A method for processing an arterial blood pressure waveform signal by a residual neural network

Abstract

The present application relates to the technical field of vital sign detection, in particular to a method for processing arterial blood pressure waveform signals through a residual neural network. The method comprises pre-processing arterial blood pressure waveform signals to extract characteristic values therefrom; performing short-time Fourier transform on the arterial blood pressure waveform signals to generate amplitude and phase images of the signals in the time-frequency domain; inputting the images and the characteristic values into a residual neural network model to obtain predicted values of hemodynamic parameters including cardiac output and stroke volume; and using the time series prediction capability of the residual neural network to predict blood pressure trends in a certain future time. The present application provides more accurate prediction of cardiac output parameters and more accurate prediction of hypotension events. In addition, the present application also increases the quantification of uncertainty in the prediction of hypotension events.

Description

Technical Field

[0001] This invention relates to the field of vital sign detection technology, specifically a method for processing arterial blood pressure waveform signals using a residual neural network. Background Technology

[0002] Timely and accurate blood pressure monitoring is crucial for preventing hypertension and its complications. Currently, traditional blood pressure measurement methods are typically daytime sporadic measurements, which cannot be performed during sleep at night. To achieve continuous dynamic blood pressure monitoring, foreign researchers have proposed cuffless blood pressure monitoring technology. However, because these devices require contact with the wearer's body, prolonged nighttime wear can cause discomfort or skin irritation, resulting in low comfort and affecting the accuracy of nighttime blood pressure measurements. Furthermore, neither daytime sporadic blood pressure measurements nor cuffed blood pressure monitoring can obtain sufficient hemodynamic parameters, nor can they predict the probability of future hypotension events. Chinese patent application 2022800559278 discloses a "Hypotension Prediction Device, Method, and Training Method for a Hypotension Prediction Model Based on Arterial Blood Pressure Wavelet Transform," in which the device receives arterial blood pressure data from a subject and uses a hypotension prediction model to determine whether the subject has hypotension based on the changes in each measurement interval of the arterial blood pressure data. The hypotension prediction model includes: a first layer trained to extract trend data for each interval while performing wavelet transform on the training arterial blood pressure data to compress it; and a subsequence data generation module for generating subsequence data for each trend data based on behavior. The hypotension prediction model is trained by low-pass filtering the training arterial blood pressure data using training parameters learned in the first layer. The hypotension prediction model further includes: a second layer trained to assign weights to training intervals required for hypotension prediction within multiple intervals of the training arterial blood pressure data; and a third layer trained to calculate similarity feature values ​​between the subsequence data of the training intervals and the trend data based on the assigned weights. This patent application can use compressed arterial blood pressure (ABP) data containing overall blood pressure trend information to determine whether hypotension is present. Furthermore, hypotension can be determined based on changes in each measurement interval of compressed arterial blood pressure (ABP) data containing overall blood pressure trend information, so as to provide appropriate treatment by comparing these changes in hypotension occurring during surgery with general patterns. However, the reliability of such deterministic prediction results needs to be evaluated in conjunction with uncertainty estimation. Uncertainty estimation is an important concept in machine learning, referring to a measure of a model's confidence in its predictions. Residual neural networks are a type of deep learning model. In traditional networks, as network depth increases, the model may become difficult to train due to vanishing / exploding gradients, and its performance may saturate or even decline. By introducing residual blocks and skip connections, residual neural networks effectively solve the vanishing gradient problem in deep neural network training, enabling the network to scale to thousands of layers and achieving breakthrough results in many computer vision tasks. Summary of the Invention

[0003] The purpose of this invention is to provide a method for processing arterial blood pressure waveform signals using residual neural networks, which differs from existing technologies. Based on the arterial blood pressure waveform, this method uses convolution operators and deep residual neural networks to directly and accurately predict cardiac output, thereby predicting the uncertainty of hypotension events and providing a range of prediction uncertainty, thus providing richer evidence for clinical use.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: The method for processing arterial blood pressure waveform signals using a residual neural network according to the present invention includes the following steps: 1) Continuously measure and acquire the subject's arterial blood pressure waveform signal; 2) Preprocess the arterial blood pressure waveform signal to eliminate interference and noise, extract feature values ​​from the arterial blood pressure waveform signal, and calculate pulse pressure and Liljestrand-Zander parameters based on the feature values; 3) Perform a short-time Fourier transform on the preprocessed arterial blood pressure waveform signal to generate amplitude and phase images of the signal in the time-frequency domain. Input the amplitude and phase images in the time-frequency domain, as well as the pulse pressure and Liljestrand-Zander parameters, into the residual neural network model to obtain predicted values ​​of hemodynamic parameters, including cardiac output and stroke volume. 4) Utilize the time-series prediction capabilities of residual neural networks to predict blood pressure trends over a certain period. First, calculate the mean arterial pressure (MAP) based on real-time arterial blood pressure. Then, standardize the MAP of the new sample using the mean and standard deviation calculated from the MAP of the training samples. The standardized MAP = (MAP - mean) / standard deviation. Subsequently, perform inference using a Bayesian residual neural network on the standardized MAP. Using the Monte Carlo samples inferred by Bayesian inference, calculate the average probability of hypotension and the uncertainty range of this average probability.

[0005] The aforementioned technical solution directly and accurately predicts cardiac output based on arterial blood pressure waveforms using convolution operators and deep residual neural networks, achieving higher accuracy than current signal processing-based cardiac output prediction algorithms. Furthermore, this solution combines Bayesian methods and deep convolutional neural networks to predict the likelihood of future hypotension events. Compared to existing commercial HPI methods and cutting-edge methods in the literature, this method offers more accurate predictions and provides a range of prediction uncertainty, providing richer evidence for clinical application.

[0006] Preferably, the residual neural network model consists of multiple residual blocks, each containing a convolutional layer, a batch normalization layer, and an activation function layer; during the training of the residual neural network model, historical data labeled with hemodynamic parameters are used to supervise the learning of the network, the hemodynamic parameters including cardiac output and stroke volume.

[0007] The above technical solution effectively solves the gradient vanishing problem in deep neural network training, thereby improving prediction accuracy.

[0008] Preferably, the residual neural network model undergoes hyperparameter optimization. The hyperparameter optimization improves the generalization ability and prediction accuracy of the residual neural network model by adjusting the optimization parameters through grid search and random search. The optimization parameters include learning rate, batch size, and number of network layers.

[0009] By adopting the above technical solutions, the generalization ability and prediction accuracy of the residual neural network model can be improved, which is conducive to improving the accuracy of prediction.

[0010] Preferably, in step 2), the characteristic values ​​include systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse rate, maximum pressure change rate, and minimum pressure change rate. The formula for calculating pulse pressure is: Pulse pressure = Systolic pressure - Diastolic pressure; The formula for calculating the Liljestrand-Zander parameter is: pulse pressure / (systolic blood pressure + diastolic blood pressure).

[0011] By using the above technical solution, the input data for the residual neural network model can be obtained.

[0012] Preferably, in step 3), the residual neural network model outputs a cardiac index, the cardiac index is multiplied by the subject's body surface area to calculate cardiac output, and the cardiac output is divided by the heart rate to calculate stroke volume.

[0013] Using the above technical solution, the hemodynamic parameters output by the residual neural network model can be obtained.

[0014] Compared with the prior art, the beneficial effects of the present invention are: providing more accurate prediction of cardiac output parameters and more accurate prediction of hypotension events; in addition, it also increases the quantification of uncertainty in the prediction of hypotension events. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0016] Figure 2 The flowchart shows the Bayesian inference algorithm for hypotension events. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, the method for processing arterial blood pressure waveform signals using a residual neural network according to the present invention includes the following steps: 1) Continuously measure and acquire the subject's arterial blood pressure waveform signal. Arterial blood pressure, abbreviated as ABP, refers to the lateral pressure of blood against the aortic wall per unit area. The arterial blood pressure waveform signal is the ABP signal, and the sampling frequency in this embodiment is 500Hz; 2) Preprocess the arterial blood pressure waveform signal to eliminate interference and noise, extract feature values ​​from the arterial blood pressure waveform signal, and calculate pulse pressure and Liljestrand-Zander parameters based on the feature values; In this step, the characteristic values ​​include systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse rate, maximum pressure change rate, and minimum pressure change rate.

[0019] Systolic blood pressure, or SBP for short, refers to the pressure inside the arteries that rises during cardiac contraction, reaching its highest value in the middle of the contraction. It represents the pressure of blood against the walls of the blood vessels at this point. Often referred to as high blood pressure, it is an important component of blood pressure measurement, reflecting the systolic function of the left ventricle and the elasticity of the aorta and other large blood vessels. Under normal circumstances, the normal range for systolic blood pressure in adults is 90-140 mmHg.

[0020] Diastolic blood pressure, abbreviated as DBP, is the pressure generated when a person's heart relaxes, due to the elastic recoil of the arteries. It is also called low blood pressure. Under normal circumstances, the normal range of diastolic blood pressure for adults is 60-90 mmHg.

[0021] Mean arterial pressure (MBP) is the average blood pressure in the arteries during a cardiac cycle. It also reflects average tissue perfusion. This value is calculated using the formula: Mean Arterial Pressure = (Systolic Pressure + 2 × Diastolic Pressure) / 3, or Mean Arterial Pressure = Diastolic Pressure + 1 / 3 of the pulse pressure. The normal range for mean arterial pressure in normal adults is 70-105 mmHg.

[0022] Pulse rate, abbreviated as PR, refers to the frequency of arterial pulsations, that is, the number of pulses per minute. Pulse rate is usually consistent with heart rate, and the normal resting pulse rate for adults is 60-100 beats per minute.

[0023] The rate of change of maximum blood pressure refers to the maximum rate of change of arterial blood pressure over a certain period of time. The rate of change of maximum blood pressure is related to various factors such as blood pressure fluctuations, cardiac function, and vascular elasticity.

[0024] The minimum rate of change of arterial blood pressure refers to the minimum rate of change of arterial blood pressure over a given period of time. The formula for calculating pulse pressure based on the above-mentioned characteristic values ​​is: Pulse pressure = Systolic pressure - Diastolic pressure; The Liljestrand-Zander parameter is a formula or method used to calculate cardiac output. This parameter or method can be used to assess the heart's pumping function. In calculating cardiac output, the Liljestrand-Zander formula takes into account physiological parameters such as arterial blood pressure waveforms, thus enabling a relatively accurate estimation of cardiac output.

[0025] The formula for calculating the Liljestrand-Zander parameter in this embodiment is: pulse pressure / (systolic blood pressure + diastolic blood pressure).

[0026] The systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse rate, maximum pressure change rate, minimum pressure change rate, pulse pressure, and Liljestrand-Zander parameters are concatenated into an 8-dimensional feature vector.

[0027] The preprocessed arterial blood pressure waveform signal is truncated for a period of time and subjected to a short-time Fourier transform to generate amplitude and phase images of the signal in the time-frequency domain. In this embodiment, a 50-second segment of the ABP signal is truncated for the short-time Fourier transform. The amplitude and phase images in the time-frequency domain, along with the aforementioned 8-dimensional feature vector, are input into a residual neural network model so that the residual neural network model can obtain the output cardiac index based on the input information. The amplitude and phase images in the time-frequency domain are processed by a ResNet-based convolutional neural network to extract high-level features. The feature values, along with the pulse pressure and Liljestrand-Zander parameters calculated based on the feature values, are input through an LSTM network to capture time-series characteristics. The cardiac index is the value obtained by dividing the volume of blood pumped by the heart by the body surface area. The significance of this index lies in taking into account the differences in body size among different individuals, thereby enabling a more accurate assessment of cardiac performance. The cardiac index is determined by two factors: heart rate and the volume of blood pumped by the heart per stroke. Cardiac output is obtained by multiplying the cardiac index output by the residual neural network model by the subject's body surface area. Stroke volume (BSA) can then be calculated from cardiac output and pulse rate using a simple formula. The subject's body surface area can be calculated using the Du Bois formula: BSA(m²) = 0.007184 × body weight (kg). 0.425 × Height (cm) 0.725 .

[0028] The residual neural network model consists of multiple residual blocks, each containing a convolutional layer, a batch normalization layer, and an activation function layer; the specific architecture of the residual neural network model is existing well-known technology and will not be detailed here. Figure 1 As shown, during the training of the residual neural network model, historical data labeled with hemodynamic parameters were used for supervised learning of the network. These hemodynamic parameters include cardiac output and stroke volume.

[0029] The residual neural network model has undergone hyperparameter optimization. The hyperparameter optimization is achieved by adjusting the optimization parameters through grid search and random search to improve the generalization ability and prediction accuracy of the residual neural network model. The optimization parameters include learning rate, batch size, and number of network layers.

[0030] 4) Finally, the time series prediction capability of residual neural networks is used to predict the blood pressure trend over a certain period of time in the future.

[0031] This step aims to use mean arterial blood pressure to predict hypotension events within the next 5 / 15 minutes and provide an estimate of the uncertainty of the model prediction for the predicted probability, thus providing richer evidence for clinical use.

[0032] First, prepare new samples by calculating the mean arterial pressure (MAP) based on real-time arterial blood pressure. Then, standardize the MAP of the new samples using the mean and standard deviation calculated from the MAP of the training samples. Standardized MAP = (MAP - mean) / standard deviation.

[0033] Subsequently, Bayesian residual neural network inference is performed on the standardized mean arterial pressure. Using the Monte Carlo samples obtained through Bayesian inference, the average probability of hypotension is calculated, and the uncertainty range of this average probability is also calculated. The Bayesian residual neural network refers to incorporating Bayesian methods into the residual blocks of a residual neural network model for inference.

[0034] The so-called Bayesian method is a probabilistic reasoning framework based on Bayes' theorem, which is an existing technology and will not be described in detail here.

[0035] Finally, based on the model's final predictions, corresponding hypotension classification results are generated according to the optimized classification rules. If a high level of uncertainty is assessed, it is marked as a low-confidence result.

[0036] Contents not described in detail in this specification are existing technologies known to those skilled in the art. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for processing arterial blood pressure waveform signals using a residual neural network, characterized in that... Includes the following steps: 1) Continuously measure and acquire the subject's arterial blood pressure waveform signal; 2) Preprocess the arterial blood pressure waveform signal to eliminate interference and noise, extract feature values ​​from the arterial blood pressure waveform signal, and calculate pulse pressure and Liljestrand-Zander parameters based on the feature values; 3) Perform a short-time Fourier transform on the preprocessed arterial blood pressure waveform signal to generate amplitude and phase images of the signal in the time-frequency domain. Input the amplitude and phase images in the time-frequency domain, as well as the pulse pressure and Liljestrand-Zander parameters, into the residual neural network model to obtain predicted values ​​of hemodynamic parameters, including cardiac output and stroke volume. 4) Utilize the time series prediction capabilities of residual neural networks to predict blood pressure trends over a certain period of time in the future.

2. The method for processing arterial blood pressure waveform signals using a residual neural network according to claim 1, characterized in that: The residual neural network model consists of multiple residual blocks, each containing a convolutional layer, a batch normalization layer, and an activation function layer. During the training of the residual neural network model, historical data labeled with hemodynamic parameters are used to supervise the learning of the network. The hemodynamic parameters include cardiac output and stroke volume.

3. The method for processing arterial blood pressure waveform signals using a residual neural network according to claim 2, characterized in that: The residual neural network model undergoes hyperparameter optimization, which involves adjusting the optimization parameters through grid search and random search to improve the generalization ability and prediction accuracy of the residual neural network model. The optimization parameters include learning rate, batch size, and number of network layers.

4. The method for processing arterial blood pressure waveform signals using a residual neural network according to claim 1 or 2, characterized in that: In step 2), the characteristic values ​​include systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse rate, maximum pressure change rate, and minimum pressure change rate. The formula for calculating pulse pressure is: Pulse pressure = Systolic pressure - Diastolic pressure; The formula for calculating the Liljestrand-Zander parameter is: pulse pressure / (systolic blood pressure + diastolic blood pressure).

5. The method for processing arterial blood pressure waveform signals using a residual neural network according to claim 1 or 2, characterized in that: In step 3), the residual neural network model outputs a cardiac index, which is multiplied by the subject's body surface area to calculate cardiac output, and cardiac output is divided by heart rate to calculate stroke volume.

6. The method for processing arterial blood pressure waveform signals using a residual neural network according to claim 1 or 2, characterized in that: In step 4), firstly, the mean arterial pressure is calculated based on real-time arterial blood pressure; then, the mean arterial pressure of the new sample is standardized using the mean and standard deviation calculated on the mean arterial pressure of the training sample; the standardized mean arterial pressure = (mean arterial pressure - mean) / standard deviation; subsequently, Bayesian residual neural network inference is performed on the standardized mean arterial pressure, and the mean hypotension probability is calculated using the Monte Carlo samples inferred by Bayesian inference, and the uncertainty range of the mean hypotension probability is calculated.

Images