Method, device, terminal equipment and medium for predicting risk of disseminated intravascular coagulation

Abstract

The application provides a diffuse intravascular coagulation risk prediction method, device, terminal equipment and medium, comprising: acquiring an original pathological data sequence; extracting time domain statistical features of the original pathological data sequence; constructing a DIC dynamic balance index; constructing a multi-modal feature vector according to the time domain statistical features and the DIC dynamic balance index, inputting the multi-modal feature vector into a double-channel decoupling network based on a pathological adversarial mechanism, using an orthogonal query vector generated by a static baseline feature as semantic guidance, mapping the multi-modal feature vector to a thrombus focus channel and a bleeding focus channel which are physically isolated, capturing the evolution process of thrombus formation and hyperfibrinolysis through an implicit pathological adversarial tensor, and outputting a risk probability prediction value of diffuse intravascular coagulation of a to-be-tested object in a future preset time window. The application can improve the accuracy of diffuse intravascular coagulation risk prediction.

Description

Technical Field

[0001] This invention relates to the field of data analysis and disease risk prediction technology, specifically to a method, device, terminal equipment, and medium for predicting the risk of disseminated intravascular coagulation. Background Technology

[0002] Disseminated intravascular coagulation (DIC) is an acquired coagulation disorder syndrome triggered by various underlying diseases (such as sepsis, trauma, malignant tumors, and obstetric pathology). It is characterized by widespread microvascular thrombosis, massive consumption of coagulation factors and platelets, leading to life-threatening bleeding or multiple organ failure. In the intensive care unit (ICU), DIC is a key factor contributing to patient mortality, especially sepsis-induced DIC (SI-DIC), with a mortality rate as high as 28%–43%.

[0003] Currently, clinical diagnosis of DIC mainly relies on scoring systems developed by organizations such as the International Society for Thrombosis and Haemostasis (ISTH) (e.g., the ISTH Overt DIC score). However, these scoring systems have limitations:

[0004] Static scoring: Scoring is based on laboratory test results at a single point in time (such as platelet count, prothrombin time (PT), fibrinogen, and D-dimer), and cannot capture the dynamic evolution of coagulation function over time. DIC is a continuous pathological process, and static scoring is difficult to identify early, compensated coagulation dysfunction. Diagnosis is often only made when the patient has entered the decompensated overt DIC stage, thus missing the best intervention opportunity.

[0005] Lag: Laboratory test results are delayed, which cannot meet the needs of real-time monitoring in the ICU.

[0006] Limited scope: Traditional scoring mainly relies on a limited number of coagulation indicators, failing to make full use of other high-dimensional data that are readily available in the ICU, such as high-frequency vital signs, inflammatory markers (PCT, CRP), organ function indicators (lactate, creatinine), and parameters of more advanced overall coagulation assessment tools such as thromboelastography (TEG).

[0007] Lack of predictive ability: Existing methods are diagnostic rather than predictive and cannot proactively warn patients of their future risk of developing DIC.

[0008] In recent years, although some studies have attempted to use machine learning (such as Logistic Regression and Random Forest) to build DIC prediction models, most models are still based on static features and have failed to effectively handle the multimodal and high-missing-rate time-series data commonly found in clinical settings. This results in low accuracy in predicting the risk of disseminated intravascular coagulation (DIC). Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method, device, terminal equipment and medium for predicting the risk of disseminated intravascular coagulation (DIC), so as to improve the accuracy of DIC risk prediction.

[0010] In a first aspect, the present invention provides a method for predicting the risk of disseminated intravascular coagulation, the method comprising the following steps:

[0011] Multimodal dynamic time-series data of the subjects under test are obtained from the monitoring system and standardized preprocessing is performed to obtain the original pathological data sequence; the multimodal dynamic time-series data includes vital signs data, coagulation index data and viscoelastic coagulation test data;

[0012] The sliding window technique was used to extract the temporal statistical features of the original pathological data sequence; the temporal statistical features were used to characterize the overall level and trend of the original pathological data sequence.

[0013] The DIC dynamic balance index is constructed based on the real-time dynamic imbalance between coagulation activation and coagulation consumption indicators in the original pathological data sequence; the DIC dynamic balance index is used to quantify the dynamic imbalance state of the coagulation system.

[0014] Multimodal feature vectors are constructed based on time-domain statistical characteristics and the DIC dynamic equilibrium index. These multimodal feature vectors are then input into a dual-channel decoupled network based on pathological adversarial mechanisms. Orthogonal query vectors generated from static baseline features are used as semantic guides to map the multimodal feature vectors to physically isolated thrombosis and bleeding attention channels, respectively. The evolution of thrombosis and hyperfibrinolysis is captured through implicit pathological adversarial tensors, and the predicted risk probability of disseminated intravascular coagulation (DIC) in the target subject within a preset time window is output.

[0015] Optionally, a dynamic balance index for DIC can be constructed based on the real-time dynamic imbalance between coagulation activation and coagulation consumption indicators in the original pathological data sequence, including:

[0016] Through calculation formula

[0017]

[0018]

[0019]

[0020] DIC dynamic equilibrium index is obtained ;in, This represents a set of coagulation activation-related indicators, including D-dimer and FDP. Indicates coagulation activation-related indicators, , Indicates the learnable weight parameters. Indicators related to coagulation activation At any moment Observations The result after Z-score normalization express The mean, express standard deviation This represents a set of coagulation-related indicators, including PLT, FIB, and AT-III. Indicators related to coagulation consumption At any moment Observations The result after Z-score normalization express The mean, express standard deviation Indicates the bias term. This represents the hyperbolic tangent activation function. Positive values ​​tend towards hypercoagulability / fibrinolysis. The negative value tends to indicate a depletion-related low condensation rate.

[0021] Optional, static baseline characteristics include demographic information, history of underlying diseases, severity score at admission, and presence of underlying causes that could induce disseminated intravascular coagulation;

[0022] Orthogonal query vectors include coagulation risk query vectors and bleeding risk query vector ;in, ; ; and Represents a binary mask. Indicates static baseline characteristics, The projection weight matrix for the coagulation risk query is a learnable parameter matrix used to map the masked static baseline features to the coagulation query vector space, so that the generated... In subsequent attention mechanisms, it is possible to specifically query and match features related to thrombosis in dynamic data; The projection weight matrix for the bleeding risk query is a learnable parameter matrix used to map the masked static baseline features to the bleeding query vector space, so that the generated... It can specifically query dynamic data for features related to hyperfibrinolysis or coagulation factor consumption.

[0023] Optionally, the thrombosis focus channel is used to search for signs of coagulation factor activation and microthrombus formation in multimodal data; the expression for the thrombosis focus channel is: ;in, This represents the thrombosis characterization vector, which is the final output of this channel. It represents the high-level features extracted by the model from multimodal dynamic data, reflecting the risk of coagulation activation and microthrombosis. Representing a multimodal feature vector, it is a dimensional concatenation of a time-domain statistical feature sequence and a DIC dynamic balance index. It integrates the time-domain features of the physiological trend of the test subject with the coagulation imbalance state, and is the core input data for the model to understand the dynamic evolution of the patient's condition. This represents the learnable key projection weight matrix of the thrombus concern channel, which is a trainable parameter matrix; The dimension of the key vector represents the query vector. Q and key vector K Dimension size; This represents the learnable projected weight matrix of the thrombus concern channel, which is also a trainable parameter matrix.

[0024] The bleeding monitoring channel is used to search for signs of coagulation factor depletion, increased fibrinolysis, and a sudden drop in platelets; the expression for the bleeding monitoring channel is: ;in, The learnable key-projection weight matrix representing the bleeding concern channel is trainable and used to... The learnable projected weight matrix representing the bleeding concern channel is trainable and used to... .

[0025] Optionally, the latent pathological antagonism tensor is obtained by performing feature difference operations on the output features of the thrombosis concern channel and the output features of the bleeding concern channel, and is used to characterize the degree of coagulation-fibrinolysis antagonism that cannot be reflected by a single clinical indicator; the value of the latent pathological antagonism tensor is... , This represents the characteristic difference operator.

[0026] Optionally, the dual-channel decoupled network is constrained during training using a composite loss function. The composite loss function includes an early prediction loss with time decay weights, a dual-channel auxiliary supervision loss, and orthogonal decoupling constraints. The early prediction loss is used to improve the sensitivity of the dual-channel decoupled network to pre-diagnosis latency signals, the dual-channel auxiliary supervision loss is constructed based on medical heuristics, and the orthogonal decoupling constraints are used to ensure that the two channels remain independent in the semantic space.

[0027] Composite loss function The expression is as follows:

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] in, Indicates early prediction of loss, This indicates the loss from dual-channel auxiliary monitoring. This represents an orthogonal decoupling constraint. Indicates hyperparameters, , Indicates the total number of samples; Indicates sample weights; Indicates the first The true label of a sample is 1 if the sample comes from a patient diagnosed with DIC, and 0 otherwise. Indicates the first The predicted probability of a sample is the probability value output by the model that the sample is expected to have a DIC within a preset time window in the future; Indicates the time point at which DIC was diagnosed; This indicates the current prediction time point, i.e., the current moment of this set of pathological data sequences; This represents the time decay coefficient, used to control the weights. Hyperparameters that vary over time; This represents the binary cross-entropy loss function, used to measure the difference between the predicted probability distribution of the model output and the true label distribution; This indicates the auxiliary classification head at the end of the channel. This indicates a false label promoting coagulation. This indicates a false label indicating bleeding. This is the peak gain coefficient. The larger the value, the greater the penalty when the model makes a wrong prediction.

[0034] Optionally, the viscoelastic coagulation test data should include at least the R time, K time, Angle angle, maximum amplitude MA value, and LY30 parameter from the thromboelastography (TEG) or rotational thromboelastography (ROTEM).

[0035] Secondly, the present invention provides a device for predicting the risk of disseminated intravascular coagulation, comprising:

[0036] The data acquisition module is used to acquire multimodal dynamic time-series data of the subject from the monitoring system and perform standardized preprocessing to obtain the original pathological data sequence; the multimodal dynamic time-series data includes vital signs data, coagulation index data and viscoelastic coagulation test data;

[0037] The data feature extraction module is used to extract the temporal statistical features of the original pathological data sequence using the sliding window technique; the temporal statistical features are used to characterize the overall level and trend of the original pathological data sequence.

[0038] The DIC index construction module is used to construct the DIC dynamic balance index based on the real-time dynamic imbalance between coagulation activation indicators and coagulation consumption indicators in the original pathological data sequence; the DIC dynamic balance index is used to quantify the dynamic imbalance state of the coagulation system.

[0039] The risk prediction module is used to construct multimodal feature vectors based on time-domain statistical characteristics and the DIC dynamic balance index. The multimodal feature vectors are input into a dual-channel decoupled network based on pathological adversarial mechanisms. Using orthogonal query vectors generated from static baseline features as semantic guidance, the multimodal feature vectors are mapped to physically isolated thrombosis and bleeding attention channels, respectively. The evolution of thrombosis and hyperfibrinolysis is captured through implicit pathological adversarial tensors, and the predicted risk probability of disseminated intravascular coagulation (DIC) in the target subject within a preset time window is output.

[0040] Thirdly, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0041] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0042] The present invention has at least the following beneficial effects:

[0043] By integrating multimodal dynamic time-series data, this approach overcomes the limitations and lag of traditional diagnosis, which relies on static data at a single time point, and can accurately capture the dynamic evolution trajectory of the disease. By constructing a dynamic equilibrium index for DIC, it explicitly quantifies the pathological imbalance between procoagulation and fibrinolysis, and combines it with a dual-channel decoupling network to achieve physical isolation of contradictory signals of thrombosis and bleeding, effectively solving the problem of feature confusion under complex disease courses. At the same time, it uses static baseline guidance to achieve context-aware personalized prediction, and supplements it with implicit adversarial tensors to capture hidden signs before abnormal indicators. From multiple dimensions such as data dimensions, pathological logic, and model architecture, it deeply fits the complex pathological nature of DIC, significantly improving the early sensitivity and global accuracy of DIC risk prediction. Attached Figure Description

[0044] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0045] Figure 1 This is a flowchart of a method for predicting the risk of disseminated intravascular coagulation in one embodiment of this application;

[0046] Figure 2 This is a structural diagram of a disseminated intravascular coagulation risk prediction device according to another embodiment of this application;

[0047] Figure 3 This is a structural diagram of the terminal device in another embodiment of this application;

[0048] Wherein, 200 represents the disseminated intravascular coagulation risk prediction device, 201 represents the data acquisition module, 202 represents the data feature extraction module, 203 represents the DIC index construction module, and 204 represents the risk prediction module; D10 represents the terminal device, D100 represents the processor, D101 represents the memory, and D102 represents the computer program running on D100. Detailed Implementation

[0049] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0051] It should be noted that the disseminated intravascular coagulation risk prediction method provided by this invention belongs to computer-based information processing technology. Its direct output is a "risk probability prediction value" of the disease occurring within a preset time window in the future, rather than a definite diagnostic conclusion. Essentially, it is a technical means of automatically associating, extracting features, and calculating models from multi-source heterogeneous pathological data. The entire process does not involve direct contact with the human body and aims to provide clinicians with prospective auxiliary decision support information, rather than replacing doctors in exercising the final medical judgment. Therefore, it does not fall under the disease diagnosis method in the sense of patent law.

[0052] Example 1

[0053] like Figure 1 As shown, the method for predicting the risk of disseminated intravascular coagulation provided by the present invention includes steps 11 to 14.

[0054] Step 11: Obtain multimodal dynamic time-series data of the subject from the monitoring system and perform standardized preprocessing to obtain the original pathological data sequence.

[0055] In this embodiment of the invention, the multimodal dynamic time-series data includes vital signs data, coagulation index data, and viscoelastic coagulation test data.

[0056] In one feasible implementation, multimodal dynamic time-series data of the subject (i.e., the patient) can be obtained from the monitoring system, laboratory information system (LIS), and electronic medical record (EMR) of a hospital's intensive care unit (ICU). For example, an interface service based on the HL7 v2 or FHIR R4 standard is configured to map key fields (hospitalization number, visit number, examination request number, specimen number, bed number, etc.) from the EMR, LIS, and ICU monitoring systems to a unified data dictionary, and a timestamp accurate to the minute is added to each record. A unique ICU hospitalization identifier, patient_uid, is generated by combining "hospital number + visit number + ICU admission time" to distinguish multiple hospitalizations or multiple ICU transfers / discharges of the same patient at different times and in different departments. This patient_uid field is added to all data from different systems to achieve cross-system association. For fields with the same meaning but inconsistent naming in different systems (such as SBP / systolic blood pressure / arterial systolic blood pressure), a unified mapping is implemented, using a unified unit and value range; laboratory items are uniquely coded according to "item code + test method" to distinguish test results with the same name but different methods.

[0057] In this embodiment of the invention, vital signs data include, but are not limited to, heart rate, respiratory rate, blood pressure (systolic, diastolic, and mean arterial pressure), blood oxygen saturation, and body temperature; these data are typically collected at high frequency by bedside monitors (e.g., per minute or per hour). Coagulation parameters include platelet count, prothrombin time, international normalized ratio, activated partial thromboplastin time, fibrinogen, D-dimer, fibrin(ogen) degradation products, and antithrombin III (AT-III). These data are obtained from laboratory tests. Viscoelastic coagulation test data are obtained using thromboelastography (TEG) or rotational thromboelastography (ROTEM). Viscoelastic coagulation test data include at least: reaction time (R time), clotting time (K time), coagulation angle (Angle angle), maximum amplitude (MA value), and fibrinolysis index LY30 (reflecting the percentage of clot dissolution at 30 minutes). These tests provide dynamic information on the entire coagulation process.

[0058] In practice, due to the different sampling frequencies of data from different sources, they need to be aligned to a unified time grid (e.g., at 1-hour intervals). For a given time grid point, if the original observation exists, it is used directly; if it is missing, it is estimated using linear interpolation or cubic spline interpolation based on the observations at the preceding and following time points to ensure the continuity of the time series. For low-frequency data such as laboratory indicators, a maximum tolerance time window (e.g., ±6 hours) can be set; missing points outside this window are not interpolated. After time series alignment and interpolation, to eliminate the influence of dimensional differences between different modal data, the values ​​of each indicator also need to be standardized, specifically using Z-score standardization.

[0059] All the indicator data, processed as described above and arranged in chronological order, are combined to form a multidimensional original pathological data sequence. ,in Indicates time A vector of all indicator values, , Indicates the length of the original pathological data sequence.

[0060] Step 12: Use the sliding window technique to extract the temporal statistical features of the original pathological data sequence.

[0061] In this embodiment of the invention, time-domain statistical features are used to characterize the overall level and trend of change of the original pathological data sequence.

[0062] In one feasible implementation, step 12 specifically includes steps 12.1 to 12.3.

[0063] Step 12.1: Define a fixed-length sliding window and extract fixed-length subsequences sequentially starting from the beginning of the sequence.

[0064] For example, setting the window length Hours, sliding step The subsequences extracted sequentially by hour are: .

[0065] Step 12.2: For each indicator within the window, calculate a set of time-domain statistics.

[0066] Specifically, for each indicator within the window (such as heart rate, platelet count, D-dimer, etc.), the mean (reflecting the average level of the indicator within the window period), standard deviation (reflecting the degree of fluctuation of the indicator within the window period), extreme values ​​(reflecting the extreme cases of the indicator), and linear trend slope (obtained by linearly fitting the data points within the window, which is used to determine the overall upward or downward trend of the indicator within the window period) are calculated.

[0067] Step 12.3: Concatenate the time-domain statistics calculated for each indicator to form a high-dimensional time-domain statistical feature vector.

[0068] For example, if there are Each indicator is calculated as follows: For each time-domain statistic, the time-domain statistical eigenvector is... The dimension is .

[0069] The sliding window technique can transform continuous time-series data into a series of fixed-length feature segments rich in statistical information. The extracted time-domain statistical features can effectively summarize the central tendency, dispersion, and direction of change of the data within a short window, providing a basic quantitative description of the coagulation system state for the model.

[0070] Step 13: Construct the DIC dynamic balance index based on the real-time dynamic imbalance between coagulation activation indicators and coagulation consumption indicators in the original pathological data sequence.

[0071] The core pathophysiology of DIC is the imbalance between the activation and depletion of the coagulation system. To explicitly quantify this dynamic imbalance, this invention constructs a dynamic balance index for DIC based on the relationship between coagulation activation and coagulation depletion indicators in the original pathological data sequence.

[0072] In one feasible implementation, key coagulation indicators are first extracted from the standardized raw pathological data sequence and divided into two opposing sets of indicators: a set of coagulation activation-related indicators. Set of indicators related to coagulation consumption Among them, the set of coagulation activation-related indicators Indicators reflecting the degree of thrombin generation and fibrin degradation were selected, including D-dimer and fibrin(ogen) degradation products; a set of coagulation consumption-related indicators were also included. Indicators reflecting the degree of coagulation substrate and platelet consumption were selected, including platelet count, fibrinogen, and antithrombin III (AT-III).

[0073] Subsequently, through the calculation formula

[0074]

[0075]

[0076]

[0077] DIC dynamic equilibrium index is obtained ;in, This represents a set of coagulation activation-related indicators; elevated levels of these indicators usually suggest a hypercoagulable state or hyperfibrinolysis. Indicates coagulation activation-related indicators, , Indicates the learnable weight parameters. Indicators related to coagulation activation At any moment Observations The result after Z-score standardization express The mean, express standard deviation This represents a set of indicators related to coagulation consumption; a decrease in these indicators suggests consumptive hypocoagulation. Indicators related to coagulation consumption At any moment Observations The result after Z-score standardization express The mean, express standard deviation Indicates the bias term. This represents the hyperbolic tangent activation function. Positive values ​​tend towards hypercoagulability / fibrinolysis. The negative value tends to indicate a depletion-related low condensation rate.

[0078] The value is , The closer the value is The stronger the trend of coagulation activation / fibrinolysis hyperactivity, the better. The closer the value is This indicates a stronger trend in the consumption of clotting factors / platelets. It's worth noting the DIC dynamic balance index. It is a composite feature with clear pathophysiological significance. It integrates multiple related laboratory indicators into a single comprehensive indicator, which not only simplifies the feature space, but also directly reflects the core contradiction in the DIC disease process—the imbalance between coagulation and fibrinolysis. This provides the model with strong prior knowledge guidance and significantly improves the robustness and interpretability of deep learning models in complex pathological environments.

[0079] Step 14: Construct a multimodal feature vector based on time-domain statistical features and the DIC dynamic equilibrium index. Input the multimodal feature vector into a dual-channel decoupled network based on pathological adversarial mechanism. Use the orthogonal query vector generated by static baseline features as semantic guidance to map the multimodal feature vector to the physically isolated thrombosis concern channel and bleeding concern channel, respectively. Capture the evolution process of thrombosis formation and hyperfibrinolysis through the implicit pathological adversarial tensor and output the predicted risk probability value of the subject to undergo disseminated intravascular coagulation within the future preset time window.

[0080] In this embodiment, the core logic of predicting the risk of disseminated intravascular coagulation lies in simulating the contradictory evolution of "thrombosis (hypercoagulability)" and "fibrinolysis (hypocoagulability / consumption)" in the pathological process of DIC.

[0081] Specifically, step 14 includes steps 14.1 to 14.5.

[0082] Step 14.1: The time-domain statistical feature sequence extracted in Step 12 is concatenated with the DIC dynamic balance index constructed in Step 13 in a dimensional manner to form a high-dimensional dynamic feature matrix that integrates macroscopic physiological trends and microscopic coagulation imbalance states. .

[0083] in, , Indicates the feature dimension.

[0084] Step 14.2: Use the static baseline features of the object to be tested as prior guidance to generate an orthogonal query vector.

[0085] First, we define two complementary medical knowledge mask matrices. (Coagulation-related mask) and (Bleeding-related mask). For example: for the "malignant tumor" feature, The corresponding bit is 1; for the characteristic of "history of gastrointestinal ulcers", The corresponding bit is 1.

[0086] Subsequently, the static feature vector Each element is multiplied with the mask element-wise and projected onto a specific query space to generate an orthogonal query vector; the orthogonal query vector includes the coagulation risk query vector. and bleeding risk query vector ;in, ; ; and Represents a binary mask. Indicates static baseline characteristics, The projection weight matrix for the coagulation risk query is a learnable parameter matrix used to map the masked static baseline features to the coagulation query vector space, so that the generated... In subsequent attention mechanisms, it is possible to specifically query and match features related to thrombosis in dynamic data; The projection weight matrix for the bleeding risk query is a learnable parameter matrix used to map the masked static baseline features to the bleeding query vector space, so that the generated... It can specifically query dynamic data for features related to hyperfibrinolysis or coagulation factor consumption.

[0087] Step 14.3, convert the high-dimensional dynamic feature matrix The input is fed into two parallel attention mechanism channels.

[0088] Specifically, the two parallel attention mechanism channels are the thrombosis attention channel and the bleeding attention channel. The thrombosis attention channel utilizes... As a query term, feature signals related to thrombosis and procoagulant activation are retrieved from the dynamic feature sequence, and a thrombosis characterization vector is output. Bleeding monitoring channel utilization As a query term, feature signals related to coagulation factor consumption and hyperfibrinolysis are retrieved from the dynamic feature sequence, and a bleeding characterization vector is output. .

[0089] In one feasible implementation, the expression for the thrombus concern channel is: ;in, This represents the thrombosis characterization vector, which is the final output of this channel. It represents the high-level features extracted by the model from multimodal dynamic data, reflecting the risk of coagulation activation and microthrombosis. Representing a multimodal feature vector, it is a dimensional concatenation of a time-domain statistical feature sequence and a DIC dynamic balance index. It integrates the time-domain features of the physiological trend of the test subject with the coagulation imbalance state, and is the core input data for the model to understand the dynamic evolution of the patient's condition. This represents the learnable key projection weight matrix of the thrombus concern channel, which is a trainable parameter matrix; The dimension of the key vector represents the query vector. Q and key vector K Dimension size; This represents the learnable projection weight matrix of the thrombus focus channel, which is also a trainable parameter matrix.

[0090] The expression for the bleeding concern channel is: ;in, The learnable key-projection weight matrix representing the bleeding concern channel is trainable and used to... The learnable projected weight matrix representing the bleeding concern channel is trainable and used to... .

[0091] Step 14.4, calculate the latent pathological adversarial tensor.

[0092] To capture the crucial "decompensation" signs in DIC, this invention performs feature difference operations to construct a latent pathological adversarial tensor. Specifically, the latent pathological antagonism tensor is obtained by performing feature difference operations on the output characteristics of the thrombosis concern channel and the output characteristics of the bleeding concern channel, and is used to characterize the degree of coagulation-fibrinolysis antagonism that cannot be reflected by a single clinical indicator.

[0093] In one feasible implementation, the latent pathology countertensor , This represents the characteristic difference operator. It should be noted that when the model simultaneously captures strong procoagulant signals ( High) and strong bleeding / consumption signals ( When the concentration is high, the resulting antagonistic perturbation value will increase significantly. This simulates the pathophysiological process of intense conflict between the coagulation and fibrinolytic systems before DIC enters the overt phase in clinical practice, and is key to capturing latent risks.

[0094] Step 14.5, prediction of disseminated intravascular coagulation risk.

[0095] Specifically, thrombosis characteristics bleeding symptoms Latent pathological countertensor and DIC dynamic balance index Multi-level fusion is performed, the input is fed into a multilayer perceptron (MLP), and finally mapped to a sigmoid activation function. The continuous values ​​between these ranges provide the probability of DIC occurring within a preset time window (e.g., the next 24 hours). .

[0096] By constructing a dual-channel decoupled network and using static baseline features to generate orthogonal query vectors for semantic guidance, the physical isolation and specific feature extraction of the two contradictory pathological processes of "thrombosis" and "bleeding tendency" were achieved. The latent pathological adversarial tensor further amplified the signal of imbalance between the two. This design deeply aligns with the complex and dynamically evolving pathological nature of DIC, enabling the model to capture the precursory information of disease deterioration more precisely and earlier, thereby significantly improving the accuracy and timeliness of prediction.

[0097] It should be noted that the method provided by this invention is essentially a computer-based data processing and pattern recognition technology. It acquires multimodal dynamic time-series data of the patient from the monitoring system through an automated interface, and sequentially performs a series of information processing steps, including standardized data preprocessing, temporal feature extraction, dynamic equilibrium index calculation, and feature mapping and fusion based on a specific neural network architecture. The final output is a predicted risk probability value for the onset of illness within a future time window. This process is entirely completed by the machine through calculations of digital signals and mathematical models, without involving direct contact with the human body or medical actions that determine the nature of the disease. Its output only provides clinicians with quantitative and prospective auxiliary reference information; the doctor's comprehensive clinical judgment remains a necessary prerequisite for making a final diagnosis and formulating a treatment plan. Therefore, this method does not fall under the disease diagnosis method in the sense of patent law.

[0098] Example 2

[0099] To enable those skilled in the art to reproduce and train the dual-channel decoupling network provided by the present invention, the training process of the dual-channel decoupling network in Embodiment 1 is described in this embodiment.

[0100] In this embodiment of the invention, by introducing multidimensional constraints during the model training phase, the dual-channel decoupling network is ensured to accurately distinguish between "procoagulant activation" and "fibrinolytic hemorrhage" features. Specifically, the training of the dual-channel decoupling network employs a composite loss function. The constraints are applied to the composite loss function, which includes an early prediction loss with time decay weights, a dual-channel auxiliary supervision loss, and orthogonal decoupling constraints.

[0101] In one feasible implementation, to address the issue that clinical diagnosis of DIC often lags behind pathological progression, the early prediction loss assigns higher weight to time points before diagnosis, forcing the model to learn the subtle features of the latency period. The expression for the early prediction loss is as follows:

[0102]

[0103]

[0104] in, , Indicates the total number of samples; Indicates sample weights; Indicates the first The true label of a sample is 1 if the sample comes from a patient diagnosed with DIC, and 0 otherwise. Indicates the first The predicted probability of a sample is the probability value output by the model that the sample is expected to have a DIC within a preset time window in the future; Indicates the time point at which DIC was diagnosed; This indicates the current prediction time point, i.e., the current moment of this set of pathological data sequences; This represents the time decay coefficient, used to control the weights. Hyperparameters that change over time. Through this weighting mechanism, the model faces a greater penalty for errors when predicting samples from 24-48 hours before diagnosis, thereby improving early warning capabilities. This is the peak gain coefficient. The larger the value, the greater the penalty when the model makes a wrong prediction.

[0105] To guide the thrombosis and bleeding attention channels to focus on the correct pathological signals, the dual-channel auxiliary monitoring loss uses medical heuristics to generate pseudo-labels.

[0106] In one feasible implementation, medical heuristic rules for generating pseudo-labels include:

[0107] Coagulation-promoting false labels If D-dimer or thrombin-anthirombin complex ,but ,otherwise 0.

[0108] Bleeding False Label : If the number of platelets or fibrinogen ,but ,otherwise 0.

[0109] The expression for the dual-channel auxiliary supervision loss is as follows:

[0110]

[0111] in, This represents the binary cross-entropy loss function, used to measure the difference between the predicted probability distribution of the model output and the true label distribution; This indicates the auxiliary classification head at the end of the channel. This indicates a false label promoting coagulation. This indicates a false label indicating bleeding.

[0112] This loss ensures that the "thrombosis pathway" focuses on coagulation activation characteristics, rather than mistakenly focusing on consumptive indicators such as platelet count decline.

[0113] In one feasible implementation, in order to prevent the features extracted from the two channels from becoming redundant or semantically confusing, the present invention introduces orthogonal decoupling constraints.

[0114] The expression for orthogonal decoupling constraints is as follows:

[0115]

[0116] This constraint forces the model to physically isolate the "activation" and "consumption" signals in the semantic space, so that the subsequent "latent pathological adversarial tensor" can accurately capture the difference signal between the two, that is, the degree of imbalance in coagulation function.

[0117] The final total loss is the weighted sum of the above three parts: . This represents hyperparameters.

[0118] In actual training, stochastic gradient descent (SGD) or the Adam optimizer are used to perform end-to-end joint training of all network parameters (including feature extraction weights, attention weights, classifier weights, and learnable weights in the DIC exponent) on the prepared training dataset. The total loss is minimized using the backpropagation algorithm. .

[0119] Example 3

[0120] This embodiment is based on a retrospective cohort study using a real intensive care unit (MIMIC) database. It elaborates on the data screening and feature engineering process, and compares the performance of the proposed method with mainstream machine learning models.

[0121] 3.1 Data sources and selection of research subjects.

[0122] Data for this study were obtained from the MIMIC public database (a single-center retrospective cohort). Initial screening: Data from 432 DIC patients were extracted. Inclusion and exclusion criteria: Patients admitted to the ICU for the first time were screened, leaving 347 patients; patients with a hospital stay of less than 24 hours were excluded, and 271 patients were finally included in the core study cohort. Primary outcome measure: 30-day all-cause mortality.

[0123] 3.2 Data preprocessing and missing value management.

[0124] 248 initial variables were extracted from the original database. After rigorous data cleaning, variables with no clinical significance or duplicate definitions were removed, leaving 134 variables. Variables with a missing rate >80% (33 in total) were directly removed. For variables with a missing rate <30%, median (for continuous variables) or mode (for categorical variables) were used for imputation. Based on ICD-9 codes and clinical records, etiology classification labels (etiology_class) were established: 1=Infection, 2=Obstetrics, 3=Oncology, 4=Trauma and Surgery, 5=Other.

[0125] 3.3 Multidimensional feature engineering and dimensionality reduction analysis.

[0126] To address the redundancy issue in high-dimensional medical data, this embodiment employs a coarse-to-fine "funnel-shaped" feature selection strategy: univariate analysis and principal component analysis (PCA) for dimensionality reduction. First, univariate analysis was performed on the selected variables, revealing 58 continuous variables significantly correlated with 30-day mortality (P<0.05). These 58 variables were then incorporated into principal component analysis (PCA). The first 10 principal components explained approximately 70% of the population variance, with the first 5 principal components exhibiting clear pathological indications. LASSO regression feature secondary screening used LASSO regression (minimum absolute contraction and selection operator) to regularize the above variables, and selected 12 key predictors at the optimal value of log(lambda): ① Scoring indicators: APSIII, SOFA; ② Vital signs: maximum body temperature (temperature_max), mean systolic blood pressure (sbp_avg), mean blood oxygen (spo2_avg); ③ Biochemistry and coagulation: anion gap (aniongap_avg), mean PTT, mean fibrinogen, erythrocyte distribution width (rdw_avg), maximum lactate dehydrogenase (ld_ldh_max), alkaline phosphatase (alp_avg); Other: etiology classification (etiology_class). Multivariate Logistic Regression Modeling: The variables selected by LASSO were incorporated into the multivariate logistic regression model, and six core features with independent predictive value were finally identified (P<0.05): SOFA score (OR=1.186, P<0.001); maximum body temperature (OR=0.443, P<0.001); red blood cell distribution width (RDW) (OR=1.113, P=0.025); lactate dehydrogenase (LDH) (OR=1.019, P=0.036); mean systolic blood pressure (SBP) (OR=0.975, P=0.022); and alkaline phosphatase (ALP) (OR=1.004, P=0.001).

[0127] 3.4 Model performance comparison experiment.

[0128] To verify the performance of the "dual-channel decoupling network based on pathological adversarial mechanism" proposed in this invention, it was compared with nine mainstream machine learning models constructed based on the above-mentioned screening features: Logistic Regression (LR) AUC was 0.802; GBM (Gradient Boosting Machine) AUC was 0.792; Random Forest (RF) AUC was 0.788; SVM (Support Vector Machine) AUC was 0.785; KNN (K-Nearest Neighbors) AUC was 0.774; XGBoost AUC was 0.766; Naive Bayes (NB) AUC was 0.757; Neural Network (NN) AUC was 0.741; and the basic transformer AUC was 0.832.

[0129] 3.5 Experimental Results.

[0130] Effectiveness of Feature Engineering: The six core metrics (SOFA, Temp, RDW, LDH, SBP, ALP) selected through PCA and LASSO encompass key information such as organ failure, infection response, coagulation loss, and tissue perfusion, enabling even a simple LR model (AUC=0.802) to achieve good results. However, the basic transformer model with an AUC of 0.832 performed best. Further improvements were made by adjusting model parameters, increasing network depth, and incorporating prior medical knowledge (i.e., the DIC dynamic equilibrium index and dual-channel decoupling). This not only utilized the selected core features but also captured the dynamic evolution relationship between LDH (tissue damage) and RDW / SBP (microcirculatory disturbance) through implicit adversarial tensors, achieving a further performance leap based on the basic transformer model.

[0131] Example 4

[0132] like Figure 2 As shown, an embodiment of the present invention provides a disseminated intravascular coagulation risk prediction device, the device 200 comprising:

[0133] The data acquisition module 201 is used to acquire multimodal dynamic time-series data of the subject from the monitoring system and perform standardized preprocessing to obtain the original pathological data sequence; the multimodal dynamic time-series data includes vital signs data, coagulation index data and viscoelastic coagulation test data;

[0134] The data feature extraction module 202 is used to extract the temporal statistical features of the original pathological data sequence using the sliding window technique; the temporal statistical features are used to characterize the overall level and trend of the original pathological data sequence.

[0135] The DIC index construction module 203 is used to construct the DIC dynamic balance index based on the real-time dynamic imbalance relationship between coagulation activation indicators and coagulation consumption indicators in the original pathological data sequence; the DIC dynamic balance index is used to quantify the dynamic imbalance state of the coagulation system.

[0136] The risk prediction module 204 is used to construct a multimodal feature vector based on time-domain statistical characteristics and the DIC dynamic balance index. The multimodal feature vector is input into a dual-channel decoupled network based on pathological adversarial mechanism. Using orthogonal query vectors generated from static baseline features as semantic guidance, the multimodal feature vectors are mapped to physically isolated thrombosis concern channels and bleeding concern channels respectively. The evolution process of thrombosis formation and hyperfibrinolysis is captured through implicit pathological adversarial tensors, and the predicted risk probability value of the subject to be tested for disseminated intravascular coagulation within a preset time window is output.

[0137] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that, for the sake of convenience and brevity, the division of the above-mentioned functional units and modules is only used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0138] like Figure 3 As shown, embodiments of the present invention provide a terminal device, such as... Figure 3 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 3 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0139] Specifically, when the processor D100 executes the computer program D102, it acquires multimodal dynamic time-series data of the subject from the monitoring system and performs standardized preprocessing to obtain the original pathological data sequence; it extracts the temporal statistical features of the original pathological data sequence using sliding window technology; it constructs a DIC dynamic balance index based on the real-time dynamic imbalance between coagulation activation indicators and coagulation consumption indicators in the original pathological data sequence; it constructs a multimodal feature vector based on the temporal statistical features and the DIC dynamic balance index, inputs the multimodal feature vector into a dual-channel decoupling network based on pathological adversarial mechanism, and uses an orthogonal query vector generated from static baseline features as semantic guidance to map the multimodal feature vector to physically isolated thrombosis concern channels and bleeding concern channels respectively. It captures the evolution process of thrombosis formation and hyperfibrinolysis through implicit pathological adversarial tensors and outputs the predicted risk probability value of the subject developing disseminated intravascular coagulation within a preset time window in the future. By integrating multimodal dynamic time-series data, it overcomes the limitations and lag of traditional diagnosis relying on static data at a single time point, enabling precise capture of the dynamic evolution trajectory of the disease. By constructing a dynamic equilibrium index for DIC, it explicitly quantifies the pathological imbalance between procoagulation and fibrinolysis, and combines it with a dual-channel decoupling network to achieve physical isolation of contradictory signals of thrombosis and bleeding, effectively solving the problem of feature confusion under complex disease courses. At the same time, it uses static baseline guidance to achieve context-aware personalized prediction, and supplements it with implicit adversarial tensors to capture hidden signs before abnormal indicators. It deeply fits the complex pathological nature of DIC from multiple dimensions such as data dimensions, pathological logic, and model architecture, significantly improving the early sensitivity and global accuracy of DIC risk prediction.

[0140] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0141] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0142] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0143] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0144] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0145] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for predicting the risk of disseminated intravascular coagulation, characterized in that, include: Multimodal dynamic time-series data of the subject to be tested are acquired from the monitoring system and standardized preprocessed to obtain the original pathological data sequence; the multimodal dynamic time-series data includes vital signs data, coagulation index data and viscoelastic coagulation test data; The time-domain statistical features of the original pathological data sequence were extracted using the sliding window technique; The time-domain statistical features are used to characterize the overall level and trend of change of the original pathological data sequence; A DIC dynamic balance index is constructed based on the real-time dynamic imbalance between coagulation activation and coagulation consumption indicators in the original pathological data sequence; the DIC dynamic balance index is used to quantify the dynamic imbalance state of the coagulation system; the expression of the DIC dynamic balance index is: in, Indicates the dynamic equilibrium index of DIC. This represents a set of coagulation activation-related indicators, including D-dimer and FDP. Indicates coagulation activation-related indicators, , Indicates the learnable weight parameters. Indicators related to coagulation activation At any moment Observations The result after Z-score normalization express The mean, express standard deviation This represents a set of coagulation-related indicators, including PLT, FIB, and AT-III. Indicators related to coagulation consumption At any moment Observations The result after Z-score normalization express The mean, express standard deviation Indicates the bias term. This represents the hyperbolic tangent activation function. Positive values ​​tend towards hypercoagulability / fibrinolysis. Negative values ​​tend to indicate depletion-related low condensation rates; A multimodal feature vector is constructed based on the time-domain statistical features and the DIC dynamic balance index. The multimodal feature vector is input into a dual-channel decoupling network based on pathological adversarial mechanism. An orthogonal query vector generated from static baseline features is used as semantic guidance to map the multimodal feature vector to physically isolated thrombosis concern channels and bleeding concern channels, respectively. The evolution process of thrombosis formation and hyperfibrinolysis is captured by the implicit pathological adversarial tensor, and the predicted risk probability value of the subject to be tested to develop disseminated intravascular coagulation within a future preset time window is output. The static baseline characteristics include demographic information, history of underlying diseases, severity score at admission, and presence of underlying causes that could induce disseminated intravascular coagulation. The orthogonal query vector includes a coagulation risk query vector. and bleeding risk query vector ;in, ; ; and Represents a binary mask. Indicates static baseline characteristics, The projection weight matrix for the coagulation risk query is a learnable parameter matrix used to map the masked static baseline features to the coagulation query vector space, so that the generated... In subsequent attention mechanisms, it is possible to specifically query and match features related to thrombosis in dynamic data; The projection weight matrix for the bleeding risk query is a learnable parameter matrix used to map the masked static baseline features to the bleeding query vector space, so that the generated... It can specifically query dynamic data for features related to hyperfibrinolysis or coagulation factor consumption; The thrombosis focus channel is used to search for signs of coagulation factor activation and microthrombus formation in multimodal data; the expression for the thrombosis focus channel is: ;in, This represents the thrombosis characterization vector, which is the final output of this channel. It represents the high-level features extracted by the model from multimodal dynamic data, reflecting the risk of coagulation activation and microthrombosis. Representing a multimodal feature vector, it is a dimensional concatenation of a time-domain statistical feature sequence and a DIC dynamic balance index. It integrates the time-domain features of the physiological trend of the test subject with the coagulation imbalance state, and is the core input data for the model to understand the dynamic evolution of the patient's condition. This represents the learnable key projection weight matrix of the thrombus concern channel, which is a trainable parameter matrix; The dimension of the key vector represents the query vector. and key vector Dimension size; This represents the learnable projected weight matrix of the thrombus concern channel, which is also a trainable parameter matrix. The bleeding monitoring channel is used to search for signs of coagulation factor depletion, hyperfibrinolysis, and a sudden drop in platelets; the expression for the bleeding monitoring channel is: ;in, The learnable key-projection weight matrix representing the bleeding concern channel is trainable and used to... The learnable projected weight matrix representing the bleeding concern channel is trainable and used to... .

2. The method for predicting the risk of disseminated intravascular coagulation according to claim 1, characterized in that, The latent pathological antagonism tensor is obtained by performing feature difference operations on the output features of the thrombosis concern channel and the output features of the bleeding concern channel, and is used to characterize the degree of coagulation-fibrinolysis antagonism that cannot be reflected by a single clinical indicator; the value of the latent pathological antagonism tensor is... , This represents the characteristic difference operator.

3. The method for predicting the risk of disseminated intravascular coagulation according to claim 2, characterized in that, The dual-channel decoupling network is constrained during training using a composite loss function. This composite loss function includes an early prediction loss with time decay weights, a dual-channel auxiliary supervision loss, and orthogonal decoupling constraints. The early prediction loss is used to improve the sensitivity of the dual-channel decoupling network to pre-diagnosis latency signals. The dual-channel auxiliary supervision loss is constructed based on medical heuristics. The orthogonal decoupling constraints are used to ensure that the two channels remain independent in the semantic space. Composite loss function The expression is as follows: in, Indicates early prediction of loss, This indicates the loss from dual-channel auxiliary monitoring. This represents an orthogonal decoupling constraint. Indicates hyperparameters, , Indicates the total number of samples; Indicates sample weights; Indicates the first The true label of a sample is 1 if the sample comes from a patient diagnosed with DIC, and 0 otherwise. Indicates the first The predicted probability of a sample is the probability value output by the model that the sample is expected to have a DIC within a preset time window in the future; Indicates the time point at which DIC was diagnosed; This indicates the current prediction time point, i.e., the current moment of this set of pathological data sequences; This represents the time decay coefficient, used to control the weights. Hyperparameters that vary over time; This represents the binary cross-entropy loss function, used to measure the difference between the predicted probability distribution of the model output and the true label distribution; This indicates the auxiliary classification head at the end of the channel. This indicates a false label promoting coagulation. This indicates a false label indicating bleeding. This is the peak gain coefficient. The larger the value, the greater the penalty when the model makes a wrong prediction.

4. The method for predicting the risk of disseminated intravascular coagulation according to claim 3, characterized in that, The viscoelastic coagulation test data shall include at least the R time, K time, Angle angle, maximum amplitude MA value, and LY30 parameter from the thromboelastography (TEG) or rotational thromboelastography (ROTEM).

5. A device for predicting the risk of disseminated intravascular coagulation, characterized in that, include: The data acquisition module is used to acquire multimodal dynamic time-series data of the subject from the monitoring system and perform standardized preprocessing to obtain the original pathological data sequence; the multimodal dynamic time-series data includes vital signs data, coagulation index data and viscoelastic coagulation test data; The data feature extraction module is used to extract the temporal statistical features of the original pathological data sequence using sliding window technology; The time-domain statistical features are used to characterize the overall level and trend of change of the original pathological data sequence; The DIC index construction module is used to construct a DIC dynamic balance index based on the real-time dynamic imbalance between coagulation activation and coagulation consumption indicators in the original pathological data sequence; the DIC dynamic balance index is used to quantify the dynamic imbalance state of the coagulation system; the expression of the DIC dynamic balance index is: in, Indicates the dynamic equilibrium index of DIC. This represents a set of coagulation activation-related indicators, including D-dimer and FDP. Indicates coagulation activation-related indicators, , Indicates the learnable weight parameters. Indicators related to coagulation activation At any moment Observations The result after Z-score normalization express The mean, express standard deviation This represents a set of coagulation-related indicators, including PLT, FIB, and AT-III. Indicators related to coagulation consumption At any moment Observations The result after Z-score normalization express The mean, express standard deviation Indicates the bias term. This represents the hyperbolic tangent activation function. Positive values ​​tend towards hypercoagulability / fibrinolysis. Negative values ​​tend to indicate depletion-related low condensation rates; The risk prediction module is used to construct a multimodal feature vector based on the time-domain statistical features and the DIC dynamic balance index. The multimodal feature vector is input into a dual-channel decoupled network based on pathological adversarial mechanism. Using an orthogonal query vector generated from static baseline features as semantic guidance, the multimodal feature vector is mapped to physically isolated thrombosis concern channels and bleeding concern channels respectively. The evolution process of thrombosis formation and hyperfibrinolysis is captured through implicit pathological adversarial tensors. The module outputs the predicted risk probability value of the subject to be tested developing disseminated intravascular coagulation within a preset time window in the future. The static baseline characteristics include demographic information, history of underlying diseases, severity score at admission, and presence of underlying causes that could induce disseminated intravascular coagulation. The orthogonal query vector includes a coagulation risk query vector. and bleeding risk query vector ;in, ; ; and Represents a binary mask. Indicates static baseline characteristics, The projection weight matrix for the coagulation risk query is a learnable parameter matrix used to map the masked static baseline features to the coagulation query vector space, so that the generated... In subsequent attention mechanisms, it is possible to specifically query and match features related to thrombosis in dynamic data; The projection weight matrix for the bleeding risk query is a learnable parameter matrix used to map the masked static baseline features to the bleeding query vector space, so that the generated... It can specifically query dynamic data for features related to hyperfibrinolysis or coagulation factor consumption; The thrombosis focus channel is used to search for signs of coagulation factor activation and microthrombus formation in multimodal data; the expression for the thrombosis focus channel is: ;in, This represents the thrombosis characterization vector, which is the final output of this channel. It represents the high-level features extracted by the model from multimodal dynamic data, reflecting the risk of coagulation activation and microthrombosis. Representing a multimodal feature vector, it is a dimensional concatenation of a time-domain statistical feature sequence and a DIC dynamic balance index. It integrates the time-domain features of the physiological trend of the test subject with the coagulation imbalance state, and is the core input data for the model to understand the dynamic evolution of the patient's condition. This represents the learnable key projection weight matrix of the thrombus concern channel, which is a trainable parameter matrix; The dimension of the key vector represents the query vector. and key vector Dimension size; This represents the learnable projected weight matrix of the thrombus concern channel, which is also a trainable parameter matrix. The bleeding monitoring channel is used to search for signs of coagulation factor depletion, hyperfibrinolysis, and a sudden drop in platelets; the expression for the bleeding monitoring channel is: ;in, The learnable key-projection weight matrix representing the bleeding concern channel is trainable and used to... The learnable projected weight matrix representing the bleeding concern channel is trainable and used to... .

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, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4.

Images