Dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters

Abstract

The present invention discloses a dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters, which includes: a physiological parameter sensing subsystem, an early warning model updating subsystem, a dynamic early warning subsystem, a post-traumatic hemorrhagic shock onset judgment subsystem and a silent subsystem; the dynamic early warning subsystem is used to calculate the probability of occurrence of post-traumatic hemorrhagic shock in patients, and the post-traumatic hemorrhagic shock onset judgment subsystem will correct the predicted probability value according to the medical environment data of the location where it is located. The silent subsystem controls the early warning result of the silent subsystem controlling the dynamic early warning subsystem according to the differences among the test data, the training data and the label data. The present invention only uses common non-invasive parameters and does not require laboratory data. It can be used in remote areas, public health emergencies and battlefield frontline situations, eliminating the damage to individual patients caused by frequent collection of laboratory parameters, and reducing the system use cost and early warning error probability.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology and medical health, and in particular to a dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters. Background Art

[0002] Hemorrhagic shock is a type of hypovolemic shock that results from severe blood loss leading to insufficient oxygen delivery at the cellular level. If blood loss is not controlled, the patient will die rapidly, and deaths from hemorrhagic shock caused by trauma account for 10 to 40% of all deaths from trauma. Retrospective studies of multiple databases have shown that nearly a quarter of patients admitted to the hospital after trauma will experience symptoms of hemorrhagic shock during their hospital stay. For most patients, the effects of shock on the body are initially reversible, but repeated or prolonged hypotension and the use of high-dose vasopressors may worsen the prognosis.

[0003] Medical staff intermittently evaluate the monitored patient vital signs and rely on individual physiological parameters to identify patients at risk of deterioration. These early warning systems do not use comprehensive patient information, so alarms are often inaccurate and prone to alarm fatigue. Low-precision early warning systems are also not conducive to clinicians identifying and interpreting relevant information, and it is difficult for nurses or doctors to continuously monitor or evaluate ICU patients. Low-frequency and low-precision early warning systems will be difficult to adapt to patients with acute changes in their conditions. Due to the lack of sufficiently accurate prediction tools, clinicians may resort to subjective judgment, which increases the risk of doctors taking action based on relevant information. Prophylactic transfusions or the use of 100% mechanical ventilation have a significant effect on reducing the mortality rate of hemorrhagic shock. If there is an early warning assessment system that can dynamically predict the probability of future hemorrhagic shock in post-traumatic patients, then timely preventive treatment plans can be provided for high-risk patients, thereby greatly reducing mortality and medical costs. Doctors in intensive care units can obtain a large amount of patient physiological data and measurement indicators from multiple monitoring systems, but relying solely on the limited ability of humans to process complex information will hinder the early identification of patients' deterioration, and conventional identification and early warning methods are difficult to monitor patients' conditions in real time. Deep learning techniques excel in analyzing complex signals in data-rich environments. The large amount of data collected in the ICU and the public availability of large medical critical care datasets such as the Medical Information Mart for Intensive Care III (MIMIC III) database, eICU, and AmsterdamUMC are key to developing the use of machine learning in this environment.

[0004] Patent CN201910570791 provides a time series prediction method for traumatic hemorrhagic shock injuries, which can extract traumatic hemorrhagic shock injury data from a database and process the traumatic hemorrhagic shock injury data. Its functions include: processing data outliers, linear filling and cluster filling of data; designing stepped indicators for processed data; applying the stepped indicator results and different types of classifiers to build a prediction model, and predicting the results after a preset time length through the prediction model. This invention uses indicators that can be monitored in real time to implement real-time dynamic prediction and warning based on time series for traumatic hemorrhagic shock, but does not reserve time for clinical intervention and does not perform interpretability analysis on the model.

[0005] In addition, in the field of medical and health big data, patient clinical data can be divided into cross-sectional data with only one cross-section and time series data with multiple cross-sections. The latter has the characteristics of large amount of information and trend changes, so the time series prediction accuracy is higher than the cross-sectional prediction, and can realize sliding prediction, real-time monitoring and early warning of the disease. However, due to the different measurement indicators of the wounded, the different measurement times of the wounded indicators, and the fact that most laboratory indicators will not be measured multiple times in a short period of time, the problem of sparse and missing medical data is extremely serious. The existing technology has at least the following problems:

[0006] (1) There is no mature data filling technology system, and most of them use mean filling or linear filling. The filling method is single, and after filling, there are still problems such as poor data quality and large difference from the real data;

[0007] (ii) Existing technical methods mostly use cross-sectional prediction methods, such as taking the average of the data to obtain the cross-sectional data and then making predictions. The results obtained are “final” and cannot achieve rolling prediction and real-time monitoring of the disease;

[0008] (3) Existing methods use a small number of time series for prediction and only select vital signs that have low measurement costs and high frequency, such as heart rate and blood pressure, which results in poor prediction results;

[0009] (IV) Existing early warning methods do not adopt relevant processing to reduce the false alarm rate. Frequent model false alarms lead to alarm fatigue among medical staff, which in turn affects the active clinical treatment of patients in critical conditions.

[0010] Although the existing publicly available technologies can predict the possible onset of post-traumatic hemorrhagic shock in patients through their physiological parameters, they lack the clinical intervention time reserved between the running time point of the warning model and the onset warning time point, and lack effective methods to reduce the false alarm rate. Summary of the invention

[0011] In view of the problems existing in the existing dynamic early warning system for post-traumatic hemorrhagic shock, the present invention discloses a dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters, which can predict the possibility of post-traumatic hemorrhagic shock in patients in the future within multiple time scales, reduce the false alarm rate and achieve effective early warning.

[0012] The present invention discloses a post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters, which comprises: a physiological parameter sensing subsystem, a warning model updating subsystem, a dynamic warning subsystem, a post-traumatic hemorrhagic shock onset judgment subsystem and a silent subsystem;

[0013] The physiological parameter sensing subsystem is connected to the dynamic early warning subsystem, the dynamic early warning subsystem is connected to the early warning model updating subsystem and the post-traumatic hemorrhagic shock onset judgment subsystem respectively, and the silent subsystem is connected to the post-traumatic hemorrhagic shock onset judgment subsystem, the dynamic early warning subsystem and the early warning model updating subsystem respectively.

[0014] The physiological parameter sensing subsystem is used to monitor and collect the patient's physiological parameter data information in real time, pre-process the collected patient's physiological parameter data, and then send the pre-processed data to the dynamic early warning subsystem.

[0015] The patient's physiological parameter data information includes the patient's conventional non-invasive physiological parameters and the patient's physiological parameter time series data within the learning window.

[0016] The preprocessing of the collected patient physiological parameter data includes, for the discrete sampling values ​​of the collected patient physiological parameters, using the discrete sampling values ​​to form a sampling vector, calculating the cross-correlation matrix of the sampling vector, processing the cross-correlation matrix using a feature extraction method to obtain its eigenvalue vector, and using the eigenvalue vector to weight the sampling vector to obtain a smoothed value of the patient's physiological parameter as the preprocessed data, thereby effectively reducing the interference of the collected wild value and other error values ​​on the subsequent early warning judgment process, and reducing the false alarm probability of the early warning judgment process.

[0017] The preprocessing of the collected patient physiological parameter data includes: the sampling vector composed of discrete sampling values ​​of the patient physiological parameters collected within a period of time is expressed as [x1, x2, ..., x N ], N is the number of discrete sampling values ​​of the patient's physiological parameters collected within a period of time, the mutual correlation matrix C of the sampling vector is calculated, and the eigenvalue decomposition of the mutual correlation matrix C is performed to obtain:

[0018] C=VDV H ,

[0019] Among them, C is the eigenvector matrix, D is the eigenvalue matrix, the diagonal elements of the matrix D are normalized and used as weight vectors, the sampling vectors are weighted summed, and the smoothed values ​​of the patient's physiological parameters are obtained as preprocessed data.

[0020] The dynamic early warning subsystem is used to receive and clean the conventional non-invasive physiological parameters of the patient collected by the physiological parameter sensing subsystem and the physiological parameter time series data within the patient learning window, calculate the original probability of the patient suffering from post-traumatic hemorrhagic shock within the prediction time window, and then calculate the probability of the patient suffering from post-traumatic hemorrhagic shock within the future prediction time window according to the conventional non-invasive physiological parameters collected by the physiological parameter sensing subsystem and the physiological parameter time series data within the patient learning window, obtain the predicted probability value, and send the predicted probability value to the post-traumatic hemorrhagic shock onset judgment subsystem. The post-traumatic hemorrhagic shock onset judgment subsystem will correct the predicted probability value according to the medical environment data of its location and the doctor's treatment experience evaluation value. If the corrected predicted probability value is greater than the first predicted probability threshold, the post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters judges that the patient will suffer from post-traumatic hemorrhagic shock within the future prediction time window. If the corrected predicted probability value is less than or equal to the second predicted probability threshold, the system judges that the patient will not suffer from post-traumatic hemorrhagic shock within the future prediction time window.

[0021] The early warning model updating subsystem monitors the medical environment where the system is located. If the early warning model updating subsystem detects changes in the medical environment where the system is located, the subsystem uses an incremental learning method to update the weights of the deep learning model in the dynamic early warning subsystem based on the patient's routine non-invasive physiological parameters collected by the physiological parameter sensing subsystem and the prognosis information including the patient's mortality, post-traumatic hemorrhagic shock, and length of hospitalization, so that the early warning results of the system adapt to the medical environment where the system is located.

[0022] The silent subsystem determines the differences among the test data, training data and label data of the dynamic early warning of post-traumatic hemorrhagic shock of the intelligent dynamic warning subsystem. If there is a significant difference between the three types of data, the silent subsystem controls the dynamic warning subsystem not to output the warning result, and collects the warning result data of the dynamic warning subsystem within a period of time, extracts the warning result data with the maximum prediction probability within the period of time, and outputs it to the post-traumatic hemorrhagic shock onset judgment subsystem.

[0023] The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including:

[0024] The data of the early warning model update subsystem is regarded as a stationary random process. Corresponding autoregressive-moving average models, namely, ARMA models, are established for the test data, label data and training data, respectively, and the first ARMA model, the second ARMA model and the third ARMA model are obtained respectively. The cross-correlation matrix of the coefficients of the three ARMA models is calculated, and the maximum eigenvalue of the cross-correlation matrix is ​​calculated. The maximum eigenvalue is used to distinguish the differences between the test data, label data and training data.

[0025] When the maximum eigenvalue is greater than the difference discrimination threshold, it is determined that significant differences occur among the test data, label data and training data of the early warning model updating subsystem.

[0026] The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including:

[0027] The silent subsystem uses the training data and the deep learning model to predict the predicted value of the labeled data, and uses the credibility evaluation method to calculate the credibility between the test data, training data and the predicted value of the labeled data. When the credibility is lower than the credibility threshold, it is determined that there is a significant difference between the warning data, label data and training data of the early warning model update subsystem.

[0028] The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including:

[0029] A sample set is formed by using training data and label data prediction values. According to the label category to which each test data belongs, the sample set is divided into a first reference sample set and a second reference sample set corresponding to the test data. The first reference sample set is a set composed of samples corresponding to the label category to which the test data belongs, and the second reference sample set is a set composed of other samples in the sample set excluding the first reference sample set. A consistency measurement calculation function is used to calculate the distance between each sample in the first reference sample set and the reference sample set to obtain a first sample set distance value vector. A consistency measurement calculation function is used to calculate the distance between each test data and its corresponding first reference sample set and second reference sample set, respectively, to obtain the first set distance value and the second set distance value of the test data. The first set distance value of the test data is sorted in ascending order in the first sample set distance value vector to obtain a sequence value. The ratio of the sequence value to the total number of samples in the first reference sample set is calculated as the credibility of the test data. If the credibility of the test data is lower than the credibility threshold, it is determined that there is a significant difference between the warning data, label data and training data of the warning model update subsystem; the credibility of each test data is calculated and compared with the credibility threshold.

[0030] The deep learning model may adopt a stacked deep learning model, a convolutional neural network module, etc.

[0031] The stacked deep learning model includes a convolutional layer, a bidirectional long short-term memory layer, and a self-attention layer.

[0032] The dynamic early warning subsystem uses an adaptive interpolation method to interpolate missing patient physiological parameter time series data within the minimum time unit, and the specific steps include:

[0033] S1, calculate the median Midparam and quartile IQRidparam of the sampling frequency of each physiological parameter from admission to discharge for each patient;

[0034] S2, selecting physiological parameters with missing values ​​of patients with missing data and recording the parameters as param parameters, and sorting all the collected param parameters in ascending order of collection time, and selecting the time position of the empty value or abnormal value in the collected param parameters as the selected position;

[0035] S3, interpolate data for the selected position in the following priority order: first, take the median of the valid values ​​of the first (Midparam+IQRidparam) sampling time lengths of the selected position as the data interpolation value of the selected position; secondly, take the median of the valid values ​​of the first (Midparam+2*IQRidparam) sampling time lengths of the selected position as the data interpolation value of the selected position; thirdly, take the median of the valid values ​​of the collected param parameters as the data interpolation value of the selected position; finally, take the median of the param parameters recorded in the database as the data interpolation value of the selected position.

[0036] The post-traumatic hemorrhagic shock onset judgment subsystem receives the original probability of the patient suffering from post-traumatic hemorrhagic shock within the prediction time window. The post-traumatic hemorrhagic shock onset judgment subsystem obtains the classification threshold D of the stacked deep learning model based on the sensitivity and specificity of the stacked deep learning model in the dynamic early warning subsystem on the test set data, so that when the classification threshold is D, the difference between the sensitivity and specificity of the stacked deep learning model on the test set data is minimized.

[0037] The post-traumatic hemorrhagic shock onset judgment subsystem modifies the probability value of post-traumatic hemorrhagic shock according to the medical environment data of the location and the doctor's treatment experience evaluation value. The calculation formula of the modified post-traumatic hemorrhagic shock onset probability is:

[0038]

[0039] Where C is the probability threshold.

[0040] After the early warning model update subsystem detects changes in the local medical environment, it collects the patient's physiological parameters and prognosis information, or uses the input patient's physiological parameters and prognosis information, combined with the incremental training method, to update the model parameters without changing the stacked deep learning model structure, and transmits the patient's physiological parameters within the learning window in real time to the dynamic early warning subsystem after the stacked deep learning model parameters are updated.

[0041] The beneficial effects of the present invention are:

[0042] The present invention discloses a dynamic early warning deep learning system for post-traumatic hemorrhagic shock based on non-invasive parameters. Only common non-invasive parameters are used, and laboratory data is not required, which expands the scope of application of the system, making it possible to use it in remote areas, public health emergencies and battlefield front-line situations, eliminating the harm caused to individual patients by frequent collection of laboratory parameters, and reducing the cost of system use. Compared with the traditional linear addition regression model, the automated and process-based intelligent dynamic warning subsystem has a higher computational complexity, is more suitable for common nonlinear problems in real problems, and can provide better post-traumatic hemorrhagic shock onset warning capabilities. The warning interval reserved for clinical intervention can provide sufficient time for doctors to design patient treatment plans. The analysis results of the system of the present invention are consistent with the results of clinical studies, and the model performance is excellent. The system can automatically update the model weights according to environmental changes to adapt to different clinical environments, while effectively reducing the false probability of post-traumatic hemorrhagic shock onset prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a block diagram of a dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters disclosed in the present invention;

[0044] Figure 2 A relationship diagram of the observation window, the delay window, and the warning window in the dynamic warning subsystem of the present invention;

[0045] Figure 3 This is a workflow diagram of the silent subsystem of the present invention. DETAILED DESCRIPTION

[0046] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or equipment that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or equipment.

[0048] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0049] Figure 1 It is a block diagram of a dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters disclosed in the present invention; Figure 2 This is a relationship diagram of the observation window, delay window, and warning window in the dynamic warning subsystem of the present invention. Figure 2 In the above, learning window: the time range with available data, used as the input of the intelligent dynamic warning submodule. Prediction window: the time period used to determine whether post-traumatic hemorrhagic shock has occurred. Delay window: the time difference between the prediction window and the learning window. Where T0 is the time point when the patient is admitted to the ICU, T1 is set to the current time point, satisfying T1≥T0, x1 is the time length of the learning window, satisfying T1-x1≥T0, x2 is the delay window length, the time period of the prediction window is [T1+x2,T1+x2+x3], and x3 is the prediction window length. Figure 3 This is a workflow diagram of the silent subsystem of the present invention.

[0050] The non-invasive parameters described in the present invention refer to parameters that can be measured without causing trauma to the patient's body or requiring a laboratory environment, such as heart rate, non-invasive blood pressure, respiratory rate, gender, age, etc. Non-invasive parameters are different from laboratory parameters.

[0051] Embodiment 1:

[0052] The present invention discloses a post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters, which comprises: a physiological parameter sensing subsystem, a warning model updating subsystem, a dynamic warning subsystem, a post-traumatic hemorrhagic shock onset judgment subsystem and a silent subsystem;

[0053] The physiological parameter sensing subsystem is used to monitor and collect the patient's physiological parameter data information in real time, pre-process the collected patient's physiological parameter data, and then send the pre-processed data to the dynamic early warning subsystem.

[0054] The physiological parameter sensing subsystem is connected to the dynamic early warning subsystem, the dynamic early warning subsystem is connected to the early warning model updating subsystem and the post-traumatic hemorrhagic shock onset judgment subsystem respectively, and the silent subsystem is connected to the post-traumatic hemorrhagic shock onset judgment subsystem, the dynamic early warning subsystem and the early warning model updating subsystem respectively.

[0055] The patient's physiological parameter data information includes the patient's conventional non-invasive physiological parameters and the patient's physiological parameter time series data within the learning window.

[0056] The preprocessing of the collected patient physiological parameter data includes, for the discrete sampling values ​​of the collected patient physiological parameters, using the discrete sampling values ​​to form a sampling vector, calculating the cross-correlation matrix of the sampling vector, processing the cross-correlation matrix using a feature extraction method to obtain its eigenvalue vector, and using the eigenvalue vector to weight the sampling vector to obtain a smoothed value of the patient's physiological parameter as the preprocessed data, thereby effectively reducing the interference of the collected wild value and other error values ​​on the subsequent early warning judgment process, and reducing the false alarm probability of the early warning judgment process.

[0057] The preprocessing of the collected patient physiological parameter data includes: the sampling vector composed of discrete sampling values ​​of the patient physiological parameters collected within a period of time is expressed as [x1, x2, ..., x N ], N is the number of discrete sampling values ​​of the patient's physiological parameters collected within a period of time, the mutual correlation matrix C of the sampling vector is calculated, and the eigenvalue decomposition of the mutual correlation matrix C is performed to obtain:

[0058] C=VDV H ,

[0059] Among them, C is the eigenvector matrix, D is the eigenvalue matrix, the diagonal elements of the matrix D are normalized and used as weight vectors, the sampling vectors are weighted summed, and the smoothed values ​​of the patient's physiological parameters are obtained as preprocessed data.

[0060] The dynamic early warning subsystem is used to receive and clean the conventional non-invasive physiological parameters of the patient collected by the physiological parameter sensing subsystem and the physiological parameter time series data within the patient learning window, calculate the original probability of the patient suffering from post-traumatic hemorrhagic shock within the prediction time window, and then calculate the probability of the patient suffering from post-traumatic hemorrhagic shock within the future prediction time window according to the conventional non-invasive physiological parameters collected by the physiological parameter sensing subsystem and the physiological parameter time series data within the patient learning window, obtain the predicted probability value, and send the predicted probability value to the post-traumatic hemorrhagic shock onset judgment subsystem. The post-traumatic hemorrhagic shock onset judgment subsystem will correct the predicted probability value according to the medical environment data of its location and the doctor's treatment experience evaluation value. If the corrected predicted probability value is greater than the first predicted probability threshold, the post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters judges that the patient will suffer from post-traumatic hemorrhagic shock within the future prediction time window. If the corrected predicted probability value is less than or equal to the second predicted probability threshold, the system judges that the patient will not suffer from post-traumatic hemorrhagic shock within the future prediction time window.

[0061] The early warning model updating subsystem monitors the medical environment where the system is located. If the early warning model updating subsystem detects changes in the medical environment where the system is located, the subsystem uses an incremental learning method to update the weights of the deep learning model in the dynamic early warning subsystem based on the patient's routine non-invasive physiological parameters collected by the physiological parameter sensing subsystem and the prognosis information including the patient's mortality, post-traumatic hemorrhagic shock, and length of hospitalization, so that the early warning results of the system adapt to the medical environment where the system is located.

[0062] The silent subsystem determines the differences among the test data, training data and label data of the dynamic early warning of post-traumatic hemorrhagic shock of the intelligent dynamic warning subsystem. If there is a significant difference between the three types of data, the silent subsystem controls the dynamic warning subsystem not to output the warning result, and collects the warning result data of the dynamic warning subsystem within a period of time, extracts the warning result data with the maximum prediction probability within the period of time, and outputs it to the post-traumatic hemorrhagic shock onset judgment subsystem.

[0063] The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including:

[0064] The data of the early warning model update subsystem is regarded as a stationary random process. Corresponding autoregressive-moving average models, namely, ARMA models, are established for the test data, label data and training data, respectively, and the first ARMA model, the second ARMA model and the third ARMA model are obtained respectively. The cross-correlation matrix of the coefficients of the three ARMA models is calculated, and the maximum eigenvalue of the cross-correlation matrix is ​​calculated. The maximum eigenvalue is used to distinguish the differences between the test data, label data and training data.

[0065] When the maximum eigenvalue is greater than the difference discrimination threshold, it is determined that significant differences occur among the test data, label data and training data of the early warning model updating subsystem.

[0066] The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including:

[0067] The silent subsystem uses the training data and the deep learning model to predict the predicted value of the labeled data, and uses the credibility evaluation method to calculate the credibility between the test data, training data and the predicted value of the labeled data. When the credibility is lower than the credibility threshold, it is determined that there is a significant difference between the warning data, label data and training data of the early warning model update subsystem.

[0068] The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including:

[0069] A sample set is formed by using training data and label data prediction values. According to the label category to which each test data belongs, the sample set is divided into a first reference sample set and a second reference sample set corresponding to the test data. The first reference sample set is a set composed of samples corresponding to the label category to which the test data belongs, and the second reference sample set is a set composed of other samples in the sample set excluding the first reference sample set. A consistency measurement calculation function is used to calculate the distance between each sample in the first reference sample set and the reference sample set to obtain a first sample set distance value vector. A consistency measurement calculation function is used to calculate the distance between each test data and its corresponding first reference sample set and second reference sample set, respectively, to obtain the first set distance value and the second set distance value of the test data. The first set distance value of the test data is sorted in ascending order in the first sample set distance value vector to obtain a sequence value. The ratio of the sequence value to the total number of samples in the first reference sample set is calculated as the credibility of the test data. If the credibility of the test data is lower than the credibility threshold, it is determined that there is a significant difference between the warning data, label data and training data of the warning model update subsystem; the credibility of each test data is calculated and compared with the credibility threshold.

[0070] The deep learning model may adopt a stacked deep learning model, a convolutional neural network module, etc.

[0071] The stacked deep learning model includes a convolutional layer, a bidirectional long short-term memory layer, and a self-attention layer.

[0072] The dynamic early warning subsystem uses an adaptive interpolation method to interpolate missing patient physiological parameter time series data within the minimum time unit, and the specific steps include:

[0073] S1, calculate the median Midparam and quartile IQRidparam of the sampling frequency of each physiological parameter from admission to discharge for each patient;

[0074] S2, selecting physiological parameters with missing values ​​of patients with missing data and recording the parameters as param parameters, and sorting all the collected param parameters in ascending order of collection time, and selecting the time position of the empty value or abnormal value in the collected param parameters as the selected position;

[0075] S3, interpolate data for the selected position in the following priority order: first, take the median of the valid values ​​of the first (Midparam+IQRidparam) sampling time lengths of the selected position as the data interpolation value of the selected position; secondly, take the median of the valid values ​​of the first (Midparam+2*IQRidparam) sampling time lengths of the selected position as the data interpolation value of the selected position; thirdly, take the median of the valid values ​​of the collected param parameters as the data interpolation value of the selected position; finally, take the median of the param parameters recorded in the database as the data interpolation value of the selected position.

[0076] The post-traumatic hemorrhagic shock onset judgment subsystem receives the original probability of the patient suffering from post-traumatic hemorrhagic shock within the prediction time window. The post-traumatic hemorrhagic shock onset judgment subsystem obtains the classification threshold D of the stacked deep learning model based on the sensitivity and specificity of the stacked deep learning model in the dynamic early warning subsystem on the test set data, so that when the classification threshold is D, the difference between the sensitivity and specificity of the stacked deep learning model on the test set data is minimized.

[0077] The post-traumatic hemorrhagic shock onset judgment subsystem modifies the probability value of post-traumatic hemorrhagic shock according to the medical environment data of the location and the doctor's treatment experience evaluation value. The calculation formula of the modified post-traumatic hemorrhagic shock onset probability is:

[0078]

[0079] Where C is the probability threshold.

[0080] After the early warning model update subsystem detects changes in the local medical environment, it collects the patient's physiological parameters and prognosis information, or uses the input patient's physiological parameters and prognosis information, combined with the incremental training method, to update the model parameters without changing the stacked deep learning model structure, and transmits the patient's physiological parameters within the learning window in real time to the dynamic early warning subsystem after the stacked deep learning model parameters are updated.

[0081] The working principle of the dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters is shown in Figure 1 The relationship between the observation window, delay window, and warning window in the dynamic warning subsystem is shown in Figure 2 . Figure 2 In the above, learning window: the time range with available data, used as the input of the intelligent dynamic warning submodule. Prediction window: the time period used to determine whether post-traumatic hemorrhagic shock has occurred. Delay window: the time difference between the prediction window and the learning window. Where T0 is the time point when the patient is admitted to the ICU, T1 is set to the current time point, satisfying T1≥T0, x1 is the time length of the learning window, satisfying T1-x1≥T0, x2 is the delay window length, the time period of the prediction window is [T1+x2,T1+x2+x3], and x3 is the prediction window length.

[0082] The physiological parameter sensing subsystem is used to monitor the conventional non-invasive physiological parameters of the patient in real time, and collect the conventional non-invasive physiological parameters of the patient and the time series data of the physiological parameters within the patient's learning window. The dynamic early warning subsystem is used to receive the conventional non-invasive physiological parameters of the patient and the time series data of the physiological parameters within the patient's learning window collected by the physiological parameter sensing subsystem, and then calculate the probability of occurrence of post-traumatic hemorrhagic shock in the patient within the prediction window in the future according to the conventional non-invasive physiological parameters and the time series data of the physiological parameters within the patient's learning window collected by the physiological parameter sensing subsystem, obtain the predicted probability value, and send the predicted probability value to the post-traumatic hemorrhagic shock onset judgment subsystem. The post-traumatic hemorrhagic shock onset judgment subsystem will correct the predicted probability value according to the medical environment data of its location and the doctor's treatment experience evaluation value. If the corrected predicted probability value is greater than 0.5, the post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters judges that the patient will have post-traumatic hemorrhagic shock in the future within the prediction window. If the corrected predicted probability value is less than or equal to 0.5, the system judges that the patient will not have post-traumatic hemorrhagic shock in the future within the prediction window. If the early warning model update subsystem detects changes in the medical environment, the subsystem will use the incremental learning method to update the weights of the stacked deep learning model in the dynamic early warning subsystem based on the patient's routine non-invasive physiological parameters collected by the physiological parameter perception subsystem, as well as clinical prognosis information such as patient mortality, post-traumatic hemorrhagic shock, and hospitalization duration, so that the early warning results of the system can adapt to the medical environment in which the system is located. The stacked deep learning model includes a convolutional layer, a bidirectional long short-term memory layer, and a self-attention layer.

[0083] The physiological ventilation parameter sensing subsystem has the function of real-time monitoring of non-invasive physiological parameters of patients, and can realize real-time monitoring of parameters such as age, gender, BMI, status of Mechanical ventilation, Glasgow Coma Score (gcs), gcs-verbal, gcs-motor, gcs-eyes, fiO2, PEEP, etco2, tidal volume, urineoutput, Heartrate, Respiration rate, temperature, SpO2, Non-invasive systolic blood pressure, Non-invasive diastolic blood pressure, etc. The physiological ventilation parameter sensing subsystem is connected to the local hospital database or the medical staff manually inputs the data.

[0084] The dynamic early warning subsystem combines the patient's routine non-invasive physiological parameters within the learning window obtained by the physiological ventilation parameter sensing subsystem to calculate the original probability that the patient will develop post-traumatic hemorrhagic shock symptoms within the prediction window in the future.

[0085] The post-traumatic hemorrhagic shock onset judgment subsystem will correct the original probability output by the dynamic warning subsystem based on the local medical environment and doctor's treatment experience, and provide feedback to medical staff on whether the early dynamic warning system will develop post-traumatic hemorrhagic shock symptoms within the prediction window in the future.

[0086] After ICU patients are connected to the dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters, the physiological ventilation parameter sensing subsystem monitors the patient's physiological parameters in real time and transmits the patient's parameters within the learning window to the dynamic warning subsystem in real time.

[0087] Taking 1 hour as the minimum time unit, if there are multiple valid sampling points within a one-hour span in the original data, the dynamic early warning subsystem calculates the total urine volume of patients within the hour for urine volume, and calculates the median for other parameters, and then interpolates the missing data values.

[0088] The dynamic warning subsystem uses an adaptive interpolation method to interpolate the missing physiological parameter time series data within the minimum time unit. The specific steps include:

[0089] S1, calculate the median Midparam and quartile IQRidparam of the sampling frequency of each physiological parameter from admission to discharge for each patient;

[0090] S2, selecting physiological parameters with missing values ​​of patients with missing data and recording the parameters as param parameters, and sorting the physiological parameters in ascending order of recording time, selecting the time position of the null value or abnormal value in the collected param parameters as the selected position; the abnormal value is the data outside the 95% CI of the parameter in the database;

[0091] S3, interpolate data for the selected position in the following priority order, first use the median of the valid values ​​of the first (Midparam+IQRidparam) sampling time lengths of the selected position as the data interpolation value of the selected position, secondly use the median of the valid values ​​of the first (Midparam+2*IQRidparam) sampling time lengths of the selected position as the data interpolation value of the selected position, thirdly use the median of the valid values ​​of the param parameter as the data interpolation value of the selected position, and finally use the median of the param parameter recorded in the database as the data interpolation value of the selected position. For example, for a patient with ID 32572156, there is a missing value in the heart rate parameter at the 8th hour after admission. The 8th hour is taken as the selected position. It is calculated that the median of the sampling of the patient's heart rate parameter is one valid value every 2 hours, and the upper quartile is one valid value every 1 hour. Then the median of the valid values ​​of the heart rate parameter from (8-2-1) to the 8th hour after the patient's admission is calculated as the interpolation value of the missing value at the 8th hour.

[0092] It should be noted that interpolated data are not considered as valid value points.

[0093] The stacked deep learning model in the dynamic early warning subsystem uses the time series data of the patient's physiological parameters within the cleaned learning window to obtain the original probability of the patient suffering from post-traumatic hemorrhagic shock within the prediction window.

[0094] The original probability of post-traumatic hemorrhagic shock in patients within the prediction window is imported into the post-traumatic hemorrhagic shock onset judgment subsystem. The post-traumatic hemorrhagic shock onset judgment subsystem obtains the classification threshold D of the stacked deep learning model based on the sensitivity and specificity of the stacked deep learning model in the dynamic early warning subsystem on the test set data, so that when the classification threshold is D, the difference in sensitivity and specificity of the stacked deep learning model on the test set data is minimized.

[0095] If the corrected probability value of the incidence of post-traumatic hemorrhagic shock is greater than 0.5, the early dynamic warning system judges that the patient will develop symptoms of post-traumatic hemorrhagic shock within the prediction window in the future; otherwise, the patient will not develop symptoms of post-traumatic hemorrhagic shock.

[0096] The silent subsystem, in the classification problem of machine learning, assumes that the training data in the sample set is x i ∈X,i=1,2,…,n, the label data in the sample set is y i ∈Y,i=1,2,…,n, then define the sample space z i =(x i ,y i ), i = 1, 2, ..., n are elements in the sample space Z = X × Y. In general, when a new object x is given n+1When the machine learning model predicts x based on the law between sample X and label Y in this space Z = X × Y n+1 The label is y n+1 =F(x1,x2,…,x n ,x n+1 ,y1,y2,…,y n ).

[0097] The silent subsystem is designed to evaluate the test data x n+1 With warning label data The silent system suppresses the output of warning results by the post-traumatic hemorrhagic shock onset judgment subsystem for labels with low credibility. Set the parameter ε∈(0,1) that reflects the significance level. The method by which the silent subsystem calculates the credibility between the object and the warning label is the consistency measurement method, that is, the test data x n+1 The degree of similarity with existing samples. In the silent subsystem, the degree of similarity is the minimum Euclidean distance between the test data and the reference sample set.

[0098] Taking the binary classification task as an example, we first define the positive reference sample set in the reference sample set Negative reference sample set And suppose the consistency metric calculation function is F α , which is used to calculate the minimum Euclidean distance between the test data and the reference sample set. The minimum Euclidean distance between the test data and the reference sample set refers to the minimum Euclidean distance between the elements of the test data and the reference sample set.

[0099] Calculate the test data x n+1 The minimum distance of the reference sample set that is opposite to the corresponding predicted label, and is marked as α n+1 The distance from the sample to the set is calculated using the Euclidean distance calculation method. Similarly, the metric values ​​of the positive reference sample set and the negative reference sample set are calculated as follows: and The metric value of the reference sample set is the set of distances from each element in the reference sample set to the set. n+1 If the model warns that the sample is positive, then i+1 In the positive sample metric set α pos Sort in ascending order to obtain the sorting number result, and calculate the ratio of the sorting number result to the total number of positive samples. The calculation formula is:

[0100]

[0101] If p is less than the set threshold (such as 0.05), the silent subsystem controls the model warning result to be temporarily not output, and collects the model's predicted probability output results within the next 15 minutes to the silent subsystem. The silent subsystem returns the maximum predicted probability output result within 15 minutes to the post-traumatic hemorrhagic shock onset judgment subsystem.

[0102] In case of changes in the clinical medical environment, ICU patients are connected to an interpretable post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters. The physiological ventilation parameter sensing subsystem monitors the patient's physiological parameters in real time.

[0103] Taking hour as the minimum time unit, if there are multiple valid sampling points within one hour in the original data, the dynamic warning subsystem calculates the total urine volume of the patient within the hour for urine volume, and calculates the median for other parameters, and then interpolates the missing data values. The warning model update subsystem collects the patient's physiological parameters and prognostic information, or uses the input physiological parameters and prognostic information, combined with the incremental training method, to update the model parameters without changing the stacked deep learning model structure, and transmits the patient's physiological parameters within the learning window to the dynamic warning subsystem after updating the stacked deep learning model parameters in real time.

[0104] The dynamic early warning subsystem transmits the patient data within the cleaned learning window to the trained machine learning model. After processing by the model, the original probability of the patient suffering from post-traumatic hemorrhagic shock within the prediction window can be obtained.

[0105] The original probability of post-traumatic hemorrhagic shock within the prediction window is imported into the post-traumatic hemorrhagic shock onset judgment subsystem. The classification threshold D is obtained by combining the warning performance of the machine learning model on the test set data in the dynamic warning subsystem with the doctor's experience. When the classification threshold is D, the sensitivity or specificity of the model on the test set data is the expert recommended value.

[0106] It can be seen that the present invention discloses a dynamic early warning depth system for post-traumatic hemorrhagic shock based on non-invasive parameters. It only uses common non-invasive parameters and does not require laboratory data, which expands the scope of application of the system, making it possible to use it in remote areas, public health emergencies and battlefield front-line situations, eliminating the harm caused to individual patients by frequent collection of laboratory parameters, and reducing the cost of system use. Compared with the traditional linear addition regression model, the automated and process-based intelligent dynamic warning subsystem has a higher computational complexity, is more suitable for common nonlinear problems in real problems, and can provide better post-traumatic hemorrhagic shock onset warning capabilities. The warning interval reserved for clinical intervention can provide sufficient time for doctors to design patient treatment plans. The analysis results of the system of the present invention are consistent with the results of clinical studies, and the model performance is excellent. The system can automatically update the model weights according to environmental changes to adapt to different clinical environments, while effectively reducing the false probability of post-traumatic hemorrhagic shock onset prediction.

[0107] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters, comprising: Physiological parameter sensing subsystem, early warning model updating subsystem, dynamic early warning subsystem, post-traumatic hemorrhagic shock onset judgment subsystem and silent subsystem; The physiological parameter sensing subsystem is used to monitor and collect the patient's physiological parameter data information in real time, and pre-process the collected patient's physiological parameter data, and then send the pre-processed data to the dynamic early warning subsystem; The physiological parameter sensing subsystem is connected to the dynamic early warning subsystem, the dynamic early warning subsystem is connected to the early warning model updating subsystem and the post-traumatic hemorrhagic shock onset judgment subsystem respectively, and the silent subsystem is connected to the post-traumatic hemorrhagic shock onset judgment subsystem, the dynamic early warning subsystem and the early warning model updating subsystem respectively; The dynamic early warning subsystem is used to receive and clean the conventional non-invasive physiological parameters of the patient collected by the physiological parameter sensing subsystem and the physiological parameter time series data within the patient learning window, calculate the original probability of the patient suffering from post-traumatic hemorrhagic shock within the prediction time window, and then calculate the probability of the patient suffering from post-traumatic hemorrhagic shock within the future prediction time window according to the conventional non-invasive physiological parameters collected by the physiological parameter sensing subsystem and the physiological parameter time series data within the patient learning window, obtain the predicted probability value, and send the predicted probability value to the post-traumatic hemorrhagic shock onset judgment subsystem. The post-traumatic hemorrhagic shock onset judgment subsystem will correct the predicted probability value according to the medical environment data of its location and the doctor's treatment experience evaluation value. If the corrected predicted probability value is greater than the first predicted probability threshold, the post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters judges that the patient will suffer from post-traumatic hemorrhagic shock within the future prediction time window. If the corrected predicted probability value is less than or equal to the second predicted probability threshold, the system judges that the patient will not suffer from post-traumatic hemorrhagic shock within the future prediction time window. The silent subsystem determines the differences among the test data, training data and label data of the dynamic early warning of post-traumatic hemorrhagic shock of the intelligent dynamic early warning subsystem. If there is a significant difference between the three types of data, the silent subsystem controls the dynamic early warning subsystem not to output the early warning result, and collects the early warning result data of the dynamic early warning subsystem within a period of time, extracts the early warning result data with the maximum prediction probability within the period of time, and outputs it to the post-traumatic hemorrhagic shock onset judgment subsystem; The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including: A sample set is formed by using training data and label data prediction values. According to the label category to which each test data belongs, the sample set is divided into a first reference sample set and a second reference sample set corresponding to the test data. The first reference sample set is a set composed of samples corresponding to the label category to which the test data belongs, and the second reference sample set is a set composed of other samples in the sample set except the first reference sample set. A consistency measurement calculation function is used to calculate the distance between each sample in the first reference sample set and the reference sample set to obtain a first sample set distance value vector. A consistency measurement calculation function is used to calculate the distance between each test data and its corresponding first reference sample set and second reference sample set, respectively, to obtain the first set distance value and the second set distance value of the test data. The first set distance value of the test data is sorted in ascending order in the first sample set distance value vector to obtain a sequence value. The ratio of the sequence value to the total number of samples in the first reference sample set is calculated as the credibility of the test data. If the credibility of the test data is lower than the credibility threshold, it is determined that there is a significant difference between the warning data, label data and training data of the warning model update subsystem. The credibility of each test data is calculated and compared with the credibility threshold. The patient's physiological parameter data information includes the patient's conventional non-invasive physiological parameters and the patient's physiological parameter time series data within the learning window; The preprocessing of the collected patient physiological parameter data includes, for the discrete sampling values ​​of the collected patient physiological parameters, using the discrete sampling values ​​to form a sampling vector, calculating the cross-correlation matrix of the sampling vector, processing the cross-correlation matrix using a feature extraction method to obtain its eigenvalue vector, and weighting the sampling vector using the eigenvalue vector to obtain a smoothed value of the patient's physiological parameter as the preprocessed data.

2. The post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters as claimed in claim 1, The preprocessing of the collected patient physiological parameter data includes: the sampling vector composed of discrete sampling values ​​of the patient physiological parameters collected within a period of time is expressed as [x1, x2, ..., x N ], N is the number of discrete sampling values ​​of the patient's physiological parameters collected within a period of time, the mutual correlation matrix C of the sampling vector is calculated, and the eigenvalue decomposition of the mutual correlation matrix C is performed to obtain: C=VDV H , in, C is the eigenvector matrix, D is the eigenvalue matrix, the diagonal elements of the matrix D are normalized and used as weight vectors, the sampling vectors are weighted summed, and the smoothed values ​​of the patient's physiological parameters are obtained as preprocessed data.

3. According to the dynamic early warning system for post-traumatic hemorrhagic shock based on non-invasive parameters as described in claim 1, the warning model updating subsystem monitors the medical environment where the system is located. If the warning model updating subsystem detects that the medical environment has changed, the subsystem uses an incremental learning method to update the weights of the deep learning model in the dynamic warning subsystem based on the patient's routine non-invasive physiological parameters collected by the physiological parameter sensing subsystem and the prognostic information including the patient's mortality, the occurrence of post-traumatic hemorrhagic shock, and the length of hospitalization, so that the warning results of the system adapt to the medical environment where the system is located.

4. The post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters as claimed in claim 1, The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including: The silent subsystem uses the training data and the deep learning model to predict the predicted value of the labeled data, and uses the credibility evaluation method to calculate the credibility between the test data, training data and the predicted value of the labeled data. When the credibility is lower than the credibility threshold, it is determined that there is a significant difference between the warning data, label data and training data of the early warning model update subsystem.

5. The post-traumatic hemorrhagic shock dynamic early warning system based on non-invasive parameters as claimed in claim 1, The silent subsystem determines the difference between the test data, label data and training data of the early warning model update subsystem, including: The data of the early warning model update subsystem is regarded as a stationary random process. For the test data, label data and training data, the corresponding autoregressive-moving average model, i.e., ARMA model, is established respectively. The first ARMA model, the second ARMA model and the third ARMA model are obtained respectively. The coefficients of the three ARMA models are calculated, and the maximum eigenvalue of the cross-correlation matrix is ​​calculated. The maximum eigenvalue is used to distinguish the differences between the test data, label data and training data. When the maximum eigenvalue is greater than the difference discrimination threshold, it is determined that significant differences occur among the test data, label data and training data of the early warning model updating subsystem.

Images