CN-121971057-A - Severe learning-based critical patient blood pressure trend monitoring and intervention decision-making system

CN121971057ACN 121971057 ACN121971057 ACN 121971057ACN-121971057-A

Abstract

The invention discloses a deep learning-based critical patient blood pressure trend monitoring and intervention decision-making system, which relates to the technical field of medical information processing and comprises the steps of obtaining continuous invasive arterial pressure waveforms, segmenting the continuous arterial pressure waveforms in a beat-to-beat mode, extracting average pressure and multidimensional high-frequency waveform characterization quantities in a beat-to-beat mode, calculating average pressure deviation quantity, high-frequency retention quantity and average collapse sign indexes by combining preamble analysis window data, inputting a time sequence deep learning network to obtain pseudo trend probability, further obtaining corrected average pressure and real trend slopes, respectively calculating link check priority and real low-pressure dry pre-priority, and outputting auxiliary intervention decisions. The invention can distinguish link distortion from real blood pressure deterioration, and improve the accuracy of trend monitoring and intervention decision.

Inventors

LI SHAOJI
WANG TANGYU
LIU FANGYA
CHEN CHUBING
LI SHAOXIONG
JIA LI
WANG SHIJUN
WANG CHUNZE
XING QINGTAO
LI JIAZHONG

Assignees

海南聚能科技创新研究院有限公司
陵水黎族自治县人民医院(陵水黎族自治县人民医院医共体总院)

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (9)

1. The blood pressure trend monitoring and intervention decision-making system for severe patients based on deep learning is characterized by comprising the following components: the waveform segmentation module is used for acquiring continuous invasive arterial pressure waveforms and segmenting the continuous invasive arterial pressure waveforms by pulse, and extracting the pulse-by-pulse average pressure and the multidimensional high-frequency characterization quantity of heart pulses; The reference construction module is used for determining a voltage-sharing reference and a high-frequency characteristic reference according to the preamble analysis window data, combining the beat-by-beat average voltage-sharing and multidimensional high-frequency characterization quantity, and calculating voltage-sharing deviation quantity and high-frequency retention quantity; The index calculation module is used for calculating the average collapse symptom index based on the average voltage deviation and the high-frequency reserved quantity; The model correction module is used for inputting the equalizing deviation, the high-frequency retention and the average collapse sign index into the time sequence deep learning network to obtain a pseudo trend probability, correcting the average pressure of each beat based on the pseudo trend probability and obtaining corrected equalizing and real trend slope; the auxiliary decision module is used for respectively calculating the link checking priority and the real low-voltage dry pre-priority based on the average collapse sign index, the pseudo trend probability, the correction voltage equalizing and the real trend slope, and outputting an auxiliary intervention decision.
2. The deep learning based critical patient blood pressure trend monitoring and intervention decision making system of claim 1, wherein extracting the beat-to-beat mean pressure and the multidimensional high frequency characterization quantity of the heart beat comprises: performing beat-to-beat segmentation on the continuous invasive arterial pressure waveform according to two adjacent diastolic valley points to obtain waveform data of target heart beat; Calculating an integral average value of waveform data in a heart beat period as a beat-to-beat average voltage value; the extracted waveform data has very poor amplitude over the heart cycle, as beat-to-beat pulse pressure; Extracting the maximum value of the time derivative of the waveform data as the sharpness of the gradual contraction rise; Extracting the relative amplitude between the first local minimum point pressure and the first local rising peak value after the contraction peak to be used as the beat-to-beat dicrotic notch significance; if the first local rise peak is not detected, taking the beat-to-beat dicrotic notch significance as zero; The beat-to-beat pulse pressure, the beat-to-beat shrinkage rising sharpness and the beat-to-beat dicrotic notch significance are jointly constructed into a multidimensional high-frequency characterization quantity.
3. The deep learning based critical patient blood pressure trend monitoring and intervention decision making system of claim 2, wherein the determining of the pressure equalizing reference and the high frequency characteristic reference according to the preamble analysis window data, and the calculating of the pressure equalizing deviation and the high frequency retention by combining the beat-to-beat pressure equalizing and the multidimensional high frequency characteristic quantity, comprises: Calculating historical beat-to-beat average pressure, historical beat-to-beat pulse pressure, historical beat-to-beat shrinkage rising sharpness and midrange of the notch significance of the historical beat-to-beat dicrotic beat in the preamble analysis window, and correspondingly obtaining a pressure-equalizing reference, a pulse pressure reference serving as a high-frequency characteristic reference, a rising sharpness reference and a notch significance reference; Calculating the proportion of the absolute difference between the beat-by-beat average voltage and the voltage-sharing reference relative to the voltage-sharing reference to obtain the voltage-sharing deviation; and respectively calculating the ratio of the pulse pressure, the ascending sharpness of the contraction of the beat and the notch significance of the beat-to-beat weight relative to the corresponding high-frequency characteristic reference, correspondingly obtaining the pulse pressure retention, the ascending sharpness retention and the notch significance retention, and jointly using the pulse pressure retention, the ascending sharpness retention and the notch significance as the high-frequency retention.
4. The deep learning based critical patient blood pressure trend monitoring and intervention decision making system of claim 3, wherein calculating a collapse syndrome index based on the pressure equalization deviation and the high frequency retention comprises: Calculating the mean value of the pulse pressure retention, the rising sharpness retention and the notch significance retention to obtain a high-frequency retention mean value; When the high-frequency retention mean value is smaller than one, taking a difference value between one and the high-frequency retention mean value as a high-frequency collapse term, otherwise, taking the value of the high-frequency collapse term as zero; taking the opposite number of the equalizing deviation as an index to perform natural index operation to obtain equalizing stability; taking the product of the voltage-sharing stability term and the high-frequency collapse term as a mean-stability collapse symptom index.
5. The deep learning-based critical patient blood pressure trend monitoring and intervention decision making system of claim 4, wherein the inputting of the equalization deviation, the high frequency retention and the equalization collapse sign index into the time sequence deep learning network, the obtaining of the pseudo trend probability, comprises: The method comprises the steps of forming a time sequence input sequence by using a pressure equalizing deviation amount, a pulse pressure retention amount, a rising sharpness retention amount, a notch significance retention amount and a collapse homogenizing index corresponding to a plurality of continuous heartbeats; And inputting the time sequence input sequence into a time sequence deep learning network comprising a gating circulation unit, and outputting the probability that the current abnormality belongs to a pseudo trend caused by pressure measurement link distortion.
6. The deep learning based critical patient blood pressure trend monitoring and intervention decision making system of claim 5, wherein correcting the beat-to-beat average pressure based on the pseudo trend probability to obtain corrected pressure equalization and true trend slope comprises: Taking the difference value between the first trend probability and the pseudo trend probability as the real probability, adding the product of the real probability and the average voltage of every beat and the product of the pseudo trend probability and the average voltage reference to obtain the corrected average voltage; And calculating the difference between the correction equalizing voltage of the current heart beat and the correction equalizing voltage of the previous heart beat, and taking the ratio of the difference to the adjacent heart beat time interval as a real trend slope.
7. The deep learning based critical patient blood pressure trend monitoring and intervention decision making system of claim 6, wherein calculating the link check priority and the true low pressure stem pre-priority based on the average collapse syndrome index, the pseudo trend probability, the correction pressure equalization, and the true trend slope, respectively, comprises: Taking the product of the pseudo trend probability and the average collapse sign index as the link checking priority; Calculating a difference value between the voltage equalizing reference and the correction voltage equalizing, if the difference value is larger than zero, taking the difference value as a pressure drop amplitude, otherwise, taking the pressure drop amplitude to be zero; taking the ratio of the pressure drop amplitude to the pressure equalizing reference absolute value as the average pressure drop degree; Obtaining the opposite number of the true trend slope, taking the opposite number as the descending slope if the opposite number is larger than zero, otherwise taking the descending slope to be zero; Taking the ratio of the descending slope to the absolute value of the true trend slope as the trend deterioration degree; taking the product of the real probability, the pressure equalizing decline degree and the trend deterioration degree as the real low-pressure dry pre-priority.
8. The deep learning based critical patient blood pressure trend monitoring and intervention decision making system of claim 7, wherein outputting the supplemental intervention decision comprises: outputting an auxiliary intervention decision for checking the pressure measuring link with priority when the link checking priority is greater than the real low-voltage dry pre-priority; outputting an auxiliary intervention decision for preferentially executing blood pressure intervention evaluation when the real low-pressure dry pre-priority is greater than the link checking priority; When the true low-voltage dry pre-priority is equal to the link check priority, outputting an auxiliary intervention decision that maintains the current scheme and shortens the reexamine interval.
9. The deep learning-based critical patient blood pressure trend monitoring and intervention decision making method applied to the deep learning-based critical patient blood pressure trend monitoring and intervention decision making system as claimed in any one of claims 1 to 8, wherein the method comprises the following steps: s1, acquiring continuous invasive arterial pressure waveforms, segmenting the continuous invasive arterial pressure waveforms step by step, and extracting the average pressure and the multidimensional high-frequency characterization quantity of heart beat by step; s2, determining a voltage-sharing reference and a high-frequency characteristic reference according to the preamble analysis window data, and calculating a voltage-sharing deviation amount and a high-frequency retention amount by combining a beat-to-beat voltage-sharing and multi-dimensional high-frequency characteristic amount; s3, calculating a mean-stability collapse symptom index based on the mean-pressure deviation amount and the high-frequency retention amount; S4, inputting the equalizing deviation, the high-frequency retention and the equalizing collapse sign index into a time sequence deep learning network to obtain a pseudo trend probability, and correcting the average pressure of each beat based on the pseudo trend probability to obtain corrected equalizing and real trend slope; and S5, respectively calculating the link checking priority and the real low-voltage dry pre-priority based on the average collapse sign index, the pseudo trend probability, the correction voltage equalizing and the real trend slope, and outputting an auxiliary intervention decision.

Description

Severe learning-based critical patient blood pressure trend monitoring and intervention decision-making system Technical Field The invention relates to the technical field of medical information processing, in particular to a severe patient blood pressure trend monitoring and intervention decision-making system based on deep learning. Background In intensive care, continuous invasive arterial pressure monitoring is an important basis for evaluating the circulation state, tissue perfusion state and intervention response of a patient, and medical staff usually combine continuous waveforms, average pulse-by-pulse pressures and variation trends thereof to judge whether the patient has a blood pressure worsening risk or not, and determine whether a fluid replacement evaluation, vasoactive drug adjustment or further review monitoring link is required. For a critical patient with an arterial catheter for a long time, the monitoring data is not always determined by the blood dynamics state of the patient, and the conditions such as micro bubbles, local discount, local compression, partial blockage, fiber deposition and the like can gradually reduce the pressure transmission capacity of the pressure measuring link, so that the amplitude swing, the shrinkage rising edge and the later local details in the pulse pressure waveform are continuously weakened, and the average pressure level can not have obvious change of the synchronous amplitude within a certain time. Under the situation, how to timely distinguish abnormal changes caused by progressive distortion of a pressure measuring link from real blood pressure deterioration trend of a patient in the continuous blood pressure monitoring process of a severe patient and output a more reasonable intervention decision according to the abnormal changes and the real blood pressure deterioration trend of the patient becomes a technical problem to be solved. When the prior art is used for solving the problems, trend judgment, risk early warning or intervention prompt is generally directly carried out based on an original invasive arterial pressure waveform or an original blood pressure value, and a special recognition and diversion mechanism for a progressive distortion process of a pressure measuring link is lacked, so that a system is easy to misjudge the weakening of waveform details caused by link abnormality as the degradation of the real circulation state of a patient, thereby influencing the accuracy of trend analysis results. Meanwhile, the existing scheme focuses on judging the current blood pressure level or the short-time change, so that the cooperative analysis of the average pressure change degree and the waveform detail attenuation degree is difficult to combine with the recent individuation reference state of a patient, and the reasonable distinction between the priority checking of the pressure measuring link and the priority executing of the blood pressure intervention evaluation is difficult to be realized at the output level. As a result of this, the system may still trigger unnecessary low pressure treatment cues when monitoring link anomalies are dominant, but it is difficult to timely give more targeted auxiliary decisions when the actual blood pressure drop develops, thereby limiting the accuracy and clinical practicality of blood pressure trend monitoring and intervention decisions for critically ill patients. Disclosure of Invention The invention aims to solve the defects that in the prior art, the trend judgment accuracy is insufficient and the auxiliary intervention decision is easy to deviate due to the fact that the abnormal change of the pseudo blood pressure and the real blood pressure worsening trend of a patient caused by progressive distortion of a pressure measuring link are difficult to distinguish, and provides a severe patient blood pressure trend monitoring and intervention decision system based on deep learning. In order to solve the problems existing in the prior art, the invention adopts the following technical scheme: The blood pressure trend monitoring and intervention decision-making system for severe patients based on deep learning comprises: the waveform segmentation module is used for acquiring continuous invasive arterial pressure waveforms and segmenting the continuous invasive arterial pressure waveforms by pulse, and extracting the pulse-by-pulse average pressure and the multidimensional high-frequency characterization quantity of heart pulses; The reference construction module is used for determining a voltage-sharing reference and a high-frequency characteristic reference according to the preamble analysis window data, combining the beat-by-beat average voltage-sharing and multidimensional high-frequency characterization quantity, and calculating voltage-sharing deviation quantity and high-frequency retention quantity; The index calculation module is used for calculating the average collapse symptom index based on the ave