CN-121687517-B - Method, device, terminal equipment and medium for predicting dispersive intravascular coagulation risk

CN121687517BCN 121687517 BCN121687517 BCN 121687517BCN-121687517-B

Abstract

The invention provides a method, a device, terminal equipment and a medium for predicting dispersive intravascular coagulation risk, which comprise the steps of obtaining an original pathological data sequence, extracting time domain statistical characteristics of the original pathological data sequence, constructing a DIC dynamic balance index, constructing a multi-mode feature vector according to the time domain statistical characteristics and the DIC dynamic balance index, inputting the multi-mode feature vector into a dual-channel decoupling network based on a pathological countermeasure mechanism, mapping the multi-mode feature vector to a thrombus attention channel and a bleeding attention channel which are physically isolated by taking an orthogonal query vector generated by a static baseline characteristic as semantic guidance, capturing the evolution process of thrombus formation and hyperfibrinolysis by using the implicit pathological countermeasure tensor, and outputting a risk probability predicted value of dispersive intravascular coagulation of an object to be measured in a future preset time window. The invention can improve the accuracy of the prediction of the risk of disseminated intravascular coagulation.

Inventors

LI JINXIU
MA YUXUAN
Wang Guyi
XIANG QINGYU
XIANG BINGJIE
Gong Subo
HUANG YENING
LIU ZHIXIAN
LIU JINPING

Assignees

中南大学湘雅二医院

Dates

Publication Date: 20260508
Application Date: 20260205

Claims (7)

1. A method for predicting risk of disseminated intravascular coagulation comprising: acquiring multi-mode dynamic time sequence data of an object to be tested from a monitoring system, and performing standardized preprocessing to obtain an original pathological data sequence, wherein the multi-mode dynamic time sequence data comprises vital sign data, coagulation index data and viscoelastic coagulation test data; Extracting time domain statistical features of the original pathological data sequence by utilizing a sliding window technology, wherein the time domain statistical features are used for representing the overall level and the change trend of the original pathological data sequence; establishing a DIC dynamic balance index according to a real-time dynamic unbalance relation between a blood coagulation activation index and a blood coagulation consumption index in the original pathological data sequence, wherein the DIC dynamic balance index is used for quantifying a dynamic unbalance state of a blood coagulation system, and the DIC dynamic balance index has the expression: Wherein, the Representing the DIC dynamic balance index, Represents a set of coagulation activation related indicators, including D-dimer and FDP, Represents the relevant index of the blood coagulation activation, , Representing the parameters of the weight that can be learned, Indicating a coagulation activation related index At the moment of time Is of the observed value of (2) The results after Z-score normalization treatment, Representation of Is used for the average value of (a), Representation of Is set in the standard deviation of (2), Represents a set of blood coagulation consumption related indicators including PLT, FIB and AT-III, Indicating blood coagulation consumption related index At the moment of time Is of the observed value of (2) The results after Z-score normalization treatment, Representation of Is used for the average value of (a), Representation of Is set in the standard deviation of (2), The term of the bias is indicated, Representing the hyperbolic tangent activation function, Positive values of (c) tend to hypercoagulability/hyperfibrinolysis, Tends to be consumable and congeal; Constructing a multi-modal feature vector according to the time domain statistical feature and the DIC dynamic balance index, inputting the multi-modal feature vector into a dual-channel decoupling network based on a pathology countermeasure mechanism, respectively mapping the multi-modal feature vector to a thrombus concerned channel and a bleeding concerned channel which are physically isolated by using an orthogonal query vector generated by a static baseline feature as semantic guidance, capturing the evolution process of thrombus formation and hyperfibrinolysis by using implicit pathology countermeasure tensor, and outputting a risk probability prediction value of the occurrence of dispersive intravascular coagulation of the object to be detected in a future preset time window; the static baseline characteristics include demographic information, underlying disease history, severity score at admission, and whether there is underlying cause that induces disseminated intravascular coagulation; The orthogonal query vectors include an acceleration risk query vector Bleeding risk query vector Wherein, the method comprises the steps of, ; ; And The binary mask is represented by a binary mask, Representing a static baseline characteristic of the device, Projection weight matrix representing coagulation risk inquiry is a learnable parameter matrix for mapping static baseline characteristics after masking treatment to coagulation inquiry vector space so as to generate The characteristics related to thrombus formation in the dynamic data can be specially queried and matched in a subsequent attention mechanism; representing a bleeding risk query projection weight matrix, which is a learnable parameter matrix, for mapping the masked static baseline features to a bleeding query vector space, so as to generate Can specifically inquire about the characteristics related to hyperfibrinolysis or blood coagulation factor consumption in the dynamic data; the thrombus-focusing channel is used for searching the signs of activated blood coagulation factors and micro-thrombus formation in multi-modal data, and has the expression of Wherein, the method comprises the steps of, Representing a thrombus characterization vector, which is the final output of the channel, representing the high-level features of the model extracted from the multi-modal dynamic data reflecting the risk of coagulation activation and microthrombotic formation; The multi-mode feature vector is represented by performing dimensional splicing on a time domain statistical feature sequence and a DIC dynamic balance index, and the multi-mode feature vector fuses physiological trend time domain features and a coagulation imbalance state of an object to be detected and is core input data for understanding the dynamic evolution of the patient's illness state by a model; A matrix of leachable key projection weights representing a thrombus-focused channel, which is a matrix of trainable parameters; Representing the dimensions of a key vector, representing a query vector Sum key vector Dimension size of (2); A matrix of leachable projection weights representing the thrombus-focused channel, which is also a matrix of trainable parameters; the bleeding and focusing channel is used for searching for signs of blood coagulation factor consumption, hyperfibrinolysis and platelet collapse, and has the expression of Wherein, the method comprises the steps of, A matrix of leachable key projection weights representing a bleeding-attention channel, trainable, for applying A matrix of leachable projection weights representing the bleeding-attention channel, trainable, for applying 。
2. The method according to claim 1, wherein the implicit pathological contrast tensor is obtained by performing a characteristic difference operation on the output characteristics of the thrombus-interest channel and the output characteristics of the bleeding-interest channel, and is used for characterizing the degree of coagulation-fibrinolysis contrast which cannot be represented by a single clinical index, and the implicit pathological contrast tensor is given as a value , Representing the characteristic difference operator.
3. The method for predicting the risk of disseminated intravascular coagulation of claim 2, wherein the two-channel decoupling network is constrained by a composite loss function during training, wherein the composite loss function comprises an early prediction loss for introducing time attenuation weight, a two-channel auxiliary supervision loss for improving the sensitivity of the two-channel decoupling network to pre-diagnosis latency signals and an orthogonal decoupling constraint for ensuring that the two channels remain independent in semantic space; Composite loss function The expression of (2) is as follows: Wherein, the Representing the early-stage prediction loss, Representing a two-channel auxiliary supervision loss, Representing the orthogonal decoupling constraints of the device, The super-parameter is represented by a parameter, , Representing the total number of samples; Representing sample weights; Represent the first A sample true tag, which takes a value of 1 if the sample is from a patient diagnosed with DIC, or 0 if the sample is not; Represent the first The prediction probability of each sample is a probability value which is output by a model and is considered to be DIC of the sample in a future preset time window; representing DIC time points for diagnosis; representing a current predicted point in time, i.e. the instant of the current set of pathological data sequences; Representing time decay coefficients for controlling weights A time-varying superparameter; Representing a binary cross entropy loss function for measuring the difference between the predicted probability distribution and the real label distribution output by the model; an auxiliary sorting head representing the end of the channel, Indicating that the label is a counterfeit label with a set accelerator, Indicating that the bleeding is a false label, As a result of the peak gain factor, The larger the model is, the larger the penalty in misprediction is.
4. The method of claim 3, wherein the viscoelastic coagulation test data comprises at least R-time, K-time, angle, maximum amplitude MA and LY30 parameters in a thromboelastography TEG or a rotational thromboelastography ROTEM.
5. A diffuse intravascular coagulation risk prediction device, comprising: The system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module is used for acquiring multi-mode dynamic time sequence data of an object to be detected from a monitoring system and carrying out standardized preprocessing to obtain an original pathological data sequence; the data feature extraction module is used for extracting time domain statistical features of the original pathological data sequence by utilizing a sliding window technology, wherein the time domain statistical features are used for representing the overall level and the change trend of the original pathological data sequence; the DIC index construction module is used for constructing a DIC dynamic balance index according to the real-time dynamic unbalance relation between the blood coagulation activation index and the blood coagulation consumption index in the original pathological data sequence, wherein the DIC dynamic balance index is used for quantifying the dynamic unbalance state of a blood coagulation system, and the expression of the DIC dynamic balance index is as follows: Wherein, the Representing the DIC dynamic balance index, Represents a set of coagulation activation related indicators, including D-dimer and FDP, Represents the relevant index of the blood coagulation activation, , Representing the parameters of the weight that can be learned, Indicating a coagulation activation related index At the moment of time Is of the observed value of (2) The results after Z-score normalization treatment, Representation of Is used for the average value of (a), Representation of Is set in the standard deviation of (2), Represents a set of blood coagulation consumption related indicators including PLT, FIB and AT-III, Indicating blood coagulation consumption related index At the moment of time Is of the observed value of (2) The results after Z-score normalization treatment, Representation of Is used for the average value of (a), Representation of Is set in the standard deviation of (2), The term of the bias is indicated, Representing the hyperbolic tangent activation function, Positive values of (c) tend to hypercoagulability/hyperfibrinolysis, Tends to be consumable and congeal; The risk prediction module is used for constructing a multi-modal feature vector according to the time domain statistical feature and the DIC dynamic balance index, inputting the multi-modal feature vector into a dual-channel decoupling network based on a pathology countermeasure mechanism, mapping the multi-modal feature vector to a thrombus concerned channel and a bleeding concerned channel which are physically isolated by using an orthogonal query vector generated by a static baseline feature as semantic guidance, capturing the evolution process of thrombus formation and hyperfibrinolysis by using a hidden pathology countermeasure tensor, and outputting a risk probability prediction value of the occurrence of dispersive intravascular coagulation of the object to be detected in a future preset time window; the static baseline characteristics include demographic information, underlying disease history, severity score at admission, and whether there is underlying cause that induces disseminated intravascular coagulation; The orthogonal query vectors include an acceleration risk query vector Bleeding risk query vector Wherein, the method comprises the steps of, ; ; And The binary mask is represented by a binary mask, Representing a static baseline characteristic of the device, Projection weight matrix representing coagulation risk inquiry is a learnable parameter matrix for mapping static baseline characteristics after masking treatment to coagulation inquiry vector space so as to generate The characteristics related to thrombus formation in the dynamic data can be specially queried and matched in a subsequent attention mechanism; representing a bleeding risk query projection weight matrix, which is a learnable parameter matrix, for mapping the masked static baseline features to a bleeding query vector space, so as to generate Can specifically inquire about the characteristics related to hyperfibrinolysis or blood coagulation factor consumption in the dynamic data; the thrombus-focusing channel is used for searching the signs of activated blood coagulation factors and micro-thrombus formation in multi-modal data, and has the expression of Wherein, the method comprises the steps of, Representing a thrombus characterization vector, which is the final output of the channel, representing the high-level features of the model extracted from the multi-modal dynamic data reflecting the risk of coagulation activation and microthrombotic formation; The multi-mode feature vector is represented by performing dimensional splicing on a time domain statistical feature sequence and a DIC dynamic balance index, and the multi-mode feature vector fuses physiological trend time domain features and a coagulation imbalance state of an object to be detected and is core input data for understanding the dynamic evolution of the patient's illness state by a model; A matrix of leachable key projection weights representing a thrombus-focused channel, which is a matrix of trainable parameters; Representing the dimensions of a key vector, representing a query vector Sum key vector Dimension size of (2); A matrix of leachable projection weights representing the thrombus-focused channel, which is also a matrix of trainable parameters; the bleeding and focusing channel is used for searching for signs of blood coagulation factor consumption, hyperfibrinolysis and platelet collapse, and has the expression of Wherein, the method comprises the steps of, A matrix of leachable key projection weights representing a bleeding-attention channel, trainable, for applying A matrix of leachable projection weights representing the bleeding-attention channel, trainable, for applying 。
6. A terminal device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the method according to any of claims 1 to 4 when executing the computer program.
7. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the method according to any one of claims 1 to 4.

Description

Method, device, terminal equipment and medium for predicting dispersive intravascular coagulation risk Technical Field The invention belongs to the technical field of data analysis and disease risk prediction, and particularly relates to a method, a device, terminal equipment and medium for predicting dispersive intravascular coagulation risk. Background Disseminated Intravascular Coagulation (DIC) is an acquired coagulation dysfunction syndrome triggered by a variety of underlying diseases (e.g., sepsis, trauma, malignancy, pathological obstetrics, etc.), characterized by extensive systemic microvascular thrombosis, extensive consumption of coagulation factors and platelets, and in turn, life threatening bleeding or multiple organ failure. In Intensive Care Units (ICU), DIC is one of the key factors leading to patient mortality, especially sepsis-induced DIC (SI-DIC), with mortality rates as high as 28% -43%. Currently, clinical diagnosis of DIC mainly depends on scoring systems established by institutions such as the International Society of Thrombosis and Hemostasis (ISTH) (e.g., ISTH dominant DIC scoring). However, these scoring systems have limitations: Static-scoring is based on laboratory examination results at a single time point (e.g., platelet count, prothrombin time PT, fibrinogen, D-dimer) and is unable to capture the dynamic evolution of clotting function over time. DIC is a continuous pathological process, where early, compensatory coagulation dysfunction is difficult to identify by static scoring, often diagnosis can be confirmed only when the patient has entered explicit DIC stage of decompensation, missing optimal intervention opportunities. Hysteresis, namely, hysteresis exists in the return of laboratory inspection results, and the requirements of ICU real-time monitoring cannot be met. Uniqueness: traditional scoring relies mainly on a limited number of coagulation indicators, failing to make full use of other high-dimensional data readily available in ICU, such as parameters of high frequency vital signs, inflammation indicators (PCT, CRP), organ function indicators (lactic acid, creatinine) and more advanced coagulation integrity assessment tools such as Thromboelastography (TEG). The lack of predictive capability-existing methods are diagnostic rather than predictive and do not actively pre-warn patients of future risk of developing DIC. In recent years, there have been studies attempting to construct DIC prediction models using machine learning (such as Logistic regression, random forest), but most models are still based on static features and fail to effectively process multi-modal, high-loss-rate time series data that are prevalent in clinical environments. Resulting in less accurate prediction of the risk of disseminated intravascular coagulation. Disclosure of Invention The invention aims to provide a method, a device, terminal equipment and a medium for predicting risk of disseminated intravascular coagulation so as to improve accuracy of predicting risk of disseminated intravascular coagulation. In a first aspect, the present invention provides a method for predicting risk of disseminated intravascular coagulation, the method comprising the steps of: Acquiring multi-mode dynamic time sequence data of an object to be tested from a monitoring system, and performing standardized preprocessing to obtain an original pathological data sequence, wherein the multi-mode dynamic time sequence data comprises vital sign data, coagulation index data and viscoelastic coagulation test data; Extracting time domain statistical features of the original pathological data sequence by utilizing a sliding window technology, wherein the time domain statistical features are used for representing the overall level and the change trend of the original pathological data sequence; Constructing a DIC dynamic balance index according to the real-time dynamic unbalance relation between the coagulation activation index and the coagulation consumption index in the original pathological data sequence, wherein the DIC dynamic balance index is used for quantifying the dynamic unbalance state of the coagulation system; Constructing a multi-modal feature vector according to the time domain statistical feature and the DIC dynamic balance index, inputting the multi-modal feature vector into a dual-channel decoupling network based on a pathology countermeasure mechanism, respectively mapping the multi-modal feature vector to a thrombus concerned channel and a bleeding concerned channel which are physically isolated by using an orthogonal query vector generated by a static baseline feature as semantic guidance, capturing the evolution process of thrombus formation and hyperfibrinolysis by using implicit pathology countermeasure tensor, and outputting a risk probability prediction value of the occurrence of dispersive intravascular coagulation of an object to be detected in a future preset time window. Optionally, constr