Through the dual-mode blood pressure calculation method, the multi-source signal is synchronized sampling and preprocessed, and combined with deep neural network, a comprehensive analysis of blood pressure measurement and atrial fibrillation risk assessment is achieved, solving the problems of inaccurate blood pressure measurement and independent detection in the existing technology. Improves measurement accuracy and anti-interference ability.
Patent Information
- Application Number
- CN202411099461.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-12
AI Technical Summary
The existing blood pressure measurement and atrial fibrillation detection methods are independently performed, making it difficult to obtain accurate blood pressure values and reliable atrial fibrillation risk assessments in one measurement process, and there are interference effects such as cuff noise and inflation noise, resulting in poor blood pressure measurement. precise.
The dual-mode blood pressure calculation method is adopted to synchronously sample and preprocess the pressure pulse signal, PPG signal and ECG signal to perform interference pattern recognition and quality evaluation, extract multimodal features, and combine with deep neural network for timing analysis to realize blood pressure measurement and room Comprehensive analysis of tremor risk assessment.
It improves the accuracy and reliability of blood pressure measurement, enhances anti-interference ability, provides a comprehensive cardiovascular health monitoring solution, and integrates atrial fibrillation risk assessment function.
Smart Images

Figure CN118648883B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of blood pressure measurement, and in particular to a dual-mode blood pressure calculation method, device, equipment and storage medium. Background Art
[0002] Blood pressure measurement and atrial fibrillation detection are two key links in cardiovascular health management. Traditional methods often need to be performed separately, which increases the burden on patients and medical costs. Existing single-mode blood pressure measurement methods have some limitations. For example, boost measurements may be affected by factors such as cuff noise and inflation noise, while step-down measurements may have occasional large errors. In addition, conventional atrial fibrillation detection usually relies on professional electrocardiogram equipment, which is not convenient for daily monitoring and early screening.
[0003] It is still a challenge for existing technologies to effectively and synchronously acquire and preprocess heterogeneous signals, and to accurately extract useful features in the presence of various interferences. In addition, existing blood pressure measurement and atrial fibrillation detection methods are often performed independently and lack organic integration. How to obtain accurate blood pressure values and reliable atrial fibrillation risk assessment in one measurement process, and how to adapt to changes in different users and measurement environments, is an urgent problem to be solved in current research. Summary of the invention
[0004] The present application provides a dual-mode blood pressure calculation method, device, equipment and storage medium. The present application improves the accuracy of blood pressure measurement and integrates an atrial fibrillation risk assessment function, providing users with a more comprehensive cardiovascular health monitoring solution.
[0005] In a first aspect, the present application provides a dual-mode blood pressure calculation method, the dual-mode blood pressure calculation method comprising:
[0006] Perform synchronous sampling and preprocessing on the pressure pulse signal, PPG signal and ECG signal to obtain multi-source synchronous signal data;
[0007] Performing interference pattern recognition and quality assessment on the multi-source synchronous signal data to obtain a signal quality score and an availability index;
[0008] Performing multimodal feature extraction and fusion analysis on the multi-source synchronous signal data according to the signal quality score and the availability index to obtain a target feature vector;
[0009] Performing dual-mode calculation of boost measurement and depressurization measurement on the target feature vector to obtain a target blood pressure value and measurement reliability;
[0010] Inputting the target feature vector and the target blood pressure value into a preset deep neural network model for time series analysis and atrial fibrillation detection to obtain an atrial fibrillation risk score;
[0011] The target blood pressure value, the measurement reliability and the atrial fibrillation risk score are comprehensively analyzed to obtain a comprehensive blood pressure assessment report.
[0012] In a second aspect, the present application provides a dual-mode blood pressure calculation device, the dual-mode blood pressure calculation device comprising:
[0013] A sampling module is used to synchronously sample and preprocess the pressure pulse signal, PPG signal and ECG signal to obtain multi-source synchronous signal data;
[0014] An identification module, used to perform interference pattern identification and quality assessment on the multi-source synchronous signal data to obtain a signal quality score and an availability index;
[0015] A feature extraction module, used for performing multimodal feature extraction and fusion analysis on the multi-source synchronous signal data according to the signal quality score and the availability index to obtain a target feature vector;
[0016] A dual-mode calculation module, used for performing dual-mode calculation of boost measurement and depressurization measurement on the target feature vector to obtain a target blood pressure value and measurement reliability;
[0017] an atrial fibrillation detection module, used for inputting the target feature vector and the target blood pressure value into a preset deep neural network model for time series analysis and atrial fibrillation detection to obtain an atrial fibrillation risk score;
[0018] The comprehensive analysis module is used to comprehensively analyze the target blood pressure value, the measurement reliability and the atrial fibrillation risk score to obtain a comprehensive blood pressure assessment report.
[0019] The third aspect of the present application provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned dual-mode blood pressure calculation method.
[0020] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, enable the computer to execute the above-mentioned dual-mode blood pressure calculation method.
[0021] In the technical solution provided by this application, high-frequency synchronous sampling and adaptive filtering are used to achieve high-quality acquisition and preprocessing of pressure pulse signals, PPG signals and ECG signals. Wavelet packet decomposition and multiple interference pattern recognition algorithms are used to effectively detect and evaluate various types of signal interference, thereby improving the anti-interference ability of blood pressure measurement and atrial fibrillation detection. By extracting and fusion of features from different signal sources, the complementary information of various physiological signals is fully utilized, and the expressiveness and robustness of features are enhanced. Combining the two measurement modes of boost and buck, and introducing quality assessment and adaptive calibration mechanisms, the accuracy and reliability of blood pressure measurement are significantly improved. An innovative deep neural network structure, including multi-scale convolution, cross-scale fusion and multi-task learning, is used to achieve high-precision atrial fibrillation risk assessment. Through comprehensive analysis of blood pressure values, measurement credibility and atrial fibrillation risk scores, a comprehensive health assessment report is obtained. This application improves the accuracy of blood pressure measurement and integrates atrial fibrillation risk assessment functions, providing users with a more comprehensive cardiovascular health monitoring solution. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0023] Figure 1 This is a schematic diagram of an embodiment of a dual-mode blood pressure calculation method in an embodiment of the present application;
[0024] Figure 2 Schematic diagram of an embodiment of a dual-mode blood pressure calculation device in an embodiment of the present application. DETAILED DESCRIPTION
[0025] Embodiments of the present application provide a dual-mode blood pressure calculation method, apparatus, device and storage medium. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0026] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the dual-mode blood pressure calculation method in the embodiment of the present application includes:
[0027] Step 101, synchronously sampling and preprocessing the pressure pulse signal, PPG signal and ECG signal to obtain multi-source synchronous signal data;
[0028] It is understandable that the execution subject of the present application may be a dual-mode blood pressure calculation device, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.
[0029] Specifically, the pressure pulse signal, PPG signal and ECG signal are synchronously sampled at a high frequency of 2000Hz to obtain the original multi-source signal data. The original multi-source signal data is denoised by using adaptive wavelet transform to obtain the initial denoised signal. The adaptive wavelet transform can effectively remove the noise components in the signal, retain useful signal details and improve the quality of the signal. The initial denoised signal is subjected to cubic spline interpolation baseline drift correction to obtain a baseline correction signal. By eliminating the influence of baseline drift on the signal, the stability and accuracy of the signal are ensured. According to the preset segmentation rules, the baseline correction signal is segmented into a fixed duration of 4 seconds and a fixed number of 4 pulses to obtain a duration signal segment and a pulse number signal segment. The duration signal segment and the pulse number signal segment are subjected to fast Fourier transform to obtain frequency domain feature data. The frequency domain feature data can reflect the distribution of the signal on different frequency components. According to the frequency domain feature data, the duration signal segment and the pulse number signal segment are subjected to adaptive Wiener filtering to obtain an optimized signal segment. Adaptive Wiener filtering effectively removes noise and improves the signal-to-noise ratio by combining the statistical characteristics of the signal. Dynamic time warping is performed on the optimized signal segments to obtain time-aligned signals. Dynamic time warping can align different signal segments, solve the problem of signal inconsistency in time, and ensure the synchronization of signals in the time dimension. According to the time-aligned signal, the mutual correlation coefficient between the pressure pulse signal, PPG signal and ECG signal is calculated to obtain the signal synchronization metric. The mutual correlation coefficient can quantitatively evaluate the degree of synchronization between different signals. According to the signal synchronization metric, the time-aligned signal is interpolated and resampled and phase corrected to obtain multi-source synchronous signal data. Interpolation resampling interpolates between signals so that the signals can be aligned on the same time axis to ensure the synchronization of the signals; phase correction adjusts the phase of the signal to eliminate the phase difference between different signals and improve the synchronization and consistency of the signal.
[0030] Step 102: performing interference pattern recognition and quality assessment on multi-source synchronous signal data to obtain a signal quality score and availability index;
[0031] Specifically, the multi-source synchronous signal data is decomposed by wavelet packets to obtain multi-scale signal components, providing information of each signal source in different frequency ranges. Based on the multi-scale signal components, the signal-to-noise ratio and baseline stability index of each signal source are calculated to obtain the initial signal quality index, which reflects the basic quality of the signal, including its noise level and baseline stability. Interference pattern recognition is performed on the multi-source synchronous signal data. By analyzing the specific patterns in the signal, interference patterns such as cuff jitter, body movement and breathing are identified. By constructing interference pattern feature vectors, each feature vector represents an interference pattern and contains the characteristic information of the pattern in the signal. According to the interference pattern feature vectors, the occurrence location, frequency and intensity of various interferences are calculated to obtain the interference degree score, which reflects the influence of different interferences in the signal. At the same time, the pressure signal in the multi-source synchronous signal data is slope analyzed to detect the abnormality of the inflation rate and obtain the abnormal index of the pressurization control, which helps to identify the abnormality in the inflation process caused by equipment problems or other reasons. The PPG signal in the multi-source synchronous signal data is subjected to motion artifact detection, the degree of motion interference is evaluated, and the reliability index of the PPG signal is obtained. This index reflects the reliability of the PPG signal under motion interference. The signal quality score of each signal is calculated by combining the initial signal quality index, interference level score, pressure control abnormality index and PPG signal reliability index. The availability index is obtained by performing dynamic threshold analysis and adaptive weighted averaging on each 4-second time window of multi-source synchronous signal data. These indices reflect the availability of the signal in different time windows.
[0032] Step 103: performing multimodal feature extraction and fusion analysis on the multi-source synchronous signal data according to the signal quality score and the availability index to obtain a target feature vector;
[0033] Specifically, the multi-source synchronous signal data is weighted and segmented according to the signal quality score and availability index to obtain high-quality signal segments. The pressure pulse signal in the high-quality signal segment is subjected to peak detection and waveform analysis, and features such as pulse interval, pulse amplitude, rise time and fall time are extracted to construct a pressure pulse feature vector to describe the time and amplitude changes of the pulse signal. At the same time, the PPG signal in the high-quality signal segment is subjected to waveform analysis to extract the peak, valley, area and slope features of the waveform to construct a PPG feature vector to reflect the morphological changes of blood flow and pulse wave. The ECG signal in the high-quality signal segment is subjected to R wave detection and interval analysis, and the RR interval and QT interval are calculated to construct an ECG feature vector to reflect the electrical activity of the heart, which is an important indicator for evaluating the health status of the heart. The second-order derivative of the pressure pulse feature vector is calculated to obtain the second-order feature vector of the pressure pulse. The PPG feature vector is subjected to frequency domain analysis to obtain the PPG frequency domain feature vector. The frequency domain feature can reflect the distribution of the PPG signal on different frequency components. The pressure pulse feature vector, PPG feature vector, ECG feature vector, pressure pulse second-order feature vector and PPG frequency domain feature vector are cascaded to obtain a multimodal feature vector. By combining the features of different signal sources to construct a comprehensive feature vector, the information of each signal source can be fully utilized to improve the accuracy and reliability of blood pressure calculation. The multimodal feature vector is subjected to feature dimensionality reduction to obtain the target feature vector. By reducing the dimension of the feature, redundant information is removed and key features are retained, the calculation process is simplified and the calculation efficiency is improved.
[0034] Step 104, performing dual-mode calculation of boost measurement and buck measurement on the target feature vector to obtain a target blood pressure value and measurement reliability;
[0035] Specifically, the target feature vector is segmented and divided into the features of the boost stage and the features of the depressurization stage to obtain the dual-mode feature set, so as to analyze the features of each stage in the boost and depressurization process respectively. The boost systolic pressure and diastolic pressure of the boost stage features in the dual-mode feature set are calculated to obtain the initial value of the boost blood pressure. The initial value of the blood pressure is estimated by analyzing the feature changes in the boost process. The quality of the initial value of the boost blood pressure is evaluated to determine whether there is interference that affects the judgment of the systolic pressure or diastolic pressure. Through the quality evaluation, the boost measurement quality index is obtained, and then the credibility of the initial value of the boost is evaluated. At the same time, the depressurization stage features in the dual-mode feature set are calculated to obtain the initial value of the depressurization blood pressure. The feature analysis of the depressurization stage can provide key information in the process of blood pressure reduction, which is helpful to improve the overall assessment of blood pressure. The initial values of the boost blood pressure and the initial values of the depressurization blood pressure are weighted fused to obtain the initial target blood pressure value. The measurement credibility score is calculated according to the boost measurement quality index, which reflects the quality and reliability of the initial blood pressure value. The initial target blood pressure value is adaptively calibrated in combination with the measurement credibility score and physiological constraints. Physiological constraints include individual physiological characteristics and historical blood pressure data. By considering these factors, the personalization and accuracy of blood pressure measurement can be improved. During the adaptive calibration process, the initial blood pressure value is adjusted based on the credibility score to eliminate possible interference and errors, and obtain a more accurate target blood pressure value and the final measurement credibility.
[0036] Step 105: input the target feature vector and the target blood pressure value into a preset deep neural network model for time series analysis and atrial fibrillation detection to obtain an atrial fibrillation risk score;
[0037] Specifically, the target feature vector is segmented to obtain a feature time series sequence. These feature time series sequences retain the time dependency of the original data and can better reflect the dynamic changes of physiological signals. The feature time series sequence is input into the multi-scale convolution layer of the deep neural network model, and the multi-scale features are extracted by using convolution kernels of different sizes through the multi-scale convolution layer to obtain a multi-scale feature map. The multi-scale feature map is input into the cross-scale fusion layer of the deep neural network model, and the cross-scale fusion layer uses a cross-scale attention mechanism to fuse features of different scales to obtain a fused feature map. The cross-scale fusion layer enhances the richness and expression ability of feature representation by focusing on the mutual relationship between features of different scales. The fused feature map is input into the bidirectional gated recurrent unit layer of the deep neural network model, and the long-term and short-term temporal dependencies are captured by the bidirectional gated recurrent unit layer to obtain a time series feature vector. Through forward and backward cyclic calculations, the temporal dependency of the signal can be better captured and the correlation between signals at different time points can be reflected. The time series feature vector is input into the self-attention layer of the deep neural network model, and the correlation weights between features are calculated by the self-attention layer to obtain a weighted time series feature. The self-attention layer can dynamically allocate weights between features, highlight important features, suppress noise features, and improve the accuracy and robustness of feature representation. The weighted time series features are spliced with the target blood pressure value and input into the feature interaction layer of the deep neural network model. The feature interaction layer uses the cross-attention mechanism to model the interaction between features and obtain the interaction feature vector. The feature interaction layer can capture the potential interaction information between features and improve the expressiveness of the model by simulating the complex relationship between different features. The interaction feature vector is input into the residual fully connected layer of the deep neural network model, and the gradient propagation is enhanced through the jump connection of the residual fully connected layer to obtain a high-order feature representation. The residual fully connected layer can effectively alleviate the gradient vanishing problem in the deep network by introducing residual connections, ensuring the training stability and efficiency of the model. The high-order feature representation is input into the multi-task learning layer of the deep neural network model, and the probability of atrial fibrillation and pulse regularity index are simultaneously predicted through the multi-task learning layer to obtain a multi-task output vector. The multi-task learning layer can improve the generalization ability and prediction accuracy of the model by learning multiple related tasks at the same time. The multi-task output vector is input into the probability calibration layer of the deep neural network model. The predicted probability is calibrated by the probability calibration layer using the temperature scaling method to obtain the calibrated probability of atrial fibrillation. The probability calibration layer adjusts the temperature parameter of the predicted probability to make the output probability of the model more consistent with the actual probability distribution, thereby improving the credibility and accuracy of the model prediction. According to the calibrated probability of atrial fibrillation and pulse regularity indicators, combined with the preset risk assessment rules, the atrial fibrillation risk score is calculated.
[0038] The high-order feature representation is feature segmented to obtain the feature subsets of the atrial fibrillation task and the feature subsets of the pulse regularity task. By reasonably allocating high-order features, each task subset can focus on its own task features, improving the accuracy of prediction and the independence of tasks. The feature subset of the atrial fibrillation task is mapped to obtain the atrial fibrillation feature map, and the feature subset of the pulse regularity task is mapped to obtain the pulse regularity feature map. The mapping transformation makes the features more discernible in the new feature space through nonlinear transformation, and can better reflect the task-related information. The atrial fibrillation feature map is input into the atrial fibrillation task subnetwork to obtain the initial atrial fibrillation probability. The atrial fibrillation task subnetwork can effectively capture the feature information related to atrial fibrillation through multi-layer nonlinear transformation and feature extraction, and preliminarily predict the probability of atrial fibrillation. At the same time, the pulse regularity feature map is input into the pulse regularity task subnetwork to obtain the initial pulse regularity index. The pulse regularity task subnetwork can preliminarily evaluate the stability and regularity of the pulse by analyzing the regularity characteristics of the pulse signal. The initial atrial fibrillation probability and the initial pulse regularity index are subjected to feature interaction to obtain the interactive features. Feature interaction can capture more complex feature associations and improve the overall performance of the model by simulating the interaction between the features of two tasks. The interaction features are input into the shared representation layer to obtain the shared task features. The shared representation layer constructs a unified feature representation by fusing the feature information of different tasks. According to the shared task features, the initial probability of atrial fibrillation is corrected to obtain the corrected probability of atrial fibrillation. The correction process adjusts the initial prediction results by combining the shared feature information, making the prediction of the probability of atrial fibrillation more accurate and reliable. Similarly, according to the shared task features, the initial pulse regularity index is corrected to obtain the corrected pulse regularity index. The corrected pulse regularity index can more accurately reflect the regularity and stability of the pulse by supplementing the information of the shared task features. The corrected probability of atrial fibrillation and the corrected pulse regularity index are spliced to obtain a multi-task output vector, which includes both the risk assessment of atrial fibrillation and the regularity assessment of the pulse.
[0039] Step 106: Comprehensively analyze the target blood pressure value, measurement reliability, and atrial fibrillation risk score to obtain a comprehensive blood pressure assessment report.
[0040] Specifically, the target blood pressure value, measurement reliability and atrial fibrillation risk score are analyzed and integrated. The target blood pressure value is obtained by accurately calculating multi-source signals and fusing multimodal features, and is an important parameter reflecting the individual's blood pressure status. The measurement reliability is based on multiple factors such as signal quality score and interference degree score, and the reliability and accuracy of the measured blood pressure value are ensured through comprehensive analysis. The atrial fibrillation risk score is obtained after time series analysis and atrial fibrillation detection by a deep neural network model, and is used to assess the risk probability of atrial fibrillation in an individual. In the comprehensive analysis process, the target blood pressure value is preliminarily analyzed to determine whether it is in the normal range, high or low, and its accuracy is evaluated in combination with the measurement reliability. If the measurement reliability is high, it means that the target blood pressure value has a high reference value and can be directly used to assess the individual's blood pressure health status. If the measurement reliability is low, further analysis of interference factors is required, and re-measurement or calibration combined with historical data may be required. The target blood pressure value and measurement reliability are integrated and analyzed with the atrial fibrillation risk score. By associating blood pressure values with atrial fibrillation risk, a more comprehensive assessment of cardiovascular health status can be made. For example, the combination of high blood pressure values and high atrial fibrillation risk scores may indicate that an individual has a high risk of cardiovascular disease and needs to take timely medical intervention measures. If the blood pressure value is normal but the atrial fibrillation risk score is high, further monitoring of heart health is needed to prevent potential arrhythmia problems. During the analysis process, multivariate statistical analysis and machine learning models are applied to perform pattern recognition and trend analysis on a large amount of individual data to more accurately assess the health status of individuals. For example, through cluster analysis, individuals can be divided into different health risk levels, and corresponding health advice and intervention measures are provided according to different levels. At the same time, combined with the individual's historical health data and lifestyle data, personalized analysis and prediction are carried out to provide a more customized health management plan. The comprehensive blood pressure assessment report is presented in the form of charts and text descriptions, clearly and intuitively showing the comprehensive analysis results of the target blood pressure value, measurement credibility and atrial fibrillation risk score. The report not only includes the current health status assessment, but also provides corresponding health advice and preventive measures.
[0041] In the embodiment of the present application, high-frequency synchronous sampling and adaptive filtering are used to achieve high-quality acquisition and preprocessing of pressure pulse signals, PPG signals and ECG signals. Wavelet packet decomposition and multiple interference pattern recognition algorithms are used to effectively detect and evaluate various types of signal interference, thereby improving the anti-interference ability of blood pressure measurement and atrial fibrillation detection. By extracting and fusion of features from different signal sources, the complementary information of various physiological signals is fully utilized, and the expressiveness and robustness of features are enhanced. Combining the two measurement modes of boost and buck, and introducing quality assessment and adaptive calibration mechanisms, the accuracy and reliability of blood pressure measurement are significantly improved. An innovative deep neural network structure, including multi-scale convolution, cross-scale fusion and multi-task learning, is used to achieve high-precision atrial fibrillation risk assessment. Through comprehensive analysis of blood pressure values, measurement credibility and atrial fibrillation risk scores, a comprehensive health assessment report is obtained. This application improves the accuracy of blood pressure measurement, integrates atrial fibrillation risk assessment functions, and provides users with a more comprehensive cardiovascular health monitoring solution.
[0042] In a specific embodiment, the process of executing step 101 may specifically include the following steps:
[0043] (1) Perform high-frequency synchronous sampling of 2000 Hz on the pressure pulse signal, PPG signal and ECG signal to obtain the original multi-source signal data, and perform adaptive wavelet transform denoising on the original multi-source signal data to obtain the initial denoised signal;
[0044] (2) performing cubic spline interpolation baseline drift correction on the initial noise reduction signal to obtain a baseline correction signal, and segmenting the baseline correction signal into 4-second fixed duration segments and 4 pulse fixed number segments according to a preset segmentation rule to obtain duration signal segments and pulse number signal segments;
[0045] (3) performing fast Fourier transform on the duration signal segment and the pulse rate signal segment to obtain frequency domain feature data, and performing adaptive Wiener filtering on the duration signal segment and the pulse rate signal segment according to the frequency domain feature data to obtain optimized signal segments;
[0046] (4) Performing dynamic time warping on the optimized signal segments to obtain a time-aligned signal, and calculating the correlation coefficients between the pressure pulse signal, the PPG signal, and the ECG signal based on the time-aligned signal to obtain a signal synchronization metric value;
[0047] (5) The time-aligned signal is interpolated, resampled and phase corrected according to the signal synchronization metric value to obtain multi-source synchronous signal data.
[0048] Specifically, adaptive wavelet transform is used to denoise the original multi-source signal data. Wavelet transform is a signal processing tool that can localize signals in both time and frequency. Adaptive wavelet transform can effectively remove noise and retain the main features of the signal by selecting the most suitable mother wavelet and decomposition layer. represents the original signal, represents the wavelet transform coefficient, then the wavelet transform of the signal can be expressed as:
[0049] ;
[0050] in, The scale is And the displacement is The wavelet function of . The initial noise reduction signal is obtained by thresholding the wavelet transform coefficients to remove high-frequency noise components. The initial noise reduction signal is corrected for baseline drift using cubic spline interpolation. Baseline drift refers to the slow change of the signal baseline over time, which may be caused by device drift or body movement. Cubic spline interpolation can smoothly fit the signal baseline and thus perform correction. is the signal after noise reduction, is the signal after baseline correction, then the cubic spline interpolation correction can be expressed as:
[0051] ;
[0052] in, It is the baseline drift function obtained by cubic spline interpolation fitting. The baseline correction signal improves the stability and accuracy of the signal by eliminating low-frequency drift. According to the preset segmentation rules, the baseline correction signal is segmented into 4 seconds fixed time length and 4 pulse fixed number segments. The time length signal segments and pulse number signal segments provide signal features in different dimensions. After obtaining these signal segments, a fast Fourier transform (FFT) is performed on them to extract frequency domain feature data. FFT can convert time domain signals into frequency domain signals and reveal the frequency components in the signal. Assume is the time domain signal, For its frequency domain representation, the FFT can be expressed as:
[0053] ;
[0054] By analyzing the frequency domain feature data, the main frequency components of the signal can be identified. According to the frequency domain feature data, the duration signal segment and the pulse number signal segment are adaptively filtered. Wiener filtering is a filtering method based on the minimum mean square error criterion, which can effectively remove noise while retaining signal characteristics. is the input signal, is the filtered output signal, then the adaptive Wiener filter can be expressed as:
[0055] ;
[0056] in, is a filter function in the frequency domain, which is adaptively adjusted to minimize the impact of noise. The optimized signal fragment is processed by Wiener filtering to obtain a clearer and more reliable signal. Dynamic time warping is performed on the optimized signal fragment to align the time axes of different signals. Dynamic time warping is a time series alignment algorithm that can find the best matching path between two signals and eliminate time differences. and For two signals to be aligned, the goal of dynamic time warping is to minimize the distance function between them:
[0057] ;
[0058] By calculating the optimal path, a time-aligned signal is obtained. Based on the time-aligned signal, the cross-correlation coefficients between the pressure pulse signal, PPG signal, and ECG signal are calculated to obtain the signal synchronization metric. The cross-correlation coefficient can quantitatively evaluate the degree of synchronization between different signals. and is the aligned signal, the cross-correlation coefficient It can be expressed as:
[0059] ;
[0060] in, and are the average values of the signals respectively. A high cross-correlation coefficient indicates that the signals have high synchronization. According to the signal synchronization metric, the time-aligned signals are interpolated and resampled and phase corrected to obtain multi-source synchronous signal data. Interpolation resampling ensures the synchronization of signals at the same time point by interpolating on the time axis. Phase correction adjusts the phase of the signal to eliminate the phase difference between different signals and obtain consistent multi-source synchronous signal data.
[0061] In a specific embodiment, the process of executing step 102 may specifically include the following steps:
[0062] (1) Perform wavelet packet decomposition on multi-source synchronous signal data to obtain multi-scale signal components, and calculate the signal-to-noise ratio and baseline stability index of each signal source based on the multi-scale signal components to obtain the initial signal quality index;
[0063] (2) Identify the interference patterns of cuff vibration, body motion, and breathing based on multi-source synchronous signal data, and construct the interference pattern feature vector;
[0064] (3) The occurrence location, frequency and intensity of each type of interference are calculated based on the interference pattern feature vector to obtain the interference degree score. The pressure signal in the multi-source synchronous signal data is analyzed by slope, the inflation rate anomaly is detected, and the pressurization control anomaly index is obtained.
[0065] (4) Perform motion artifact detection on the PPG signal in the multi-source synchronous signal data, evaluate the degree of motion interference, and obtain the PPG signal reliability index;
[0066] (5) Calculate the corresponding signal quality score based on the initial signal quality index, interference degree score, pressurization control abnormality index and PPG signal reliability index;
[0067] (6) Perform dynamic threshold analysis and adaptive weighted averaging on each 4-second time window of multi-source synchronous signal data to obtain the availability index.
[0068] Specifically, the multi-source synchronous signal data is decomposed by wavelet packet. Wavelet packet decomposition is a signal processing technology that can decompose the signal into multiple scale signal components in different frequency ranges. Assume that there is a multi-source synchronous signal data set, which contains pressure signal, PPG signal and ECG signal, denoted as .right After wavelet packet decomposition, a series of multi-scale signal components are obtained, which can reveal the characteristics of the signal in different frequency ranges. Indicated in scale and location The wavelet packet component at , then:
[0069] ;
[0070] in, is the corresponding wavelet packet basis function. The signal-to-noise ratio (SNR) and baseline stability index of each signal source are calculated based on the multi-scale signal components. The signal-to-noise ratio is an important indicator to measure signal quality, which represents the ratio of the useful part of the signal to the noise part. is the signal power, is the noise power, then the signal-to-noise ratio can be expressed as:
[0071] ;
[0072] The baseline stability index is used to evaluate the stability of the signal baseline, which is determined by analyzing the changes in the low-frequency components of the signal. The signal quality can be preliminarily evaluated by calculating the signal-to-noise ratio and baseline stability index, and the initial signal quality index can be obtained. The interference patterns of cuff jitter, body movement and breathing are identified based on the multi-source synchronous signal data. By analyzing the signal characteristics, the interference pattern feature vector is constructed to represent the characteristic information of different interference patterns in the signal. Let the interference pattern feature vector be , the location, frequency and intensity of the interference pattern can be identified through the pattern recognition algorithm. Based on this information, the interference degree score of each type of interference is calculated. The interference degree score can quantify the impact of different interferences on the signal. At the same time, the slope analysis of the pressure signal in the multi-source synchronous signal data is performed to detect abnormal inflation rate. The slope analysis calculates the instantaneous rate of change of the pressure signal to identify whether there is an abnormality in the inflation process. Assume that the pressure signal is , its instantaneous slope can be expressed as:
[0073] Slope ;
[0074] By analyzing the slope change, the inflation rate abnormality can be detected and the pressure control abnormality index can be obtained. For the PPG signal in the multi-source synchronous signal data, motion artifact detection is performed to evaluate the degree of motion interference. Motion artifact is a signal artifact caused by individual movement. This interference can be identified and evaluated through a specific detection algorithm. Suppose the output of the motion artifact detection algorithm is , the reliability index of the PPG signal can be obtained through analysis. According to the initial signal quality index, interference level score, pressure control abnormality index and PPG signal reliability index, the corresponding signal quality score is calculated comprehensively. The signal quality score can fully reflect the overall quality of the signal and is an important indicator for evaluating the effectiveness of the signal. Dynamic threshold analysis and adaptive weighted averaging are performed on each 4-second time window of multi-source synchronous signal data to obtain the availability index. Dynamic threshold analysis identifies and excludes abnormal signals by setting an adaptive threshold in each time window. Adaptive weighted averaging obtains a comprehensive availability index by weighted averaging the signal quality scores in different time windows.
[0075] In a specific embodiment, the process of executing step 103 may specifically include the following steps:
[0076] (1) Weighted segmentation of multi-source synchronous signal data is performed according to the signal quality score and availability index to obtain high-quality signal segments;
[0077] (2) Perform peak detection and waveform analysis on the pressure pulse signal in the high-quality signal segment, extract the pulse interval, pulse amplitude, rise time and fall time, and construct the pressure pulse feature vector;
[0078] (3) Perform waveform analysis on the PPG signal in the high-quality signal segment, extract the peak, valley, area and slope features of the waveform, and construct a PPG feature vector;
[0079] (4) Perform R wave detection and interval analysis on the ECG signal in the high-quality signal segment, calculate the RR interval and QT interval, and construct the ECG feature vector;
[0080] (5) Calculating the second-order derivative of the pressure pulse feature vector to obtain the second-order feature vector of the pressure pulse, and performing frequency domain analysis on the PPG feature vector to obtain the PPG frequency domain feature vector;
[0081] (6) Concatenating the pressure pulse feature vector, the PPG feature vector, the ECG feature vector, the pressure pulse second-order feature vector, and the PPG frequency domain feature vector to obtain a multimodal feature vector;
[0082] (7) Perform feature dimensionality reduction on the multimodal feature vector to obtain the target feature vector.
[0083] Specifically, the multi-source synchronous signal data is weighted and segmented according to the signal quality score and availability index, and high-quality signal segments are selected. Each signal segment is assigned a corresponding weight, and the weight value is determined by the signal quality score and availability index. Suppose the signal quality score is , the availability index is , then the weight of the signal segment It can be expressed as:
[0084] ;
[0085] in, and is the weight coefficient, which controls the influence of the two indicators on the final weight. High-quality signal segments are selected through the weighted segmentation method. Peak detection and waveform analysis are performed on the pressure pulse signal in the high-quality signal segment. Peak detection determines the start and end time of the pulse by identifying the peak value of each pulse wave in the pressure pulse signal. Waveform analysis calculates the various features of the pulse waveform to extract features such as pulse interval, pulse amplitude, rise time and fall time. Suppose the pressure pulse signal is , its peak detection algorithm can be expressed as:
[0086] ;
[0087] The pulse interval is the time interval between two adjacent peaks. By extracting these features, the pressure pulse feature vector is constructed. Similarly, the PPG signal in the high-quality signal segment is subjected to waveform analysis to extract the peak, valley, area and slope features of the waveform. Suppose the PPG signal is , its waveform analysis algorithm can be expressed as:
[0088] ;
[0089] The waveform area is calculated by integration, and the slope feature is calculated by the derivative of the signal. By extracting these features, the PPG feature vector is constructed. R wave detection and interval analysis are performed on the ECG signal in the high-quality signal segment. R wave detection determines the time point of the heartbeat by identifying the R wave peak in the ECG signal. Interval analysis calculates the RR interval and QT interval to reflect the various characteristics of the heart cycle. Assume that the ECG signal is , its R wave detection algorithm can be expressed as:
[0090] ;
[0091] The RR interval is the time interval between two adjacent R wave peaks, and the QT interval is the time interval from the start of the Q wave to the end of the T wave. By extracting these features, the ECG feature vector is constructed. After extracting the pressure pulse feature vector, the second-order derivative is calculated to obtain the pressure pulse second-order feature vector. The second-order derivative reflects the acceleration information of the signal change and can provide more detailed dynamic features. Suppose the pressure pulse feature vector is , its second-order derivative can be expressed as:
[0092] ;
[0093] By calculating the second-order derivative, the second-order characteristic vector of the pressure pulse is constructed. At the same time, the PPG characteristic vector is analyzed in the frequency domain to obtain the PPG frequency domain characteristic vector. The frequency domain analysis converts the signal from the time domain to the frequency domain through the fast Fourier transform (FFT) to reveal the frequency components in the signal. Let the PPG characteristic vector be , which is expressed in the frequency domain as:
[0094] ;
[0095] Through frequency domain analysis, the frequency domain features of the PPG signal are extracted and the PPG frequency domain feature vector is constructed. The pressure pulse feature vector, PPG feature vector, ECG feature vector, pressure pulse second-order feature vector and PPG frequency domain feature vector are feature cascaded to obtain a multimodal feature vector. Feature cascade combines the features of different signal sources to form a unified feature vector, which can make full use of the information of each signal source and improve the representation ability of the feature. Feature dimensionality reduction is performed on the multimodal feature vector to obtain the target feature vector. Feature dimensionality reduction reduces the dimension of the feature vector through a dimensionality reduction algorithm (such as principal component analysis PCA), retains important features, and removes redundant information.
[0096] In a specific embodiment, the process of executing step 104 may specifically include the following steps:
[0097] (1) Segment the target feature vector into the boost phase feature and the depressurization phase feature to obtain a dual-mode feature set;
[0098] (2) Calculate the boost systolic and diastolic pressures for the boost phase features in the dual-mode feature set to obtain the initial boost blood pressure value;
[0099] (3) Perform a quality assessment on the initial value of the boosted blood pressure to determine whether there is interference that affects the determination of systolic or diastolic blood pressure, and obtain the boosted blood pressure measurement quality index;
[0100] (4) Calculate the blood pressure reduction based on the blood pressure reduction stage features in the dual-mode feature set to obtain the initial blood pressure reduction value;
[0101] (5) Perform weighted fusion on the initial values of boost blood pressure and the initial values of depressor blood pressure to obtain the initial target blood pressure value, and calculate the measurement credibility score based on the boost measurement quality index;
[0102] (6) Based on the measurement credibility score and physiological constraints, the initial target blood pressure value is adaptively calibrated to obtain the target blood pressure value and measurement credibility.
[0103] Specifically, the target feature vector is segmented and divided into the features of the boosting stage and the features of the depressurization stage to obtain a dual-mode feature set, and the changes of blood pressure in the process of boosting and depressurization are analyzed and processed respectively. After obtaining the dual-mode feature set, the boosting systolic pressure and diastolic pressure of the boosting stage features are calculated to obtain the initial value of the boosting blood pressure. The boosting stage features mainly reflect the changes of physiological signals in the process of blood pressure rising. By analyzing these signals, the initial values of systolic pressure and diastolic pressure are calculated. By peak detection and waveform analysis of the pressure pulse signal, the characteristics of the pulse wave are identified, so as to calculate the initial values of systolic pressure and diastolic pressure. The initial value of the boosting blood pressure is evaluated for quality, and it is determined whether there is interference that affects the judgment of systolic pressure or diastolic pressure, and the boosting measurement quality index is obtained. In the quality evaluation process, the signal quality in the boosting stage features is mainly analyzed to identify interference factors that may affect the blood pressure calculation, such as signal noise, motion artifacts or baseline drift. By comprehensively evaluating these interference conditions, the reliability of the initial value of the boosting blood pressure is determined, and the boosting measurement quality index is calculated accordingly. The blood pressure reduction stage features in the dual-mode feature set are calculated to obtain the initial blood pressure reduction value. The blood pressure reduction stage features mainly reflect the changes in physiological signals during the blood pressure reduction process. By analyzing these signals, the initial values of systolic and diastolic blood pressure in the blood pressure reduction process are calculated. Similar to the pressure increase process, the initial blood pressure value in the blood pressure reduction stage is obtained by waveform analysis and feature extraction of the pressure pulse signal. The initial blood pressure value of the pressure increase and the initial blood pressure reduction are weighted and fused to obtain the initial target blood pressure value, and the measurement credibility score is calculated according to the pressure increase measurement quality index. In the weighted fusion process, the contribution of the pressure increase and pressure reduction stage features to the blood pressure calculation is considered respectively, and the two are weighted according to the pressure increase measurement quality index to obtain a comprehensive initial target blood pressure value. The measurement credibility score is based on the pressure increase measurement quality index. By evaluating the reliability and accuracy of the signal, a score reflecting the credibility of the blood pressure measurement is calculated. Combined with the measurement credibility score and physiological constraints, the initial target blood pressure value is adaptively calibrated to obtain the final target blood pressure value and measurement credibility. Physiological constraints include individual physiological characteristics and historical blood pressure data. By combining this information, the initial target blood pressure value is calibrated to correct possible errors and ensure the accuracy and personalization of the final target blood pressure value. During the adaptive calibration process, the measurement credibility score is used to adjust the initial target blood pressure value, eliminate interference and errors, and obtain an accurate and reliable blood pressure value.
[0104] In a specific embodiment, the process of executing step 105 may specifically include the following steps:
[0105] (1) Segment the target feature vector to obtain a feature time series sequence, and input the feature time series sequence into the multi-scale convolution layer of the deep neural network model. The multi-scale convolution layer uses convolution kernels of different sizes to extract multi-scale features and obtain a multi-scale feature map.
[0106] (2) The multi-scale feature map is input into the cross-scale fusion layer of the deep neural network model, and the cross-scale fusion layer uses the cross-scale attention mechanism to fuse features of different scales to obtain a fused feature map;
[0107] (3) Input the fused feature map into the bidirectional gated recurrent unit layer of the deep neural network model, capture the long-term and short-term temporal dependencies through the bidirectional gated recurrent unit layer, and obtain the temporal feature vector;
[0108] (4) Input the time series feature vector into the self-attention layer of the deep neural network model, calculate the correlation weights between the features through the self-attention layer, and obtain the weighted time series features;
[0109] (5) The weighted time series features are concatenated with the target blood pressure value and input into the feature interaction layer of the deep neural network model. The interaction between features is modeled using the cross-attention mechanism through the feature interaction layer to obtain the interaction feature vector.
[0110] (6) Input the interactive feature vector into the residual fully connected layer of the deep neural network model, enhance the gradient propagation through the skip connection of the residual fully connected layer, and obtain a high-order feature representation;
[0111] (7) Inputting the high-order feature representation into the multi-task learning layer of the deep neural network model, and simultaneously predicting the probability of atrial fibrillation and the pulse regularity index through the multi-task learning layer to obtain a multi-task output vector;
[0112] (8) Inputting the multi-task output vector into the probability calibration layer of the deep neural network model, calibrating the predicted probability using the temperature scaling method through the probability calibration layer, and obtaining the calibrated atrial fibrillation probability;
[0113] (9) Calculate the atrial fibrillation risk score based on the calibrated atrial fibrillation probability and pulse regularity indicators combined with the preset risk assessment rules.
[0114] Specifically, the self-labeled feature vector is sliced according to the time dimension to form a feature time series sequence containing time information. The feature time series sequence can reflect the dynamic process of feature changes over time. The feature time series sequence is input into the multi-scale convolution layer of the deep neural network model, and the multi-scale features are extracted by using convolution kernels of different sizes through the multi-scale convolution layer to obtain a multi-scale feature map. In this process, different convolution kernels of the multi-scale convolution layer can capture the characteristics of the signal at different time scales and generate a multi-scale feature map containing rich multi-scale information. Suppose the input feature time series sequence is , the multi-scale convolution operation can be expressed as:
[0115] ;
[0116] in, Indicates The features extracted by the convolution kernels are It is The weight of the convolution kernel, is the bias, is a nonlinear activation function. The multi-scale feature map is input into the cross-scale fusion layer of the deep neural network model, and the cross-scale fusion layer uses the cross-scale attention mechanism to fuse the features of different scales to obtain a fused feature map. The cross-scale attention mechanism can identify and focus on the correlation between features of different scales, and generate a more comprehensive and detailed feature map by weighted fusion of feature information of different scales. Let the multi-scale feature map be , the cross-scale attention mechanism can be expressed as:
[0117] ;
[0118] in, is the weight calculated by the attention mechanism, is the fused feature map. The fused feature map is input into the bidirectional gated recurrent unit layer of the deep neural network model. The bidirectional gated recurrent unit layer captures the long-term and short-term temporal dependencies and obtains the temporal feature vector. The bidirectional gated recurrent unit can simultaneously capture the forward and backward dependencies of the signal through forward and backward loop calculations, ensuring the integrity and accuracy of the temporal features. Suppose the hidden state of the forward gated recurrent unit is , the hidden state of the backward GRU is , then the output of the bidirectional GRU can be expressed as:
[0119] ;
[0120] The bidirectional gated recurrent unit layer can effectively capture the complex dependencies of the signal in the time dimension. The time series feature vector is input into the self-attention layer of the deep neural network model, and the correlation weights between the features are calculated through the self-attention layer to obtain the weighted time series feature. The self-attention mechanism calculates the relationship between the features and assigns different weights to highlight important features and suppress irrelevant features. By calculating the similarity between the features, a weighted time series feature is generated, which can more accurately reflect the key features of the signal. Let the time series feature vector be , the self-attention mechanism can be expressed as:
[0121] Attention softmax ;
[0122] in, represents the query matrix, represents the key matrix, represents the value matrix, is the dimension of the key, and the weighted time series features are obtained through the self-attention mechanism. The weighted time series features are concatenated with the target blood pressure value and input into the feature interaction layer of the deep neural network model. The cross-attention mechanism is used to model the interaction between features through the feature interaction layer to obtain the interaction feature vector. The cross-attention mechanism can capture the complex relationship between features more comprehensively by modeling the interaction between different features. By concatenating the weighted time series features and the target blood pressure value and inputting into the feature interaction layer, all feature information can be fully utilized to improve the prediction ability of the model. Assume that the weighted time series features are The target blood pressure is , the concatenated features are expressed as:
[0123] ;
[0124] The interactive feature vector is input into the residual fully connected layer of the deep neural network model, and the gradient propagation is enhanced through the skip connection of the residual fully connected layer to obtain a high-order feature representation. The residual fully connected layer can effectively alleviate the gradient vanishing problem in the deep network by introducing residual connections, ensuring the training stability and efficiency of the model. The high-order feature representation retains important feature information through residual connections, while enhancing the expression ability of the model. Assume that the input of the residual fully connected layer is , the output is , then the residual connection can be expressed as:
[0125] ;
[0126] in, represents the weight matrix, Indicates bias, Represents a nonlinear activation function. The high-order feature representation is input into the multi-task learning layer of the deep neural network model. The multi-task learning layer simultaneously predicts the probability of atrial fibrillation and the pulse regularity index to obtain a multi-task output vector. The multi-task learning layer can improve the generalization ability and prediction accuracy of the model by learning multiple related tasks at the same time. In this way, the probability of atrial fibrillation and the pulse regularity index are predicted at the same time. Suppose the input of the multi-task learning layer is , the output is and , respectively represent the probability of atrial fibrillation and the pulse regularity index. The multi-task output vector is input into the probability calibration layer of the deep neural network model, and the predicted probability is calibrated by the probability calibration layer using the temperature scaling method to obtain the calibrated probability of atrial fibrillation. The probability calibration layer adjusts the temperature parameter of the predicted probability to make the output probability of the model more consistent with the actual probability distribution, thereby improving the credibility and accuracy of the model prediction. The temperature scaling method can make the calibrated probability of atrial fibrillation more practical by adjusting the probability distribution. Assume that the temperature parameter is , the initial probability is , then the calibrated probability can be expressed as:
[0127] ;
[0128] The AF risk score is calculated based on the calibrated AF probability and pulse regularity indicators combined with the preset risk assessment rules. The risk assessment rules are preset by medical experts and can provide a comprehensive assessment of AF risk based on clinical experience and research results. The AF risk score provides a comprehensive risk assessment result by comprehensively considering the calibrated AF probability and pulse regularity indicators.
[0129] In a specific embodiment, the execution step inputs the high-order feature representation into the multi-task learning layer of the deep neural network model, and the multi-task learning layer simultaneously predicts the probability of atrial fibrillation and the pulse regularity index to obtain the multi-task output vector, which may specifically include the following steps:
[0130] (1) Perform feature segmentation on the high-order feature representation to obtain the feature subsets for the atrial fibrillation task and the feature subsets for the pulse regularity task;
[0131] (2) performing mapping transformation on the feature subset of the atrial fibrillation task to obtain an atrial fibrillation feature map, and performing mapping transformation on the feature subset of the pulse regularity task to obtain a pulse regularity feature map;
[0132] (3) Inputting the atrial fibrillation feature map into the atrial fibrillation task sub-network to obtain the initial atrial fibrillation probability, and inputting the pulse regularity feature map into the pulse regularity task sub-network to obtain the initial pulse regularity index;
[0133] (4) Perform feature interaction on the initial atrial fibrillation probability and the initial pulse regularity index to obtain interaction features, and input the interaction features into the shared representation layer to obtain shared task features;
[0134] (5) Correcting the initial atrial fibrillation probability according to the shared task characteristics to obtain a corrected atrial fibrillation probability, and correcting the initial pulse regularity index according to the shared task characteristics to obtain a corrected pulse regularity index;
[0135] (6) Concatenate the corrected atrial fibrillation probability and the corrected pulse regularity index to obtain a multi-task output vector.
[0136] Specifically, the high-level feature representation is divided according to the task so as to handle atrial fibrillation detection and pulse regularity assessment respectively. Suppose the high-level feature representation is , the feature subset of atrial fibrillation task is obtained by feature segmentation and pulse regularity task feature subset . Divide the feature representation according to the requirements of different tasks so that each subset contains the most relevant information for the corresponding task. Perform mapping transformation to obtain atrial fibrillation feature mapping The mapping transformation is to transform the feature subset into a new feature space through a mapping function to enhance the expressiveness and recognition of the feature representation. Similarly, the feature subset of the pulse regularity task Perform mapping transformation to obtain pulse regularity feature mapping The mapping transformation can be expressed as:
[0137] ;
[0138] in, and They are the mapping functions for the atrial fibrillation task and the pulse regularity task respectively. Input the atrial fibrillation task subnetwork, and after a series of network layers, the initial atrial fibrillation probability is obtained. The atrial fibrillation task subnetwork is used to process the features related to atrial fibrillation. By analyzing the atrial fibrillation feature map, the probability of atrial fibrillation can be preliminarily predicted. Similarly, the pulse regularity feature map Input the pulse regularity task sub-network to obtain the initial pulse regularity index The pulse regularity task subnetwork can preliminarily evaluate the stability and regularity of the pulse by analyzing the pulse regularity feature map. Initial atrial fibrillation probability and initial pulse regularity index The calculation can be expressed as:
[0139] ;
[0140] in, and are the weight matrices of the atrial fibrillation task and the pulse regularity task sub-networks, and is the bias term, is the activation function. For the initial probability of atrial fibrillation and initial pulse regularity index Perform feature interaction to obtain interactive features Feature interaction is to combine the initial prediction results of two tasks by modeling the relationship between them to obtain a comprehensive feature representation. Feature interaction can be achieved through an interaction function:
[0141] ;
[0142] in, is a feature interaction function that can capture the potential relationship between atrial fibrillation probability and pulse regularity index. Input the shared representation layer to obtain shared task features The shared representation layer constructs a unified feature representation by fusing the feature information of different tasks. Initial probability of atrial fibrillation The corrected probability of atrial fibrillation is obtained. Similarly, the initial pulse regularity index is corrected according to the shared task characteristics to obtain the corrected pulse regularity index. The corrected probability of atrial fibrillation and the corrected pulse regularity index are concatenated to obtain a multi-task output vector. The output vector contains the final probability of atrial fibrillation and the pulse regularity index for further analysis and decision-making.
[0143] The above describes the dual-mode blood pressure calculation method in the embodiment of the present application. The following describes the dual-mode blood pressure calculation device in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the dual-mode blood pressure calculation device includes:
[0144] The sampling module 201 is used to synchronously sample and preprocess the pressure pulse signal, the PPG signal and the ECG signal to obtain multi-source synchronous signal data;
[0145] The identification module 202 is used to perform interference pattern identification and quality assessment on multi-source synchronous signal data to obtain a signal quality score and an availability index;
[0146] The feature extraction module 203 is used to perform multi-modal feature extraction and fusion analysis on the multi-source synchronous signal data according to the signal quality score and the availability index to obtain a target feature vector;
[0147] A dual-mode calculation module 204 is used to perform dual-mode calculations of boost measurement and depressurization measurement on the target feature vector to obtain a target blood pressure value and measurement reliability;
[0148] an atrial fibrillation detection module 205, for inputting the target feature vector and the target blood pressure value into a preset deep neural network model for time series analysis and atrial fibrillation detection to obtain an atrial fibrillation risk score;
[0149] The comprehensive analysis module 206 is used to perform a comprehensive analysis on the target blood pressure value, measurement reliability and atrial fibrillation risk score to obtain a comprehensive blood pressure assessment report.
[0150] Through the synergy of the above components, high-frequency synchronous sampling and adaptive filtering technologies are used to achieve high-quality acquisition and preprocessing of pressure pulse signals, PPG signals and ECG signals. Wavelet packet decomposition and multiple interference pattern recognition algorithms are used to effectively detect and evaluate various types of signal interference, and improve the anti-interference ability of blood pressure measurement and atrial fibrillation detection. By extracting and fusion of features from different signal sources, the complementary information of various physiological signals is fully utilized to enhance the expressiveness and robustness of features. Combining the two measurement modes of boost and buck, and introducing quality assessment and adaptive calibration mechanisms, the accuracy and reliability of blood pressure measurement are significantly improved. The innovative deep neural network structure, including multi-scale convolution, cross-scale fusion and multi-task learning, is used to achieve high-precision atrial fibrillation risk assessment. Through comprehensive analysis of blood pressure values, measurement credibility and atrial fibrillation risk scores, a comprehensive health assessment report is obtained. This application improves the accuracy of blood pressure measurement and integrates atrial fibrillation risk assessment functions to provide users with a more comprehensive cardiovascular health monitoring solution.
[0151] The present application also provides a computer device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the dual-mode blood pressure calculation method in the above-mentioned embodiments.
[0152] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the dual-mode blood pressure calculation method.
[0153] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0154] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.
[0155] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A dual-mode blood pressure calculation method, characterized in that: The dual-mode blood pressure calculation method comprises: The pressure pulse signal, the PPG signal and the ECG signal are synchronously sampled and preprocessed to obtain multi-source synchronous signal data; specifically, the method comprises: performing high-frequency synchronous sampling of the pressure pulse signal, the PPG signal and the ECG signal at 2000 Hz to obtain original multi-source signal data, and performing adaptive wavelet transform denoising on the original multi-source signal data to obtain an initial noise reduction signal; performing cubic spline interpolation baseline drift correction on the initial noise reduction signal to obtain a baseline correction signal, and performing 4-second fixed-time segmentation and 4-pulse fixed-number segmentation on the baseline correction signal according to a preset segmentation rule to obtain a time signal segment and a pulse number signal fragments; performing fast Fourier transform on the duration signal fragments and the pulse rate signal fragments to obtain frequency domain feature data, and performing adaptive Wiener filtering on the duration signal fragments and the pulse rate signal fragments according to the frequency domain feature data to obtain optimized signal fragments; performing dynamic time warping processing on the optimized signal fragments to obtain time-aligned signals, and calculating the mutual correlation coefficients between the pressure pulse signal, the PPG signal and the ECG signal according to the time-aligned signals to obtain signal synchronization measurement values; performing interpolation resampling and phase correction on the time-aligned signals according to the signal synchronization measurement values to obtain multi-source synchronous signal data; Interference pattern recognition and quality assessment are performed on the multi-source synchronous signal data to obtain a signal quality score and an availability index; specifically comprising: performing wavelet packet decomposition on the multi-source synchronous signal data to obtain multi-scale signal components, and calculating the signal-to-noise ratio and baseline stability index of each signal source based on the multi-scale signal components to obtain an initial signal quality index; identifying cuff jitter, body movement and breathing interference patterns based on the multi-source synchronous signal data, and constructing an interference pattern feature vector; calculating the occurrence location, frequency and intensity of various types of interference based on the interference pattern feature vector to obtain an interference degree score, and performing slope analysis on the pressure signal in the multi-source synchronous signal data to detect inflation rate abnormalities and obtain a pressurization control abnormality index; performing motion artifact detection on the PPG signal in the multi-source synchronous signal data to evaluate the degree of motion interference and obtain a PPG signal reliability index; calculating the corresponding signal quality score based on the initial signal quality index, the interference degree score, the pressurization control abnormality index and the PPG signal reliability index; performing dynamic threshold analysis and adaptive weighted averaging on each 4-second time window of the multi-source synchronous signal data to obtain an availability index; The method comprises the following steps: performing multimodal feature extraction and fusion analysis on the multi-source synchronous signal data according to the signal quality score and the availability index to obtain a target feature vector; specifically comprising: performing weighted segmentation on the multi-source synchronous signal data according to the signal quality score and the availability index to obtain high-quality signal segments; performing peak detection and waveform analysis on the pressure pulse signal in the high-quality signal segment to extract the pulse interval, pulse amplitude, rise time and fall time, and construct a pressure pulse feature vector; performing waveform analysis on the PPG signal in the high-quality signal segment to extract the peak value, valley value, area and slope features of the waveform, and construct a PPG characteristic vector; performing R wave detection and interval analysis on the ECG signal in the high-quality signal segment, calculating the RR interval and the QT interval, and constructing an ECG characteristic vector; calculating the second-order derivative of the pressure pulse characteristic vector to obtain the pressure pulse second-order characteristic vector, and performing frequency domain analysis on the PPG characteristic vector to obtain the PPG frequency domain characteristic vector; performing feature cascade on the pressure pulse characteristic vector, the PPG characteristic vector, the ECG characteristic vector, the pressure pulse second-order characteristic vector and the PPG frequency domain characteristic vector to obtain a multimodal characteristic vector; performing feature dimension reduction on the multimodal characteristic vector to obtain a target characteristic vector; Performing dual-mode calculation of boost measurement and depressurization measurement on the target feature vector to obtain a target blood pressure value and measurement reliability; specifically comprising: performing segmentation processing on the target feature vector to divide it into boost stage features and depressurization stage features to obtain a dual-mode feature set; performing boost systolic pressure and diastolic pressure calculation on the boost stage features in the dual-mode feature set to obtain a boost blood pressure initial value; performing quality assessment on the boost blood pressure initial value to determine whether there is interference that affects the systolic or diastolic pressure determination to obtain a boost measurement quality index; performing depressurization calculation on the depressurization stage features in the dual-mode feature set to obtain a depressurization initial value; performing weighted fusion on the boost blood pressure initial value and the depressurization initial value to obtain an initial target blood pressure value, and calculating a measurement reliability score based on the boost measurement quality index; performing adaptive calibration on the initial target blood pressure value in combination with the measurement reliability score and physiological constraints to obtain a target blood pressure value and measurement reliability; The target feature vector and the target blood pressure value are input into a preset deep neural network model for time series analysis and atrial fibrillation detection to obtain an atrial fibrillation risk score; specifically comprising: segmenting the target feature vector to obtain a feature time series sequence, and inputting the feature time series sequence into a multi-scale convolution layer of a deep neural network model, extracting multi-scale features through the multi-scale convolution layer using convolution kernels of different sizes to obtain a multi-scale feature map; inputting the multi-scale feature map into a cross-scale fusion layer of the deep neural network model, and using a cross-scale attention mechanism to fuse features of different scales through the cross-scale fusion layer to obtain a fused feature map; inputting the fused feature map into a bidirectional gated recurrent unit layer of the deep neural network model, and capturing long-term and short-term time series dependencies through the bidirectional gated recurrent unit layer to obtain a time series feature vector; inputting the time series feature vector into a self-attention layer of the deep neural network model, and calculating the correlation between features through the self-attention layer weights to obtain weighted time series features; concatenate the weighted time series features with the target blood pressure values, and input them into the feature interaction layer of the deep neural network model, and use the cross-attention mechanism to model the interaction between features through the feature interaction layer to obtain an interaction feature vector; input the interaction feature vector into the residual fully connected layer of the deep neural network model, and enhance the gradient propagation through the jump connection of the residual fully connected layer to obtain a high-order feature representation; input the high-order feature representation into the multi-task learning layer of the deep neural network model, and use the multi-task learning layer to simultaneously predict the probability of atrial fibrillation and the pulse regularity index to obtain a multi-task output vector; input the multi-task output vector into the probability calibration layer of the deep neural network model, and use the temperature scaling method to calibrate the predicted probability through the probability calibration layer to obtain a calibrated probability of atrial fibrillation; calculate the atrial fibrillation risk score according to the calibrated probability of atrial fibrillation and the pulse regularity index, combined with a preset risk assessment rule; The target blood pressure value, the measurement reliability and the atrial fibrillation risk score are comprehensively analyzed to obtain a comprehensive blood pressure assessment report.
2. The dual-mode blood pressure calculation method according to claim 1, characterized in that: The step of inputting the high-order feature representation into the multi-task learning layer of the deep neural network model, and simultaneously predicting the probability of atrial fibrillation and the pulse regularity index through the multi-task learning layer to obtain a multi-task output vector includes: Performing feature segmentation on the high-order feature representation to obtain an atrial fibrillation task feature subset and a pulse regularity task feature subset; Performing a mapping transformation on the atrial fibrillation task feature subset to obtain an atrial fibrillation feature map, and performing a mapping transformation on the pulse regularity task feature subset to obtain a pulse regularity feature map; Inputting the atrial fibrillation feature map into the atrial fibrillation task subnetwork to obtain an initial atrial fibrillation probability, and inputting the pulse regularity feature map into the pulse regularity task subnetwork to obtain an initial pulse regularity index; Performing feature interaction on the initial atrial fibrillation probability and the initial pulse regularity index to obtain interaction features, and inputting the interaction features into a shared representation layer to obtain shared task features; Correcting the initial atrial fibrillation probability according to the shared task feature to obtain a corrected atrial fibrillation probability, and correcting the initial pulse regularity index according to the shared task feature to obtain a corrected pulse regularity index; The corrected atrial fibrillation probability and the corrected pulse regularity index are concatenated to obtain a multi-task output vector.
3. A dual-mode blood pressure calculation device, characterized in that: The dual-mode blood pressure calculation device comprises: The sampling module is used to synchronously sample and preprocess the pressure pulse signal, the PPG signal and the ECG signal to obtain multi-source synchronous signal data; specifically, the sampling module comprises: performing high-frequency synchronous sampling of 2000 Hz on the pressure pulse signal, the PPG signal and the ECG signal to obtain original multi-source signal data, and performing adaptive wavelet transform denoising on the original multi-source signal data to obtain an initial noise reduction signal; performing cubic spline interpolation baseline drift correction on the initial noise reduction signal to obtain a baseline correction signal, and performing 4-second fixed-time segmentation and 4-pulse fixed-number segmentation on the baseline correction signal according to a preset segmentation rule to obtain a time signal segment and a pulse segment. beat count signal segments; performing fast Fourier transform on the duration signal segments and the pulse count signal segments to obtain frequency domain feature data, and performing adaptive Wiener filtering on the duration signal segments and the pulse count signal segments according to the frequency domain feature data to obtain optimized signal segments; performing dynamic time warping on the optimized signal segments to obtain time-aligned signals, and calculating the mutual correlation coefficients between the pressure pulse signal, the PPG signal and the ECG signal according to the time-aligned signals to obtain signal synchronization metrics; performing interpolation resampling and phase correction on the time-aligned signals according to the signal synchronization metrics to obtain multi-source synchronous signal data; An identification module is used to perform interference pattern recognition and quality assessment on the multi-source synchronous signal data to obtain a signal quality score and an availability index; specifically comprising: performing wavelet packet decomposition on the multi-source synchronous signal data to obtain a multi-scale signal component, and calculating the signal-to-noise ratio and baseline stability index of each signal source according to the multi-scale signal component to obtain an initial signal quality index; identifying cuff jitter, body movement and breathing interference patterns according to the multi-source synchronous signal data, and constructing an interference pattern feature vector; calculating the occurrence location, frequency and intensity of various types of interference according to the interference pattern feature vector to obtain an interference degree score, and performing slope analysis on the pressure signal in the multi-source synchronous signal data to detect inflation rate abnormality and obtain a pressurization control abnormality index; performing motion artifact detection on the PPG signal in the multi-source synchronous signal data to evaluate the degree of motion interference and obtain a PPG signal reliability index; calculating the corresponding signal quality score according to the initial signal quality index, the interference degree score, the pressurization control abnormality index and the PPG signal reliability index; performing dynamic threshold analysis and adaptive weighted averaging on each 4-second time window of the multi-source synchronous signal data to obtain an availability index; The feature extraction module is used to perform multimodal feature extraction and fusion analysis on the multi-source synchronous signal data according to the signal quality score and the availability index to obtain a target feature vector; specifically comprising: weighted segmentation of the multi-source synchronous signal data according to the signal quality score and the availability index to obtain high-quality signal segments; peak detection and waveform analysis of the pressure pulse signal in the high-quality signal segment to extract the pulse interval, pulse amplitude, rise time and fall time to construct a pressure pulse feature vector; waveform analysis of the PPG signal in the high-quality signal segment to extract the peak, valley, area and slope features of the waveform to construct Constructing a PPG feature vector; performing R wave detection and interval analysis on the ECG signal in the high-quality signal segment, calculating the RR interval and the QT interval, and constructing an ECG feature vector; calculating the second-order derivative of the pressure pulse feature vector to obtain a pressure pulse second-order feature vector, and performing frequency domain analysis on the PPG feature vector to obtain a PPG frequency domain feature vector; performing feature cascading on the pressure pulse feature vector, the PPG feature vector, the ECG feature vector, the pressure pulse second-order feature vector, and the PPG frequency domain feature vector to obtain a multimodal feature vector; performing feature dimension reduction on the multimodal feature vector to obtain a target feature vector; A dual-mode calculation module is used to perform dual-mode calculations of boost measurement and depressurization measurement on the target feature vector to obtain a target blood pressure value and measurement reliability; specifically comprising: segmenting the target feature vector to divide it into boost stage features and depressurization stage features to obtain a dual-mode feature set; performing boost systolic and diastolic pressure calculations on the boost stage features in the dual-mode feature set to obtain a boost blood pressure initial value; performing quality assessment on the boost blood pressure initial value to determine whether there is interference that affects the systolic or diastolic pressure determination to obtain a boost measurement quality index; performing depressurization calculations on the depressurization stage features in the dual-mode feature set to obtain a depressurization initial value; performing weighted fusion on the boost blood pressure initial value and the depressurization initial value to obtain an initial target blood pressure value, and calculating a measurement reliability score based on the boost measurement quality index; and adaptively calibrating the initial target blood pressure value in combination with the measurement reliability score and physiological constraints to obtain a target blood pressure value and measurement reliability; The atrial fibrillation detection module is used to input the target feature vector and the target blood pressure value into a preset deep neural network model for time series analysis and atrial fibrillation detection to obtain an atrial fibrillation risk score; specifically comprising: segmenting the target feature vector to obtain a feature time series sequence, and inputting the feature time series sequence into a multi-scale convolution layer of the deep neural network model, extracting multi-scale features through the multi-scale convolution layer using convolution kernels of different sizes to obtain a multi-scale feature map; inputting the multi-scale feature map into a cross-scale fusion layer of the deep neural network model, and using a cross-scale attention mechanism to fuse features of different scales through the cross-scale fusion layer to obtain a fused feature map; inputting the fused feature map into a bidirectional gated recurrent unit layer of the deep neural network model, and capturing long-term and short-term temporal dependencies through the bidirectional gated recurrent unit layer to obtain a time series feature vector; inputting the time series feature vector into a self-attention layer of the deep neural network model, and calculating feature intervals through the self-attention layer The weighted time series feature is obtained by calculating the correlation weight of the target blood pressure value; the weighted time series feature is spliced with the target blood pressure value, and input into the feature interaction layer of the deep neural network model, and the interaction between features is modeled by the cross attention mechanism through the feature interaction layer to obtain an interaction feature vector; the interaction feature vector is input into the residual fully connected layer of the deep neural network model, and the gradient propagation is enhanced by the jump connection of the residual fully connected layer to obtain a high-order feature representation; the high-order feature representation is input into the multi-task learning layer of the deep neural network model, and the atrial fibrillation probability and the pulse regularity index are predicted simultaneously through the multi-task learning layer to obtain a multi-task output vector; the multi-task output vector is input into the probability calibration layer of the deep neural network model, and the predicted probability is calibrated by the temperature scaling method through the probability calibration layer to obtain a calibrated atrial fibrillation probability; according to the calibrated atrial fibrillation probability and the pulse regularity index, combined with the preset risk assessment rules, the atrial fibrillation risk score is calculated; The comprehensive analysis module is used to perform a comprehensive analysis on the target blood pressure value, the measurement reliability and the atrial fibrillation risk score to obtain a comprehensive blood pressure assessment report.
4. A computer device, characterized in that: The computer device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the computer device to perform the bimodal blood pressure calculation method according to any one of claims 1-2.
5. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the dual-mode blood pressure calculation method according to any one of claims 1-2 is implemented.
Citation Information
Patent Citations
Method for dynamically monitoring blood pressure based on multiple physiological characteristic parameters of ECG and PPG
CN114145724A
Apparatus and method for estimating blood pressure
EP4032468A1
Patient risk assessment based on data from multiple sources in a healthcare facility
US20190307405A1
Blood pressure estimation based on pulse wave velocity
WO2017142668A1
Pulse waveform-based detection and categorization of cardiovascular anomalies
WO2023220670A2
