CN-121971045-A - Deep learning-based high-quality pulse end-to-end extraction method and system

CN121971045ACN 121971045 ACN121971045 ACN 121971045ACN-121971045-A

Abstract

The invention belongs to the technical field of medical information identification, and particularly discloses a high-quality pulse end-to-end extraction method and system based on deep learning. The method comprises the steps of collecting continuous pulse wave original signals, conducting real-time pretreatment and standardization, inputting the continuous pulse wave original signals into a trained multi-scale time sequence attention network model to obtain quality classification probability of pulse wave fragments, calculating quality scores of the pulse wave fragments according to the quality classification probability, determining a high quality threshold value, screening out the high quality pulse wave fragments according to the high quality threshold value, conducting clustering analysis according to collected pressure values, determining an optimal pressure interval, and finally selecting pulse waves with highest quality from the optimal pressure intervals to form a final output high quality pulse wave set. The invention can objectively, accurately and efficiently extract the pulse wave set with optimal quality from the original pulse wave signals. The invention can be widely applied to the extraction of pulse.

Inventors

YANG RUOZHANG
WANG DAN
QIN YI

Assignees

河北普茵智能电子有限公司

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (9)

1. The high-quality pulse end-to-end extraction method based on deep learning is characterized by comprising the following steps of: s1, collecting continuous pulse wave original signals at a plurality of pressure points; s2, carrying out real-time preprocessing and standardization on the acquired pulse wave signals; s3, inputting the preprocessed pulse wave plate segments into the trained multi-scale time sequence attention network model to obtain quality classification probability of each segment; S4, calculating the quality scores of the pulse wave fragments according to the quality classification probability, and dynamically determining a high quality threshold value based on the quality scores of all the fragments; S5, screening out high-quality pulse wave fragments according to the high-quality threshold value, and performing cluster analysis according to the acquired pressure values to determine an optimal pressure interval; s6, selecting the pulse wave with the highest quality from each optimal pressure interval to form a final output high-quality pulse wave set.
2. The method for end-to-end extraction of high quality pulse based on deep learning according to claim 1, wherein the training process of the multi-scale time-series attention network model in step S3 comprises the following steps performed in sequence: s31, constructing a pulse wave quality training data set of multi-expert collaborative labeling; s32, carrying out standardized processing, segmentation processing and data enhancement on the training data; S33, constructing a multi-scale time sequence attention network model, and optimizing the model convergence process and the final performance by adopting a training method comprising weighted cross entropy loss and cosine annealing learning rate scheduling.
3. The method of end-to-end extraction of high quality pulse based on deep learning according to claim 2, wherein the step S31 comprises: collecting massive pulse wave raw data from different individuals under different pressures, wherein the massive pulse wave raw data comprise people with different ages, sexes, physique and health conditions; the expert group carries out independent labeling comprising waveform stability, form normalization, noise level and diagnostic value on each pulse wave segment by adopting a double-blind method, and finally gives three classification labels of high, medium and low comprehensively; and (3) controlling data quality, namely reserving samples with Kappa coefficients more than 0.8 based on consistency of labels among experts.
4. The deep learning based high quality pulse end-to-end extraction method of claim 2, wherein the data enhancement includes one or more of time warping, amplitude scaling, adding noise, random clipping, and channel dropping.
5. The deep learning based high quality pulse end-to-end extraction method of any one of claims 1-4, wherein the multi-scale time-series attention network model comprises: the multi-scale feature extraction module is used for synchronously extracting macro morphological features, mesoscopic local features and microcosmic detail features of the pulse wave signals through convolution kernels of different scales which are arranged in parallel, and carrying out feature fusion; the time sequence dependency modeling module is used for capturing a long-range time sequence dependency relationship in the pulse wave sequence based on the two-way long-short-term memory network and the attention mechanism; a spatial feature enhancement module for enhancing the representation capability of the features through the residual convolution block and the extrusion excitation module, and adaptively recalibrating the channel feature response through a channel attention mechanism; And the output classification module is used for outputting the probability that the pulse wave plate segment belongs to each quality class.
6. The deep learning-based high quality pulse end-to-end extraction method of claim 5, wherein the output classification module outputs probability distributions for three quality classes through global averaging pooling, full-connected layer and Softmax activation functions.
7. The deep learning-based high quality pulse end-to-end extraction method according to claim 2, wherein the preprocessing and normalization in step S2 comprises: Baseline correction, namely removing baseline drift caused by respiration and motion by adopting a real-time self-adaptive baseline estimation and elimination algorithm; Slight filtering, namely, removing obvious high-frequency noise and retaining main characteristics of an original waveform; Signal standardization, namely carrying out real-time z-score standardization on the signal to ensure that the signal is consistent with training data distribution; segmentation processing, namely segmenting continuous signals acquired in real time into segments with the same length as training, and preparing an input model.
8. The deep learning-based high-quality pulse end-to-end extraction method according to claim 1, wherein after determining the optimal pressure interval in step S5, the information of the optimal pressure interval is fed back for optimizing the subsequent pulse wave acquisition pressure.
9. A pulse end-to-end extraction system for realizing the high-quality pulse end-to-end extraction method based on deep learning as claimed in any one of claims 1 to 8, which is characterized by comprising a pulse wave acquisition module, a preprocessing module, a quality judging module, a decision control module and a pressure control module which are connected in sequence; The pulse wave acquisition module is used for acquiring original pulse wave signals at a plurality of preset pressure points; the preprocessing module is used for carrying out baseline correction, filtering and standardization processing on the acquired original pulse wave signals; The quality judging module is internally provided with a trained multi-scale time sequence attention network model and is used for carrying out end-to-end quality classification on the preprocessed pulse wave segments and outputting quality scores; the decision control module is used for dynamically determining a high quality threshold value, identifying an optimal acquisition pressure interval, selecting a final high quality pulse wave set and generating a pressure control instruction according to the quality score output by the quality judging module; and the pressure control module is used for adjusting the pressure applied to the pulse acquisition part according to the pressure control instruction generated by the decision control module.

Description

Deep learning-based high-quality pulse end-to-end extraction method and system Technical Field The invention belongs to the technical field of medical information identification, and particularly relates to a high-quality pulse end-to-end extraction method and system based on deep learning. Background Pulse diagnosis is one of the most important diagnostic methods in TCM. With the development of technology, automatic pulse taking devices are widely used for pulse wave acquisition. However, the acquired pulse wave signal quality is uneven and is easily affected by various factors such as measurement pressure, individual difference, motion artifact, equipment noise and the like. Therefore, the high-quality pulse wave fragments are intelligently and automatically extracted from the continuously acquired sequences, and are the precondition for subsequent reliable pulse analysis and diagnosis. Currently, the following three main methods for quality assessment of pulse waves exist: The first category is a manual visual discrimination method. An operator directly observes the waveform on the display device and empirically determines the continuity, baseline stationarity, main peak sharpness, and the like. The method is highly dependent on the experience of operators, has strong subjectivity, low efficiency and non-uniform standard, and is difficult to integrate into an automatic system. And the second category is a statistical learning method based on manual characteristics. The method firstly extracts the characteristics of time domain, frequency domain and the like which are designed manually from pulse waves, and then uses a Support Vector Machine (SVM), random forest and other traditional machine learning classifiers to judge. The method has the defects that the quality of characteristic engineering is seriously depended, the complex pulse wave shape and quality connotation are difficult to comprehensively and optimally feature by manual characteristics, and the flow is complicated. And third category, shallow neural network method. Some technologies adopt a simple Convolutional Neural Network (CNN) or a cyclic neural network (RNN) for processing, but the network structure is shallow, the capability of feature extraction and time sequence modeling is limited, the learning of complex nonlinear modes and long-range dependency relationships in pulse waves is insufficient, and pretreatment and feature engineering can still be relied on. In summary, the prior art generally has the defects of strong subjectivity, complex characteristic engineering, limited model expression capability, insufficient time sequence relation modeling, poor generalization capability across individuals and the like. Therefore, there is an urgent need for a method ‌ that can automatically, objectively, accurately, and efficiently extract high-quality fragments from the original pulse wave signal. ‌ ‌ A Disclosure of Invention The invention aims to provide a high-quality pulse end-to-end extraction method based on deep learning, which can objectively, accurately and efficiently extract an optimal quality pulse wave set from an original pulse wave signal directly and optimize an acquisition process through closed loop feedback; A second object of the present invention is to provide a pulse end-to-end extraction system for implementing a high quality pulse end-to-end extraction method based on deep learning. The technical scheme adopted by the invention for realizing the purposes is as follows: A high quality pulse end-to-end extraction method based on deep learning, comprising the following steps in sequence: s1, collecting continuous pulse wave original signals at a plurality of pressure points; s2, carrying out real-time preprocessing and standardization on the acquired pulse wave signals; s3, inputting the preprocessed pulse wave plate segments into the trained multi-scale time sequence attention network model to obtain quality classification probability of each segment; S4, calculating the quality scores of the pulse wave fragments according to the quality classification probability, and dynamically determining a high quality threshold value based on the quality scores of all the fragments; S5, screening out high-quality pulse wave fragments according to the high-quality threshold value, and performing cluster analysis according to the acquired pressure values to determine an optimal pressure interval; s6, selecting the pulse wave with the highest quality from each optimal pressure interval to form a final output high-quality pulse wave set. By way of limitation, the training process of the multi-scale time-series attention network model in step S3 includes the following steps performed in order: s31, constructing a pulse wave quality training data set of multi-expert collaborative labeling; s32, carrying out standardized processing, segmentation processing and data enhancement on the training data; S33, constructing a multi-scale time se