Method, system and equipment for early prediction of disseminated intravascular coagulation

Abstract

The invention discloses an early prediction method suitable for disseminated intravascular coagulation. The method comprises the following steps: acquiring time sequence data of a target sample and extracting global statistical characteristics corresponding to the time sequence data; preprocessing the time sequence data to obtain a time sequence characteristic matrix required by model input; inputting the time sequence feature matrix and the global statistical features into a pre-trained prediction model; through a first Transform sub-module in the prediction model, extracting and aggregating from a time sequence feature matrix to obtain global time sequence features; performing fusion coding on the global time sequence features and the global statistical features of the corresponding samples through a second Transform sub-module in the prediction model to obtain global fusion features; and calculating the global fusion feature and outputting a DIC prediction probability value through a classification calculation sub-module in the prediction model. According to the method, the problem of insufficient time sequence dynamic change modeling of an existing prediction method is solved, the accuracy and the stability of a prediction result are enhanced, and the method has higher practical value.

Description

Technical Field

[0001] This invention relates to a method, system, and device for early prediction of disseminated intravascular coagulation (DIC). Specifically, it relates to an interpretable prediction model built on a multi-layer Transformer architecture, and a method, system, and device for early prediction of DIC after the prediction model has been trained. This invention belongs to the field of biomedical artificial intelligence technology. Background Technology

[0002] Disseminated intravascular coagulation (DIC) is a clinical syndrome characterized by bleeding and microcirculatory failure. It arises from various diseases and pathogenic factors that damage the microvascular system, leading to coagulation activation, systemic microvascular thrombosis, massive consumption of coagulation factors, and secondary hyperfiberemia. DIC is not a standalone disease but rather an intermediate step in the complex pathological process of numerous diseases, and its causes are numerous. High-risk groups include patients with acute severe infections, trauma, and those with malignant tumors or hematological disorders. Most cases of DIC have a rapid onset, complex progression, and are difficult to diagnose, with a poor prognosis. If not identified and treated promptly, it often endangers the patient's life, with a mortality rate as high as 30%–60%. Therefore, it is essential to conduct early prediction of disseminated intravascular coagulation (DIC) to provide a strong basis for clinical diagnosis and treatment decisions, take proactive preventive measures, promptly eliminate pathogenic factors (or triggers), and block or slow down the progression of the disease, thereby reducing the probability of high-risk individuals developing DIC.

[0003] In recent years, multi-index DIC scoring diagnostic systems have been developed. However, the accuracy and practicality of standard diagnoses remain a subject of widespread debate.

[0004] In recent years, artificial intelligence (AI) technology has been widely applied in clinical prediction and is gradually entering the complex field of disease diagnosis. Among these applications, AI-powered diagnosis, through deep learning (DL) capabilities, demonstrates speed and minimal human interference. Multiple studies have shown that machine learning (ML) algorithms, such as logistic regression, random forests, and gradient boosting decision trees, exhibit good performance in predicting disseminated intraepithelial neoplasia (DIC). AI technologies with deep learning (DL) capabilities, such as the Transformer architecture, have shown unique advantages in time-series data modeling. In particular, the multi-layered architecture of the Transformer possesses strong feature extraction capabilities and sensitivity to time-series dependencies, providing a foundation for building early DIC prediction models. However, AI diagnostic algorithms still rely on data pattern recognition and probability calculation tools and aids, resulting in a "black box" operation problem. The difficulty in explaining "why" a conclusion is reached is one of the main obstacles to its clinical application.

[0005] Existing technologies for early prediction of DIC (disseminated intravascular coagulation) mainly include the following: 1. Methods based on traditional scoring systems Currently, scoring systems such as ISTH and JAAM are widely used in clinical practice for the diagnosis of DIC. These methods typically use fixed thresholds and weighting rules based on the patient's laboratory test indicators (such as platelet count, fibrinogen level, and clotting time) to calculate a total score, thereby determining whether the patient is at risk of DIC. The advantages of this approach are its simplicity and extensive clinical validation. However, its fixed weights and thresholds cannot fully reflect the dynamic physiological changes of individual patients, and its predictive accuracy is significantly affected when data is missing or interference exists.

[0006] 2. Methods based on traditional statistics and machine learning models Besides scoring systems, some studies employ machine learning algorithms such as logistic regression, support vector machines, and random forests to model multivariate data and clinical events of ICU patients. While these methods can handle nonlinear relationships in the data to some extent and integrate some high-dimensional data features, they typically use static features or pre-defined time windows, making it insufficient for modeling dynamic changes in patients' physiological indicators and difficult to capture subtle and crucial early trends.

[0007] 3. Initial Attempts at Deep Learning Models With the increasing application of deep learning in time-series data modeling, some studies have introduced models such as RNN and LSTM to analyze ICU monitoring data, attempting to achieve early warning of DIC risk through modeling continuous time-series data. Although these models have improved the capture of dynamic information by traditional methods to some extent, they are often "black box" models, lacking intuitive interpretability, which limits their promotion and application in actual clinical practice.

[0008] In summary, existing DIC early prediction models built using artificial intelligence technology still have the following problems that urgently need to be addressed: (1) Insufficient modeling of temporal dynamic changes. Traditional scoring systems and static machine learning models mainly rely on data from a single point in time or a fixed time window, making it difficult to accurately reflect the dynamic changes of patients' physiological indicators over different time periods. This makes it possible to overlook early and subtle changes, thereby reducing the sensitivity and accuracy of DIC prediction.

[0009] (2) Limitations of Data Preprocessing and Feature Engineering. Current methods rely on traditional forward imputation or simple statistical completion strategies when dealing with missing values, noise, and high-dimensional data, making it difficult to fully explore the complex correlations and temporal dependencies contained in EHR data. The limitations of feature engineering are manifested in two aspects: First, feature selection relies on human experience, making it difficult to systematically cover potential important features or feature interactions, which can easily lead to the loss of key information; second, traditional feature engineering usually compresses continuous time series into a single statistic (such as mean, maximum, or magnitude of change), failing to preserve the complete temporal dynamic pattern, resulting in insufficient sensitivity of the model to early small fluctuations; third, when dealing with high-dimensional heterogeneous medical data, manually constructed features often fail to fully express the nonlinear dependencies and long-range relationships between different variables, thus limiting model performance.

[0010] (3) Insufficient interpretability of the model. Although existing deep learning models perform well in prediction, their "black box" nature makes it difficult for clinicians to understand why the model makes a specific judgment. Especially in the process of critical diagnosis and treatment decisions, the lack of intuitive explanations reduces the trust and acceptance of the model in practical applications. Therefore, it has not been used for clinical auxiliary diagnosis and risk screening for a long time and its practicality is poor.

[0011] For example, neither the coagulation early warning method based on disease sequences disclosed in patent application CN117497197A nor the coagulation early warning method, device, equipment, and medium based on disease sequences disclosed in patent application CN110770848A provide effective solutions to the aforementioned problems. Therefore, it is evident that the prior art does not provide technical solutions to the above or related problems. Summary of the Invention

[0012] This invention aims to provide an early prediction method for disseminated intravascular coagulation (DIC). By employing a two-layer Transformer architecture temporal neural network, it effectively captures temporal dependencies and dynamic changes at different time scales, thereby improving the sensitivity and accuracy of prediction. This addresses the problem of insufficient modeling of temporal dynamic changes in existing prediction methods, providing medical professionals with more accurate auxiliary diagnostic or DIC risk screening tools when patients exhibit early symptoms of DIC, enabling them to make more targeted and correct decisions for early intervention.

[0013] This invention is achieved through the following technical solution: A method for early prediction of disseminated intravascular coagulation, characterized in that it includes: S1. Obtain time-series data of the target sample based on the target patient's electronic health record, and extract its corresponding global statistical features. S2. Preprocess the time series data to construct a time series feature matrix; S3. Input the time-series feature matrix and the global statistical features into the pre-trained prediction model; S4. Global temporal features are obtained by extracting and aggregating them from the temporal feature matrix through the first Transformer submodule in the prediction model; S5. The global temporal features and the global statistical features of the corresponding samples are fused and encoded through the second Transformer submodule in the prediction model to obtain global fused features; S6. The global fusion features are classified and calculated using the classification calculation submodule within the prediction model, and the DIC prediction probability value is output.

[0014] Preferably, the preprocessing includes: Missing values ​​are handled by using forward imputation or random forest imputation methods to fill in missing values ​​in the data to obtain imputed time series data. Feature engineering operations are performed to extract temporal features that characterize the patient's intravascular coagulation status based on the imputed temporal data, in order to construct the temporal feature matrix.

[0015] Preferably, a heatmap is output simultaneously with the DIC prediction probability value; the heatmap is generated based on the gradient of the key layer in the calculation model of the DIC prediction probability value and using the gradient weighted class activation mapping method, and the heatmap is used to visually identify the key time segments that contribute the most to the DIC prediction result.

[0016] A training method for an early prediction model of disseminated intravascular coagulation includes: S101. Obtain time-series data as training samples based on the electronic health records of historical patient groups, and obtain the global statistical features and DIC diagnosis labels corresponding to the training samples. S102. Preprocess the training samples to obtain a temporal feature matrix; S103. Input the time-series feature matrix and the global statistical features into the prediction model to be trained to obtain the prediction result output by the prediction model; S104. Calculate the loss based on the difference between the prediction result and the DIC diagnosis label of the corresponding current sample, and optimize the model parameters based on the loss; The above steps are executed iteratively until the model training meets the set conditions, at which point a trained prediction model is obtained.

[0017] Preferably, the weighted cross-entropy loss function is used to calculate the loss in step S105.

[0018] Preferably, during the model training process, the prediction model is comprehensively evaluated using set evaluation metrics, including accuracy, precision, recall, or F1 score.

[0019] A system for early prediction of disseminated intravascular coagulation based on a multi-layer Transformer architecture temporal model includes a prediction calculation module, which integrates a pre-trained prediction model. The prediction model includes a first Transformer submodule, a second Transformer submodule, and a classification calculation submodule connected sequentially. The first Transformer submodule is configured to receive the input temporal feature matrix and output global temporal features representing long-term dependencies; The second Transformer submodule is connected to the first Transformer submodule and is configured to receive global temporal features and global statistical features corresponding to the current sample, and fuse the global temporal features and global statistical features corresponding to the current sample to output global fused features; The classification calculation submodule, connected to the second Transformer submodule, is configured to receive the global fusion features and perform calculations using its integrated classification head to output the DIC prediction probability value.

[0020] Preferably, the classification head includes a fully connected layer and a Softmax layer.

[0021] Preferred options also include: The data acquisition module is configured to acquire data samples from the received electronic health records of the target patients; The preprocessing module is communicatively connected to the data acquisition module and is configured to preprocess the data samples to obtain the time-series feature matrix and the global statistical features corresponding to the current sample. The interpretability module is communicatively connected to the prediction calculation module and is configured to calculate the gradient of the key layer in the model based on the DIC prediction probability value to generate a heat map. The results output module, connected to the early prediction module and the interpretability module, is configured to output the DIC prediction probability value and the heatmap.

[0022] A device for early prediction of disseminated intravascular coagulation (DIC) integrates the early prediction system for DIC as described above and is used for auxiliary diagnosis or risk screening of clinical DIC.

[0023] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention introduces a pre-trained temporal neural network-based model with a multi-layer Transformer architecture into the early prediction method for disseminated intravascular coagulation (DIC). The model performs operations including feature extraction, feature fusion, and classification calculations, thereby obtaining and outputting more accurate prediction results than existing methods, that is, results closer to the predicted actual disease outcome of the patient. Verification has shown that the prediction results obtained using this early DIC prediction method have a significantly higher accuracy than those of existing methods.

[0024] (2) By preprocessing the time series data, the present invention improves the quality of the input data, thereby effectively improving the running efficiency of the model and shortening the time required for prediction.

[0025] (3) This invention enhances the interpretability of the prediction method by outputting a heat map while outputting the predicted probability value of DIC. The heat map visually identifies the key time segments that contribute the most to the DIC prediction results, enabling medical workers to clearly identify the time period when the patient suffers from DIC and its corresponding characteristics. This avoids the drawbacks of "black box" operation, thereby enhancing the trust of medical workers, effectively reducing the workload of clinicians, and making the prediction method more beneficial for clinical promotion and application.

[0026] (4) The present invention uses time-series data obtained from the electronic health records of the historical patient group as training samples, and obtains the global statistical features and DIC diagnosis labels corresponding to the training samples, so that the prediction results of the early prediction model of disseminated intravascular coagulation trained by the training method can be verified. The model parameters are optimized by calculating the loss based on the difference between the prediction results and the DIC diagnosis labels of the corresponding current samples. Then, by iteratively executing the steps in the training method, the early prediction model of disseminated intravascular coagulation after training can output the DIC prediction probability value (i.e., the prediction result) based on the time-series data of the input target patient.

[0027] (5) The present invention sets the prediction calculation module in the early prediction system of disseminated intravascular coagulation and integrates the pre-trained early prediction model of disseminated intravascular coagulation in the prediction calculation module, so that the prediction system can output the DIC prediction probability value according to the input time-series feature matrix and global statistical features.

[0028] (6) This invention improves the interpretability of the prediction process in the prediction method by outputting a heatmap (generated by the gradient weighted class activation mapping method, i.e. Grad-CAM technology) synchronously with the DIC prediction probability value. That is, the key time segments that contribute the most to the DIC prediction result are visually identified by the heatmap, thereby solving the problem of "black box" operation in the existing prediction model, thus improving the acceptance of the prediction method and making it more conducive to promoting and facilitating clinical use.

[0029] (7) The present invention calculates the final fused features by a classification head composed of a fully connected layer and a Softmax layer, thereby obtaining the predicted probability value of DIC risk.

[0030] (8) In the training method of the prediction model, the present invention uses the time series data of historical patient groups as training samples, and introduces the global time series features and DIC diagnosis labels of the corresponding training samples during the model training process. Thus, the trained model can output the probability value of the corresponding DIC network prediction based on the time series data of the input target patient, thereby achieving the purpose of using the model to predict DIC. Preferably, in the training method, the loss is calculated based on the difference between the prediction result and the DIC diagnosis label of the corresponding current sample, thereby optimizing the model parameters and improving the accuracy of the prediction model. More preferably, the weighted cross-entropy loss function is used to calculate the loss, which can better highlight the influence of the loss corresponding features on the prediction result.

[0031] (9) The present invention improves the accuracy of the prediction model by using set evaluation indicators to comprehensively evaluate the prediction model during the model training process, so as to obtain conclusions that are more in line with clinical practice.

[0032] (10) The present invention sets up a prediction calculation module in a disseminated intravascular coagulation early prediction system based on a multi-layer Transformer architecture time series model, and the disseminated intravascular coagulation early prediction system based on a multi-layer Transformer architecture time series model integrates the pre-trained prediction model as described above, so that the prediction system can obtain a more accurate early prediction result suitable for early prediction of disseminated intravascular coagulation, thereby providing important data support for medical workers to make early intervention decisions; preferably, the classification calculation submodule of the prediction calculation module includes a fully connected layer and a Softmax layer.

[0033] (11) The present invention effectively improves the sensitivity and accuracy of the prediction system by setting a data acquisition module, a preprocessing module, an interpretability module and a result output module in the prediction system.

[0034] (12) The present invention integrates the above-mentioned disseminated intravascular coagulation early prediction system into a device for early prediction of disseminated intravascular coagulation, so that the prediction results output by it have higher sensitivity and accuracy, thereby increasing the trust of medical workers in its prediction results and making it more conducive to its clinical application and promotion.

[0035] In summary, this invention, through its dual-layer Transformer architecture, effectively captures long-term dependencies in the temporal data of target samples within the electronic health records of target patients. By fusing these dependencies with statistical features, it comprehensively characterizes the patient's intravascular coagulation status (including static and dynamic aspects), thereby significantly improving the accuracy of early symptom identification and prediction of disseminated intravascular coagulation (DIC). Experimental testing demonstrates that the described method for early prediction of DIC effectively improves the stability, sensitivity, and accuracy of prediction results. During model training, its prediction accuracy on historical patient datasets, such as the F1 score, reaches 0.935, a 15% improvement over traditional methods. This method not only addresses the insufficient modeling of temporal dynamic changes in existing technologies but also overcomes limitations in data preprocessing and feature engineering, as well as insufficient model interpretability. It is more widely accepted by clinicians and patients and can better assist clinicians in making timely and effective early intervention decisions when diagnosing DIC. Therefore, it has greater application and promotional value in clinical auxiliary diagnosis and risk prediction. Attached Figure Description

[0036] Figure 1 This is a flowchart of the prediction method in this invention; Figure 2 This is a flowchart of the training method in this invention; Figure 3 This is a schematic diagram of the prediction system in this invention and a diagram showing the data flow within it; Figure 4 This is a block diagram illustrating the data processing and flow in this invention; Figure 5 This is a schematic diagram illustrating the generation and visualization output of prediction results in this invention; Figure 6 This is a schematic diagram of the overall architecture of the classification head in this invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0038] Example 1: This embodiment discloses a method for early prediction of disseminated intravascular coagulation, including the following steps: S1. Obtain time-series data of the target sample based on the target patient's electronic health record, and extract its corresponding global statistical features; S2. Preprocess the time series data to construct a time series feature matrix; S3. Input the time-series feature matrix and the global statistical features into the pre-trained prediction model; S4. Global temporal features are obtained by extracting and aggregating them from the temporal feature matrix through the first Transformer submodule in the prediction model; S5. The global temporal features and the global statistical features of the corresponding samples are fused and encoded through the second Transformer submodule in the prediction model to obtain global fused features; S6. The global fusion features are classified and calculated using the classification calculation submodule within the prediction model, and the DIC prediction probability value is output.

[0039] Raw data: This originates from the electronic health records (raw data) of the target patients (primarily ICU patients), serving as the source of global information for these patients. Time-series data (also time-series monitoring data, dependent on time) is obtained from the raw data through data acquisition. This includes key physiological indicators, basic demographic information, and clinical event records of ICU patients. Simultaneously, global statistical features, i.e., static features independent of time, are extracted, including ICU hospital stay days, maximum values, average values, or standard deviations in clinical tests, and patient baseline information (age, gender), etc. From the time-series feature matrix; in this embodiment, the first Transformer submodule of the prediction model extracts and aggregates global time-series features representing long-term dependencies from the time-series feature matrix. These global time-series features include global dynamic trends (slow rise or fall), fluctuation patterns (with obvious peaks, which may correspond to specific clinical events), and coefficients of variation (with time windows) showing increased fluctuations within a certain time period. Combining these two aspects allows us to identify which fluctuations are not random but have a clear deterioration trend and are more unstable in specific periods. These trends or anomalies are crucial in supporting clinical decision-making. Therefore, introducing global fusion features in the probability value calculation of the prediction model is a crucial step. The global fusion features are obtained by fusing the global statistical features and the global temporal features, and are used to characterize the temporal features of the patient's intravascular coagulation state. This combines static and dynamic representations to accurately reflect the more comprehensive overall dynamic trend of the target patient, thereby enhancing the accuracy of probability value calculation.

[0040] Of course, if it is necessary to conduct relevant early risk screening for other patients who may have disseminated intravascular coagulation, the corresponding steps can also be used for data collection, extraction and fusion.

[0041] In this embodiment, to ensure the validity of the data, reduce prediction time, and improve prediction efficiency, the time series data is preprocessed. The preprocessing includes the following steps: 1. Missing value handling: Use forward imputation or random forest imputation methods to impute missing values ​​in the data to obtain imputed time series data; (1) Forward imputation is used to complete the continuity of time series data arranged in chronological order. Specifically, it includes: when a time series feature is missing in the observations of its consecutive time points, the value of that time point is replaced with the valid observation value of the feature at the previous time point; if the feature is missing at the starting time point, backward imputation or the feature's statistics in the overall sample (such as the mean or median) can be used to fill the gap. In this way, the continuity and trend consistency of the time series features can be maintained.

[0042] (2) Random forest imputation is used to handle features that are discontinuous or have a high proportion of missing values. Specifically, for a target feature A with missing values, other features without missing values ​​at the same time point are selected as input variables. The non-missing records of A are used as training data, and a random forest regression model is used to learn the mapping relationship between other features and A. After training, the model is used to predict the missing records of A to obtain imputation values. For cases where multiple features are missing, an iterative imputation strategy can be used, building a model for each feature sequentially and imputing values ​​step by step until all missing values ​​are filled.

[0043] 2. Feature engineering operation: Based on the imputed time series data, extract time series features to characterize the coagulation state to construct the time series feature matrix.

[0044] The imputed time-series data is organized into a three-dimensional feature matrix according to the number of features F, the number of consecutive observation days DT, and the number of time points per day T, denoted as: X∈R F×D×T

[0045] in: X is a three-dimensional temporal feature matrix composed of changes in the patient's multidimensional physiological indicators over time, used to characterize the coagulation-related dynamic information of the target patient throughout the entire observation window; R is the field of real numbers; F is the number of features; D is the number of consecutive days of observation; T represents the number of times per day; for example, 24 means data is collected once per hour.

[0046] For each feature f i Arrange the observations at T time points of each day in chronological order; arrange all features along the feature dimension F to form a complete three-dimensional time-series feature matrix.

[0047] The specific steps are as follows: Day 1 (1-24 hours)

[0048] Day N (1-24 hours)

[0049] Each element in the feature matrix represents the observation value of a certain feature at a specific number of days and time points. This matrix serves as the input to the first Transformer submodule, used to extract and aggregate global temporal features for subsequent fusion encoding.

[0050] The above method ensures that the data structure of the input model is complete and takes into account both temporal continuity and feature correlation, providing sufficient time-series information for DIC prediction.

[0051] To make this method more trustworthy to medical professionals and to more intuitively display the temporal features or time segments that play a crucial role in the prediction results, in this embodiment, the prediction result outputs at least one heatmap along with the DIC prediction probability value. The heatmap is generated based on the gradient of the key layer in the calculation model of the DIC prediction probability value and using the gradient-weighted class activation mapping method (i.e., Grad-CAM technology). The heatmap is used to visually identify the key time segments that contribute the most to the DIC prediction results, thereby making the prediction method based on the prediction model interpretable, avoiding the "black box" operation of existing prediction methods in deep learning models, improving the credibility of the prediction results, making it easier for medical professionals and patients to accept, and more conducive to clinical application and promotion.

[0052] Example 2: This embodiment discloses a training method for a prediction model applicable to the early stage of disseminated intravascular coagulation, including: S101. Obtain time-series data as training samples based on the electronic health records of historical patient groups, and obtain the global statistical features and DIC diagnosis labels corresponding to the training samples. S102. Preprocess the training samples to obtain a temporal feature matrix; S103. Input the time-series feature matrix and the global statistical features into the prediction model to be trained to obtain the prediction result output by the prediction model; S104. Calculate the loss based on the difference between the prediction result and the DIC diagnosis label of the corresponding current sample, and optimize the model parameters based on the loss; The above steps are executed iteratively until the model training meets the set conditions, at which point a trained prediction model is obtained.

[0053] The prediction model trained in this embodiment can be used as the pre-trained prediction model in Embodiment 1 and can implement the method described in Embodiment 1.

[0054] In this embodiment, to further optimize the model parameters of the prediction model, a weighted cross-entropy loss function is used to calculate the loss in step S105. The formula for the weighted cross-entropy loss function is:

[0055] in, N represents the number of training samples; C is the category; y i，c It is the true label of sample i; It is a predicted probability; w c These are class weights, used to address the problem of smoothing out class imbalance.

[0056] Of course, the optimization of model parameters can be performed before the prediction model is used, or it can be optimized within a time limit during use. Therefore, in order to ensure the accuracy of the prediction model, a set evaluation index is also used to comprehensively evaluate the prediction model. The evaluation index includes accuracy, precision, recall or F1 score, so as to ensure the accuracy of the prediction model and avoid distortion in the training process, which would affect the credibility of the prediction results output during use.

[0057] The accuracy rate refers to the overall correctness of the model's predictions, representing the percentage of correctly predicted samples out of the total sample. Its calculation formula is as follows:

[0058] in, TP (True Positive) is the number of correct samples whose model predicts a positive result: the actual result is DIC (i.e., the actual result is positive), and the prediction is also DIC (i.e., the prediction is positive). TN (True Negative) is the number of correct samples where the model predicts a negative result: the actual result is not DIC (i.e., the actual result is not), and the prediction is also not DIC (i.e., the prediction is not). FP (False Positive) is the number of incorrect samples where the model predicts a positive result: the actual result is not DIC (i.e., the actual result is not), but the prediction is DIC (i.e., the prediction is positive). FN (False Negative) is the number of incorrect samples where the model predicts a negative result: the actual result is DIC (i.e., the actual result is positive), while the predicted result is non-DIC (i.e., the predicted result is negative).

[0059] The advantage of evaluating a model using accuracy is its strong interpretability and easily understandable results, making it a good overall metric for balanced datasets. However, it can be highly misleading for severely imbalanced datasets, especially in predictive models used for DIC-assisted diagnosis and risk assessment. The overall accuracy of the model cannot accurately reflect the precision in DIC risk prediction, thus affecting the reliability of the predictive model. Therefore, in this embodiment, precision, recall, and F1 score are also introduced to compensate for the shortcomings of accuracy on imbalanced datasets, providing a more accurate evaluation of the predictive model's true performance in DIC risk prediction. Specifically, this includes: (1) The precision refers to the proportion of samples that are actually positive out of all samples predicted as positive. It focuses on the model's prediction results; it answers the question of how many samples are truly positive out of all samples predicted as positive by the model; its calculation formula is:

[0060] The precision rate measures the confidence level of a model's prediction of DIC (i.e., a positive prediction). A higher precision rate indicates a more reliable prediction when the model predicts DIC.

[0061] (2) The recall rate refers to the proportion of samples that are actually positive that are correctly predicted as positive. It focuses on the true situation and answers the question of how many samples that are actually positive were successfully predicted by the model. The calculation formula is as follows:

[0062] This measure is the predictive model's ability to detect DIC (i.e., predict the risk of DIC). A higher recall rate indicates that it is less likely to miss actual DIC patients. Given the high mortality rate of DIC, missing even one patient often leads to fatal clinical outcomes. Therefore, identifying as many real patients as possible to make the right decision for early intervention is a core focus in building this predictive model.

[0063] (3) The F1 score, or F1-Score, is the harmonic mean of precision and recall, which aims to find the balance point between precision and recall. Its calculation formula is as follows:

[0064] The F1 score is high only when both precision and recall are high, i.e., the probability of no false positives and false negatives is low. This high F1 score effectively assesses the reliability of the prediction model and promotes its clinical application.

[0065] Example 3: This embodiment discloses an early prediction system for disseminated intravascular coagulation based on a multi-layer Transformer architecture temporal model. The system includes a prediction calculation module, which integrates a pre-trained prediction model. The prediction model includes a first Transformer submodule, a second Transformer submodule, and a classification calculation submodule connected in sequence. The first Transformer submodule is configured to receive the input temporal feature matrix and output global temporal features representing long-term dependencies; The second Transformer submodule is connected to the first Transformer submodule and is configured to receive global temporal features and global statistical features corresponding to the current sample, and fuse the global temporal features and global statistical features corresponding to the current sample to output global fused features; The classification calculation submodule, connected to the second Transformer submodule, is configured to receive the global fusion features and perform calculations using its integrated classification head to output the DIC prediction probability value.

[0066] The system uses the method described in Example 1 to perform early prediction of disseminated intravascular coagulation and outputs a predicted probability value for DIC.

[0067] In this embodiment, the first Transformer submodule (i.e. Figure 5 The hourly Transformer Encoder shown includes a first neural network layer and a first multi-head self-attention mechanism layer connected in sequence. The first neural network layer extracts fine-grained temporal features representing short-term dependencies, such as key physiological indicator data in hourly units. The first multi-head self-attention mechanism layer aggregates these fine-grained temporal features daily to obtain global temporal features, which are then output to the second Transformer submodule. The key physiological indicator data includes platelet count, fibrinogen level, and clotting time.

[0068] In this embodiment, the second Transformer submodule (i.e. Figure 5 The Transformer Encoder shown includes a second neural network layer and a second multi-head self-attention mechanism layer, which are used to concatenate the global temporal features and the global statistical features in the feature dimension to form a fused feature vector, which is then input into the second Transformer submodule to obtain global fused features that represent long-term trends and overall changes, and output to the classification calculation submodule.

[0069] In this embodiment, as Figure 6 As shown, the classification calculation submodule integrates a classification head consisting of at least one fully connected layer and one Softmax layer.

[0070] The fully connected layer maps the global fusion features output by the second Transformer submodule to the category space; then the Softmax layer converts them into a probability distribution. The probability calculation formula is as follows: z = W·h fusion +b

[0071]

[0072] in, Z is the logits vector output by the fully connected layer; W is the weight matrix of the fully connected layer; h fusion It is the global fusion feature vector input to the classification head; B is the bias vector of the fully connected layer; p i It is the predicted probability of the i-th class output by the Softmax layer; e^(z_J) is the result of the exponentiation operation on the i-th element of the logits vector, and e^(z_J) is the exponentiation form of the j-th element of the logits vector.

[0073] In this embodiment, as Figure 4 and Figure 5 As shown, the early prediction system for disseminated intravascular coagulation further includes: a data acquisition module configured to acquire data samples from the received electronic health records of the target patient; The preprocessing module is communicatively connected to the data acquisition module and is configured to preprocess the data samples to obtain the time-series feature matrix and the global statistical features corresponding to the current sample. The interpretability module is communicatively connected to the prediction calculation module and is configured to calculate the gradient of the key layer in the model based on the DIC prediction probability value to generate a heat map. The result output module is connected to the prediction calculation module and the interpretability module, and is configured to output the DIC prediction probability value and the heat map.

[0074] In this embodiment, to enhance the trust of medical staff in the prediction system, an interpretability module is provided. The input of the interpretability module is communicatively connected to the second Transformer module, and its output is communicatively connected to the result output module. The interpretability module generates a heatmap based on the gradient of the key layer in the calculation model of the DIC prediction probability value and using the gradient weighted class activation mapping method (i.e., Grad-CAM technology, the same method used in the method described in Embodiment 1). The generated heatmap is transmitted to the result output module and output synchronously with the DIC prediction probability value. The heatmap is used to visually identify the key time segments that contribute the most to the DIC prediction result. Thus, the prediction method using the prediction model is interpretable, avoiding the "black box" operation of existing prediction methods in deep learning models, improving the credibility of the prediction results, making it easier for medical staff and patients to recognize and accept, and more conducive to the clinical application and promotion of the prediction system.

[0075] Example 4: This embodiment discloses an early prediction device for disseminated intravascular coagulation (DIC), which integrates the early prediction system for DIC as described in Embodiment 3, and is used for auxiliary diagnosis or risk screening of clinical DIC.

[0076] In this embodiment, the device includes an input port for data acquisition and an output port for outputting prediction results, and both the input port and the output port are network interfaces, wherein: The input port is connected to the hospital information system and is used to receive the time-series data and global statistical features from the hospital information system in real time.

[0077] The result output port is connected to a display for outputting and displaying the DIC predicted probability value and the heat map.

[0078] It should be noted that the early prediction models for disseminated intravascular coagulation (DIC) involved in the methods, systems, and devices for early prediction of DIC in this invention are all obtained by the training method described in Example 2, and the early prediction of DIC is achieved through the method described in Example 1.

[0079] Furthermore, in this invention, data collection and preprocessing are performed manually. Of course, they can also be performed using a Transformer architecture temporal neural network model, but they still need to be trained first to achieve the above-mentioned data collection and preprocessing functions. Therefore, whether it is done manually or through a trained temporal neural network model, it is applicable to this invention.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for early prediction of disseminated intravascular coagulation, characterized in that, include: S1. Obtain the time series data of the target sample and extract its corresponding global statistical features; S2. Preprocess the time series data to construct a time series feature matrix; S3. Input the time-series feature matrix and the global statistical features into the pre-trained prediction model; S4. Global temporal features are obtained by extracting and aggregating them from the temporal feature matrix through the first Transformer submodule in the prediction model; S5. The global temporal features and the global statistical features of the corresponding samples are fused and encoded through the second Transformer submodule in the prediction model to obtain global fused features; S6. The global fusion features are classified and calculated using the classification calculation submodule within the prediction model, and the DIC prediction probability value is output.

2. The prediction method according to claim 1, characterized in that, The preprocessing includes: Missing values ​​are handled by using forward imputation or random forest imputation methods to fill in missing values ​​in the data to obtain imputed time series data. Feature engineering operations are performed to extract temporal features that characterize the patient's intravascular coagulation status based on the imputed temporal data, in order to construct the temporal feature matrix.

3. The prediction method according to claim 1, characterized in that, A heatmap is output along with the DIC prediction probability value. The heatmap is generated based on the gradient of the key layer in the calculation model of the DIC prediction probability value and using the gradient weighted class activation mapping method. The heatmap is used to visually identify the key time segments that contribute the most to the DIC prediction result.

4. A training method for an early prediction model of disseminated intravascular coagulation, characterized in that, include: S101. Obtain time-series data as training samples based on the electronic health records of historical patient groups, and obtain the global statistical features and DIC diagnosis labels corresponding to the training samples. S102. Preprocess the training samples to obtain a temporal feature matrix; S103. Input the time-series feature matrix and the global statistical features into the prediction model to be trained to obtain the prediction result output by the prediction model; S104. Calculate the loss based on the difference between the prediction result and the DIC diagnosis label of the corresponding current sample, and optimize the model parameters based on the loss; The above steps are executed iteratively until the model training meets the set conditions, at which point a trained prediction model is obtained.

5. The training method according to claim 4, characterized in that, In step S105, the weighted cross-entropy loss function is used to calculate the loss.

6. The training method according to claim 4, characterized in that, During the model training process, the prediction model is comprehensively evaluated using set evaluation metrics, including accuracy, precision, recall, or F1 score.

7. A system for early prediction of disseminated intravascular coagulation based on a multi-layer Transformer architecture time-series model, characterized in that, It includes a prediction calculation module, which integrates a pre-trained prediction model. The prediction model includes a first Transformer submodule, a second Transformer submodule, and a classification calculation submodule connected in sequence. The first Transformer submodule is configured to receive the input temporal feature matrix and output global temporal features representing long-term dependencies; The second Transformer submodule is connected to the first Transformer submodule and is configured to receive global temporal features and global statistical features corresponding to the current sample, and fuse the global temporal features and global statistical features corresponding to the current sample to output global fused features; The classification calculation submodule, connected to the second Transformer submodule, is configured to receive the global fusion features and perform calculations using its integrated classification head to output the DIC prediction probability value.

8. The early prediction system for disseminated intravascular coagulation according to claim 7, characterized in that, The classification head includes a fully connected layer and a Softmax layer.

9. The early prediction system for disseminated intravascular coagulation according to claim 7, characterized in that, Also includes: The data acquisition module is configured to acquire data samples from the received electronic health records of the target patients; The preprocessing module is communicatively connected to the data acquisition module and is configured to preprocess the data samples to obtain the time-series feature matrix and the global statistical features corresponding to the current sample. The interpretability module is communicatively connected to the prediction calculation module and is configured to calculate the gradient of the key layer in the model based on the DIC prediction probability value to generate a heat map. The result output module is connected to the prediction calculation module and the interpretability module, and is configured to output the DIC prediction probability value and the heat map.

10. A device for early prediction of disseminated intravascular coagulation, characterized in that, It integrates the early prediction system for disseminated intravascular coagulation as described in claims 7 to 9, and is used for the auxiliary diagnosis or risk screening of clinical DIC.

Images