Search

CN-122004811-A - Noninvasive continuous blood pressure waveform prediction method and device, electronic equipment and storage medium

CN122004811ACN 122004811 ACN122004811 ACN 122004811ACN-122004811-A

Abstract

The invention provides a non-invasive continuous blood pressure waveform prediction method and device based on morphological perception and blood pressure guidance normalization. Firstly, acquiring a preprocessed peripheral physiological signal sequence, and extracting global features by using a bidirectional state space model. And then the systolic pressure value, the diastolic pressure value and the normalized waveform are predicted in parallel through a double-head output framework. The core is to utilize the predicted value as an anchor point to perform de-normalization transformation on the normalized waveform to reconstruct the physiological waveform. The block-level morphological consistency loss is introduced in model training, self-adaptive blocking is performed through frequency domain analysis, and accurate constraint on waveform trend, fluctuation mode and base line is realized based on dynamic weighting of gradient norms. The method effectively solves the problems of inconsistent physiological contradiction of predicted values and waveform forms, low long-sequence modeling efficiency, local form distortion and the like, remarkably improves the physiological rationality and the form fidelity of continuous blood pressure monitoring, and has extremely high clinical application value.

Inventors

  • LI YE
  • Qiu Weixiu
  • Miao fen

Assignees

  • 中国科学院深圳先进技术研究院

Dates

Publication Date
20260512
Application Date
20260209

Claims (10)

  1. 1. A method for predicting a noninvasive continuous blood pressure waveform, comprising the steps of: acquiring a plurality of peripheral physiological signal sequences acquired synchronously, and preprocessing the peripheral physiological signal sequences; inputting the preprocessed peripheral physiological signal sequence into a pre-trained feature extraction module, and extracting global features of the peripheral physiological signal sequence through the feature extraction module; processing the global features by using a prediction model, and outputting a systolic pressure value, a diastolic pressure value and a normalized blood pressure waveform in parallel; And performing de-normalization transformation on the normalized blood pressure waveform by using the systolic pressure value and the diastolic pressure value as physical anchor points, and reconstructing to obtain a continuous blood pressure waveform with physiological physical dimensions.
  2. 2. The method according to claim 1, wherein the predictive model is obtained by the training steps of: acquiring a training set containing a sample physiological signal and a corresponding real blood pressure waveform; And in the training process, constraining parameters of the prediction model by adopting a morphological consistency loss function, wherein the morphological consistency loss function comprises a correlation loss for constraining the predicted waveform and the trend of the real waveform, a variance loss for constraining the waveform amplitude fluctuation mode and a mean loss for correcting local baseline deviation.
  3. 3. The method of claim 2, wherein in the training step, weights of the correlation loss, the variance loss, and the mean loss in a total loss function are dynamically adjusted according to a gradient norm generated during back propagation for each sub-loss, and a weight magnitude is inversely proportional to the gradient norm.
  4. 4. A method according to claim 2 or 3, further comprising an adaptive partitioning step prior to calculating the morphological consistency loss function: identifying the main frequency of the real blood pressure waveform through frequency domain analysis, and calculating the corresponding cardiac cycle length according to the main frequency; Dividing the sample physiological signal and the real blood pressure waveform into a plurality of non-overlapping local waveform blocks according to the cardiac cycle length; the correlation loss, the variance loss, and the mean loss are calculated separately at the local waveform block level.
  5. 5. A method according to any of claims 1-3, wherein the feature extraction module captures causal dependencies of physiological signals from past to present through parallel forward processing branches, captures future context information of sequences through backward processing branches, and finally performs a stitching fusion of outputs of the two-way branches using a two-way state space model.
  6. 6. A method according to any one of claims 1-3, characterized in that the specific operation of the denormalization transformation is: mapping the upper numerical limit in the normalized blood pressure waveform to the systolic pressure value, mapping the lower numerical limit to the diastolic pressure value, and linearly stretching the middle fluctuation amplitude of the waveform according to the difference value between the systolic pressure value and the diastolic pressure value.
  7. 7. The method of any of claims 1-3, wherein the plurality of peripheral physiological signals comprises an electrocardiographic signal, a photoplethysmographic signal, and an arterial blood pressure waveform signal, and wherein the preprocessing comprises: filtering the electrocardiosignal by adopting a band-pass filter to remove baseline drift; filtering the photo-capacitive pulse wave signal by adopting a band-pass filter to inhibit physiological artifacts; For arterial blood pressure waveform signals as a supervision target, the original acquisition form is kept without filtering treatment so as to keep the dicrotic notch characteristics in descending branches.
  8. 8. A non-invasive continuous blood pressure waveform predicting apparatus, comprising: The data preprocessing module is used for acquiring and preprocessing a peripheral physiological signal sequence; the feature coding module is used for extracting global features of the peripheral physiological signal sequence through a bidirectional state space model; The prediction reconstruction module is used for predicting the blood pressure value and the normalized waveform in parallel, and performing normalization transformation on the normalized waveform by utilizing the blood pressure value, and reconstructing to obtain a continuous blood pressure waveform.
  9. 9. An electronic device comprising a memory and a processor, characterized in that the memory has stored thereon a computer program executable on the processor, which processor, when executing the computer program, realizes the steps of the method according to any of claims 1 to 7.
  10. 10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Description

Noninvasive continuous blood pressure waveform prediction method and device, electronic equipment and storage medium Technical Field The invention relates to the technical field of biomedical signal processing and deep learning, in particular to a non-invasive continuous blood pressure waveform prediction method, a non-invasive continuous blood pressure waveform prediction device, an electronic device and a storage medium based on morphological perception and blood pressure guiding de-normalization. Background Hypertension is a major risk factor for cardiovascular disease. Traditional cuff blood pressure measurements can only provide intermittent single point readings, cannot capture dynamic blood pressure fluctuations, and invasive arterial catheters, while gold standard, have a high risk of complications. Therefore, non-invasive continuous blood pressure monitoring based on photoplethysmography (PPG) and Electrocardiogram (ECG) is becoming a research hotspot. The current mainstream noninvasive continuous blood pressure waveform prediction method is mainly based on a deep learning model, and is mapped from PPG/ECG to a blood pressure waveform through an end-to-end mode, and representative technologies include waveform reconstruction (such as U-Net, V-Net, waveNet and the like) based on Convolutional Neural Network (CNN) extraction characteristics, and a cyclic neural network (RNN) method for capturing time series dependency (such as LSTM and BiLSTM) or long sequence data dependency based on a transducer. However, the existing schemes have various limitations in practical applications. First, the problems of supervision waste and physiological inconsistencies resulting from task separation are quite prominent. Existing methods typically model the waveform prediction of blood pressure and the numerical estimate of systolic/diastolic blood pressure as two independent tasks. This separation results in the model not being able to be supervised using the extreme point information in the waveform, and there is often a physiological inconsistency between the predicted waveform and the estimated blood pressure value, for example, the peak value of the waveform output by the model does not coincide with the predicted systolic pressure value, reducing the clinical reliability. Secondly, the prior art has insufficient modeling of the waveform structure, often resulting in morphological distortion. The mainstream method mostly adopts a self-attention mechanism based on a transducer to process long-sequence data, but when processing long-time signals (such as 120 seconds/15000 points), the calculation complexity is very high, which results in extremely high memory occupation and reasoning delay, and is difficult to be deployed on portable equipment. Meanwhile, the prior art mainly relies on point-to-point loss functions such as Mean Square Error (MSE) or Mean Absolute Error (MAE), and the like, and the functions assume that each time point is independent, so that the inherent periodic morphological structure of the blood pressure waveform is ignored. The modeling mode easily causes smooth transition, peak value deviation or key characteristic loss of the generated waveform, and particularly cannot retain morphological characteristics with physiological rationality such as dicrotic notch, peak value in systole and the like, so that the generated waveform loses clinical diagnostic value. In summary, the existing noninvasive continuous blood pressure waveform prediction technology is difficult to realize physiological coupling of a blood pressure value and a waveform form while ensuring long-sequence processing efficiency, and the generated waveform is poor in physiological rationality and form fidelity due to lack of effective constraint on local periodic structural features. Disclosure of Invention The invention aims to solve the technical problems of the prior art, provides a noninvasive continuous blood pressure waveform prediction method based on morphological perception and blood pressure guiding de-normalization, and solves the technical problems of low modeling efficiency of a long sequence, inconsistent predicted values and waveform morphology physiology, poor waveform morphology fidelity and the like in the existing noninvasive continuous blood pressure waveform prediction technology. The technical scheme adopted by the invention is characterized in that firstly, a bidirectional state space model with linear calculation complexity is used as a core of feature extraction to realize efficient capturing of global time dependence of long-term physiological signals, secondly, a double-head output structure is designed and a blood pressure guided de-normalization strategy is introduced to deeply couple blood pressure value estimation with a waveform prediction task, the predicted pressure value is used as an anchor point to force a range of amplitude of a constraint waveform, so that physiological consistency is ensure