A method and system for monitoring postoperative embolism syndrome of liver cancer patients

Abstract

The application relates to a postoperative embolism syndrome monitoring method and system for a liver cancer patient, which comprises the following steps: acquiring multi-dimensional time sequence clinical data of the patient, interpolating missing values of the data through a probability regression model and generating uncertainty estimation for each interpolation point, constructing derived features based on the interpolated data, namely, calculating a time sequence change rate gradient of at least one core biochemical index and a time attenuation weighted fluctuation value of at least one vital sign index by taking the reciprocal of the uncertainty estimation as a weight, training a tree structure ensemble learning early warning model by using an asymmetric cost-sensitive loss function which is fused with a category imbalance penalty factor and a sample-specific weight determined by the time sequence change rate gradient of positive samples, inputting patient data to be warned into the trained model to output a risk probability, and triggering a high-level risk warning when the risk probability exceeds a preset threshold and key contribution features in model explainability analysis contain core inflammatory indicators as a subset of core biochemical indicators.

Description

Technical Field

[0001] This application belongs to the field of monitoring, and in particular relates to a method and system for monitoring postoperative embolism syndrome in patients with liver cancer. Background Technology

[0002] Hepatocellular carcinoma (HCC) is one of the most common malignant tumors worldwide, with high incidence and mortality rates. Hepatic artery chemoembolization (PEE) is a non-surgical treatment for intermediate-to-advanced HCC, but it often leads to embolism syndrome post-surgery, manifesting as fever, pain, nausea, and vomiting, which can cause liver failure and even death in severe cases. Clinical assessment of PES risk relies on physician experience and static monitoring of a few key biochemical indicators. Clinical data is inherently multidimensional time-series data with missing values. Processing methods such as mean or median imputation can introduce bias. Existing techniques often use raw clinical indicators as features, but these indicators are correlated. For example, the rate of change of key biochemical indicators and recent fluctuations in vital signs can reflect the trend of disease progression. While the incidence of PES is relatively low, training samples suffer from class imbalance. Although some methods employ cost-sensitive learning, they cannot differentiate the severity of different patients' conditions. Furthermore, many machine learning models lack interpretability in their predictions, making them difficult for clinicians to trust and adopt, and failing to provide a reliable basis for clinical decision-making. Summary of the Invention

[0003] To address the problem that existing technologies fail to fully utilize the deep information in data, neglect class imbalance in training samples, and thus cannot provide a reliable basis for clinical decision-making, in a first aspect of this invention, a method for monitoring postoperative embolism syndrome in patients with liver cancer is proposed, comprising:

[0004] Acquire multidimensional time-series clinical data of patients, use a probabilistic regression model to interpolate missing values ​​in the data, and generate uncertainty estimates for each interpolation point; Based on the interpolated data, derived features are constructed, including: calculating the time series rate of change gradient of at least one core biochemical indicator in the clinical data with the reciprocal of the uncertainty estimate as the weight; and calculating the time decay weighted fluctuation value of at least one vital sign indicator in the clinical data. Using a feature set containing the derived features, a tree-structured ensemble learning early warning model is trained through a cost-sensitive loss function, which incorporates a class imbalance penalty factor and sample-specific weights determined by the gradient of the time-series rate of change of positive samples. The data of patients awaiting warning are input into the trained warning model, which outputs the risk probability. When the risk probability exceeds the threshold and the model interpretability analysis shows that the key contributing features include core inflammatory indicators, a high-level risk warning is triggered.

[0005] In a second aspect of the invention, a postoperative embolism syndrome monitoring system for liver cancer patients is provided, comprising the following modules: The generation module is used to acquire multidimensional time-series clinical data of patients, interpolate missing values ​​in the data using a probabilistic regression model, and generate uncertainty estimates for each interpolation point. The calculation module is used to construct derived features based on the interpolated data. The derived features include: calculating the time series rate of change gradient of at least one core biochemical indicator in the clinical data with the reciprocal of the uncertainty estimate as the weight; and calculating the time decay weighted fluctuation value of at least one vital sign indicator in the clinical data. An integration module is used to train a tree-structured ensemble learning early warning model using a feature set containing the derived features, through a cost-sensitive loss function, which incorporates a class imbalance penalty factor and sample-specific weights determined by the time-series rate of change gradient of positive samples. The triggering module is used to input the data of patients to be warned into the trained warning model and output the risk probability. When the risk probability exceeds the threshold and the model interpretability analysis shows that the key contributing features include core inflammatory indicators, a high-level risk warning is triggered.

[0006] This application constructs derived features that profoundly reflect disease trends. For example, when calculating the gradient of the rate of change of core biochemical indicators, the uncertainty of data points is used as a weight, ensuring the appropriateness of trend judgment. Simultaneously, by calculating the time-decay weighted fluctuations of vital signs, the recent instability of the patient's condition is detected. The asymmetric cost-sensitive loss function not only solves the problem of sample class imbalance but also, by utilizing sample-specific weights determined by the rate of disease deterioration, enables the model to focus on critically ill patients with rapidly changing conditions, improving the early warning sensitivity for high-risk groups. Integrating the model's interpretability analysis into the early warning triggering mechanism requires that high-risk predictions be supported by key inflammatory indicators, enhancing the credibility of clinical applications. Attached Figure Description

[0007] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of Gaussian process regression interpolation; Figure 3 This is a schematic diagram of the specific weights for positive samples. Detailed Implementation

[0008] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0009] In the first embodiment, see Figure 1 A method for monitoring postoperative embolism syndrome in patients with liver cancer is proposed, specifically including: S1. Obtain multidimensional time-series clinical data of patients, use a probability regression model to interpolate missing values ​​in the data, and generate uncertainty estimates for each interpolation point; Postoperatively, clinical data were collected every 6 hours to obtain multidimensional time series data. This data included core biochemical indicators such as C-reactive protein (CRP), white blood cell count, and alanine aminotransferase (ALT), as well as vital signs such as body temperature, heart rate, and respiratory rate. For individual patients, if missing points existed in the CRP time series, a Gaussian process regression model was used for interpolation. Using time points as input and existing CRP values ​​as output, a Gaussian process regression model was trained. This model not only predicted CRP values ​​at missing time points (i.e., the mean of the posterior distribution) but also provided the variance of the predicted values, which served as an estimate of the uncertainty at the interpolation points.

[0010] In an optional embodiment, the step of interpolating missing values ​​in the data using a probabilistic regression model and generating uncertainty estimates for each interpolation point includes: A Gaussian process regression model is used, and the time series of each indicator is modeled using radial basis function kernels. The predicted mean of the model output at the missing time points is used as the interpolation, and the predicted variance is used as the uncertainty estimate.

[0011] For a specific indicator, such as C-reactive protein (CRP), we collect observations and corresponding time points over a period of time. Assume that at time points t=0, 6, and 24 hours, CRP observations are 10, 30, and 80 mg / L, respectively, but data is missing at t=12 hours. In this case, we construct a Gaussian process regression model to analyze the changes in CRP over time. The radial basis function kernel is of the form... This kernel function defines the correlation between data points based on the distance between time points, where the parameters... and It is learned from existing observation data.

[0012] After the model is built, it is applied to interpolation of missing data points. For the missing point at t=12 hours, the Gaussian process regression model will provide a predicted probability distribution based on the covariance relationship of existing observation points at 0, 6, and 24 hours. This distribution is also a Gaussian distribution. The mean of the distribution is used as the possible interpolation value. For example, if the model predicts a mean of 55 mg / L, then 55 will be used as the C-reactive protein value at t=12 hours. Simultaneously, the variance of the predicted distribution is used as an uncertainty estimate; for example, the predicted variance is 9. The variance value indicates the reliability of the interpolation result. The further away the time point is from the existing observation point, the higher the uncertainty of the interpolation is generally. Figure 2 .

[0013] S2, constructing derived features based on the interpolated data, the derived features including: calculating the time series rate of change gradient of at least one core biochemical indicator in the clinical data with the reciprocal of the uncertainty estimate as the weight; and calculating the time decay weighted fluctuation value of at least one vital sign indicator in the clinical data; To construct the time-series rate of change gradient of core biochemical indicators, C-reactive protein (CRP) was selected as the core indicator, and a 24-hour time window was set. Within the window, for each CRP data point, the weight was set to the reciprocal of the uncertainty estimate; for the original observations, the uncertainty is usually very small, and to simplify the calculation, the weight benchmark can be set to 1. A weighted linear regression was performed on the data points within the window, with time as the independent variable and CRP as the dependent variable. The slope of the resulting regression line is the time-series rate of change gradient of CRP. To construct the time-decay weighted fluctuation value of vital signs, body temperature was selected as the indicator, and a 48-hour time window was set. The square of the difference between each measured body temperature and the weighted average body temperature within the window was calculated, and each squared difference was assigned a weight that decays exponentially with time, with data closer to the current time point having a larger weight. The sum and square root of the weighted squared differences were then used to obtain the time-decay weighted fluctuation value of body temperature.

[0014] In an optional embodiment, calculating the time-series rate gradient of at least one core biochemical indicator in the clinical data, using the reciprocal of the uncertainty estimate as a weight, includes: C-reactive protein, procalcitonin, and white blood cell count were selected as the core biochemical indicators. Within a preset time window, the values ​​of each index are weighted by the reciprocal of the uncertainty estimate, and the weighted time series data points are fitted with linear regression. The slope of the fitted line is then used as the gradient of the time series rate of change.

[0015] Define a time window for analysis, such as the past 24 hours, and select C-reactive protein as the analysis target. Collect all C-reactive protein observations or interpolations within this window, along with corresponding uncertainty estimates. For example, the resulting data point sequence would be: Time... Corresponding value Uncertainty ;time Corresponding value Uncertainty ;time Corresponding value Uncertainty For data points obtained through direct measurement, the uncertainty is usually very small. To simplify the calculation, the weight benchmark can be set to 1, and the weight of the predicted points can be set to the reciprocal of the uncertainty estimate.

[0016] Calculate the weight for each data point, for example , , Perform a weighted linear regression analysis to find a straight line. The weighted data points are then fitted. The slope m of the fitted line represents the gradient of the C-reactive protein's rate of change over this 24-hour window. For example, if the calculated slope is 2.5, it indicates that the average growth rate of C-reactive protein is 2.5 mg / L per hour.

[0017] In an optional embodiment, calculating the time-decay-weighted fluctuation value of at least one vital sign indicator in the clinical data includes: Heart rate and respiratory rate were selected as vital signs indicators. The calculation formula is:

[0018] in, The time-decay weighted fluctuation value is... For time points The observed values ​​of the indicators, For weighted average, For time decay weight, For early warning reference time points, is the preset time decay constant, and N is the total number of observation points within the preset time period.

[0019] Calculate the time-decay weighted fluctuation value of heart rate and set the warning reference time point. The current time is the time decay constant. The timeframe is 24 hours, and all heart rate observations from the past 48 hours are selected. Assume there are 3 observation points during this period, and the data are as follows: Heart rate 48 hours ago =80; Heart rate 24 hours ago =110 Currently, heart rate =100. Calculate the time decay weight for each data point. , , , .

[0020] The weighted average of heart rate is calculated using the aforementioned weights. Multiply each heart rate value by its corresponding weight, sum them, and then divide by the sum of all weights. Based on the weighted average and weights, calculate the weighted squared difference between each data point and the weighted average. Sum all weighted squared differences, divide by the sum of weights, and then take the square root to obtain the time-decayed weighted variability value. The value represents the recent instability of heart rate; the more recent the observation, the greater its contribution to the volatility.

[0021] S3, using a feature set containing the derived features, train a tree structure ensemble learning early warning model through a cost-sensitive loss function, wherein the loss function combines a class imbalance penalty factor and a sample-specific weight determined by the gradient of the time series rate of change of positive samples; The interpolated raw clinical data, along with the time-series rate-of-change gradients of core biochemical indicators and the derived features from the fluctuation values ​​of vital signs, were combined to obtain a feature set. LightGBM was selected as the tree-structured ensemble learning early warning model. During model training, a loss function was defined. To address the issue of far fewer positive samples than negative samples, the loss for samples belonging to the positive category was multiplied by a class imbalance penalty factor, for example, set as the ratio of the number of negative to positive samples. For each positive sample, a sample-specific weight was calculated based on the time-series rate-of-change gradient of C-reactive protein. This weight is a monotonically increasing function of the gradient, for example, implemented through a scaled linear function, so that samples with larger gradients have larger weights. The total weight of each positive sample during training is equal to the product of the class imbalance penalty factor and the sample-specific weight, thus enabling the model to focus on critically ill cases with rapidly deteriorating conditions.

[0022] In an optional embodiment, the tree structure integrates a learning early warning model, which is a gradient boosting decision tree model.

[0023] The early warning model is an ensemble learning algorithm composed of multiple sequentially constructed decision trees. Instead of training multiple independent trees in parallel, it employs an iterative approach, where each new tree is built to correct the prediction residuals of all previous trees. For example, the model's input features include the previously calculated C-reactive protein gradient, heart rate fluctuation values, and a series of other indicators. The model initializes a prediction, typically the logarithmic odds of the incidence of embolic syndrome across all training samples.

[0024] Calculate the residuals between the current ensemble predictions of all trees and the true labels; these residuals represent the model's current prediction errors. Train a new, typically shallow, decision tree to fit these residuals. The new tree learns how to predict previous model errors based on input features. Add the newly trained tree to the model ensemble with a learning rate to update the overall predictions. Repeat this process hundreds or thousands of times, each iteration attempting to further reduce the prediction error. The predictions are a weighted sum of all decision tree predictions, which is then converted into a probability of embolism syndrome risk between 0 and 1 using a logistic function.

[0025] In an optional embodiment, the cost-sensitive loss function is calculated as follows:

[0026] Where M is the total number of training samples. For the true label of sample i, Let i be the probability that the model predicts sample i to be positive. This is the class imbalance penalty factor, and its value is equal to the ratio of the number of negative samples to the number of positive samples in the training set. For negative samples, ; For positive sample i, The sample-specific weights are obtained by normalizing the gradient of the time series change rate of C-reactive protein and mapping it to a preset weight interval using a monotonic mapping function.

[0027] The loss function is used to guide the construction of each new tree during the training process of the gradient boosting decision tree model, and calculates the global class imbalance penalty factor. Assuming the training set contains 9500 negative samples (non-embolic syndrome cases) and 500 positive samples (embolic syndrome cases), then =19. When calculating the loss, the loss term for all positive samples is multiplied by 19, which makes the model impose a large penalty on the behavior of missing positive samples during training.

[0028] Calculate the specificity weight for each positive sample i Obtain the time-series change rate gradient of C-reactive protein for all positive samples. Perform min-max normalization on the gradient values, scaling them to the interval between 0 and 1. Map the normalized gradient values ​​to a preset weight interval, such as 1 to 3, using a pre-defined monotonically increasing mapping function, such as the linear function f(x) = 1 + 2x. Higher C-reactive protein gradient values ​​indicate positive samples with faster disease progression. The larger the value, for example, closer to 3. During training, the total weight of a positive sample will be the global factor. and individual weights The product of these factors allows the model to focus not only on identifying cases of embolism syndrome, but also on critically ill cases with rapidly deteriorating disease indicators, such as... Figure 3 .

[0029] S4 inputs the data of patients to be warned into the trained warning model and outputs the risk probability; when the risk probability exceeds the threshold and the model interpretability analysis shows that the key contribution features include core inflammatory indicators, a high-level risk warning is triggered.

[0030] For a patient awaiting warning, clinical data is collected in real time, and data interpolation and feature derivation are performed according to steps S1 and S2. The processed feature vector is input into a trained LightGBM model, and the model outputs a value between 0 and 1, representing the risk probability of the patient developing embolism syndrome. The preset risk threshold is determined based on the model's performance on the validation set, for example, set to 0.7. When the model's output risk probability exceeds the threshold (e.g., 0.8), model interpretability analysis is performed. The SHAP algorithm is used to analyze the prediction results, calculating the contribution of each input feature to the risk probability. The list of the top five contributing key features is reviewed. C-reactive protein and procalcitonin are defined as core inflammatory markers. If C-reactive protein or a rate of change gradient appears in the list, the second condition is met. At this point, because the risk probability exceeds the threshold and the key attribution comes from core inflammatory markers, a high-level risk warning is triggered.

[0031] In an optional embodiment, triggering a high-level risk warning when the risk probability exceeds a threshold and model interpretability analysis shows that key contribution features include core inflammatory markers includes: The model interpretability analysis method is used to interpret the model's single prediction and obtain the preset number of features with the highest contribution. A high-level risk warning is triggered when the predicted risk probability is greater than the threshold and the contributing feature includes at least one of the core inflammatory indicators.

[0032] When the model performs a real-time assessment of a patient, it generates a risk probability of embolism syndrome. Assume the model outputs a risk probability of 0.8, while the preset risk threshold is 0.75. This meets the first condition for triggering a high-level warning. Instead of immediately issuing an alarm, the system proceeds to the verification process. The SHAP analysis tool is then invoked to analyze the prediction. SHAP calculates the contribution value for each feature and lists the top 5 features with the highest contribution. Assume the analysis results show that these 5 features, in descending order of contribution, are: C-reactive protein rate of change gradient, time-decay weighted variability of heart rate, current procalcitonin level, respiratory rate, and body temperature. This list is checked to determine if it contains preset core inflammatory markers. In this case, the first feature on the list is the C-reactive protein rate of change gradient, fully satisfying the second condition. Since both the risk probability and feature contribution conditions are met, the system confirms and triggers a high-level risk warning.

[0033] In a second embodiment, this application also proposes a postoperative embolism syndrome monitoring system for liver cancer patients, comprising the following modules: The generation module is used to acquire multidimensional time-series clinical data of patients, interpolate missing values ​​in the data using a probabilistic regression model, and generate uncertainty estimates for each interpolation point. The calculation module is used to construct derived features based on the interpolated data. The derived features include: calculating the time series rate of change gradient of at least one core biochemical indicator in the clinical data with the reciprocal of the uncertainty estimate as the weight; and calculating the time decay weighted fluctuation value of at least one vital sign indicator in the clinical data. An integration module is used to train a tree-structured ensemble learning early warning model using a feature set containing the derived features, through a cost-sensitive loss function, which incorporates a class imbalance penalty factor and sample-specific weights determined by the time-series rate of change gradient of positive samples. The triggering module is used to input the data of patients to be warned into the trained warning model and output the risk probability. When the risk probability exceeds the threshold and the model interpretability analysis shows that the key contributing features include core inflammatory indicators, a high-level risk warning is triggered.

[0034] In an optional embodiment, the step of interpolating missing values ​​in the data using a probabilistic regression model and generating uncertainty estimates for each interpolation point includes: A Gaussian process regression model is used, and the time series of each indicator is modeled using radial basis function kernels. The predicted mean of the model output at the missing time points is used as the interpolation, and the predicted variance is used as the uncertainty estimate.

[0035] In an optional embodiment, calculating the time-series rate gradient of at least one core biochemical indicator in the clinical data, using the reciprocal of the uncertainty estimate as a weight, includes: C-reactive protein, procalcitonin, and white blood cell count were selected as the core biochemical indicators. Within a preset time window, the values ​​of each index are weighted by the reciprocal of the uncertainty estimate, and the weighted time series data points are fitted with linear regression. The slope of the fitted line is then used as the gradient of the time series rate of change.

[0036] In an optional embodiment, calculating the time-decay-weighted fluctuation value of at least one vital sign indicator in the clinical data includes: Heart rate and respiratory rate were selected as vital signs indicators. The calculation formula is:

[0037] in, The time-decay weighted fluctuation value is... For time points The observed values ​​of the indicators, For weighted average, For time decay weight, For early warning reference time points, is the preset time decay constant, and N is the total number of observation points within the preset time period.

[0038] In an optional embodiment, the tree structure integrates a learning early warning model, which is a gradient boosting decision tree model.

[0039] In an optional embodiment, the cost-sensitive loss function is calculated as follows:

[0040] Where M is the total number of training samples. For the true label of sample i, Let i be the probability that the model predicts sample i to be positive. This is the class imbalance penalty factor, and its value is equal to the ratio of the number of negative samples to the number of positive samples in the training set. For negative samples, ; For positive sample i, The sample-specific weights are obtained by normalizing the gradient of the time series change rate of C-reactive protein and mapping it to a preset weight interval using a monotonic mapping function.

[0041] In an optional embodiment, triggering a high-level risk warning when the risk probability exceeds a threshold and model interpretability analysis shows that key contribution features include core inflammatory markers includes: The model interpretability analysis method is used to interpret the model's single prediction and obtain the preset number of features with the highest contribution. A high-level risk warning is triggered when the predicted risk probability is greater than the threshold and the contributing feature includes at least one of the core inflammatory indicators.

[0042] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0043] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method and system for monitoring postoperative embolism syndrome in patients with liver cancer, characterized in that, Includes the following steps: Acquire multidimensional time-series clinical data of patients, use a probabilistic regression model to interpolate missing values ​​in the data, and generate uncertainty estimates for each interpolation point; Based on the interpolated data, derived features are constructed, including: calculating the time series rate of change gradient of at least one core biochemical indicator in the clinical data with the reciprocal of the uncertainty estimate as the weight; and calculating the time decay weighted fluctuation value of at least one vital sign indicator in the clinical data. Using a feature set containing the derived features, a tree-structured ensemble learning early warning model is trained through a cost-sensitive loss function, which incorporates a class imbalance penalty factor and sample-specific weights determined by the gradient of the time-series rate of change of positive samples. The data of patients awaiting warning are input into the trained warning model, which outputs the risk probability. When the risk probability exceeds the threshold and the model interpretability analysis shows that the key contributing features include core inflammatory indicators, a high-level risk warning is triggered.

2. The method according to claim 1, characterized in that, The step of using a probabilistic regression model to interpolate missing values ​​in the data and generating uncertainty estimates for each interpolation point includes: A Gaussian process regression model is used, and the time series of each indicator is modeled using radial basis function kernels. The predicted mean of the model output at the missing time points is used as the interpolation, and the predicted variance is used as the uncertainty estimate.

3. The method according to claim 1, characterized in that, The step of calculating the time-series rate gradient of at least one core biochemical indicator in the clinical data, using the reciprocal of the uncertainty estimate as a weight, includes: C-reactive protein, procalcitonin, and white blood cell count were selected as core biochemical indicators. Within a preset time window, the values ​​of each index are weighted by the reciprocal of the uncertainty estimate, and the weighted time series data points are fitted with linear regression. The slope of the fitted line is then used as the gradient of the time series rate of change.

4. The method according to claim 1, characterized in that, The calculation of the time-decrease-weighted fluctuation value of at least one vital sign indicator in the clinical data includes: Heart rate and respiratory rate were selected as vital signs indicators. The calculation formula is: in, The time-decay weighted fluctuation value is... For time points The observed values ​​of the indicators, For weighted average, For time decay weight, For early warning reference time points, is the preset time decay constant, and N is the total number of observation points within the preset time period.

5. The method according to claim 1, characterized in that, The tree-structured integrated learning early warning model is a gradient boosting decision tree model.

6. The method according to claim 3, characterized in that, The cost-sensitive loss function is calculated using the following formula: Where M is the total number of training samples. For the true label of sample i, Let i be the probability that the model predicts sample i to be positive. This is the class imbalance penalty factor, and its value is equal to the ratio of the number of negative samples to the number of positive samples in the training set. For negative samples, ; For positive sample i, The sample-specific weights are obtained by normalizing the gradient of the time series change rate of C-reactive protein and mapping it to a preset weight interval using a monotonic mapping function.

7. The method according to any one of claims 1-6, characterized in that, When the risk probability exceeds a threshold and model interpretability analysis shows that key contribution features include core inflammatory markers, a high-level risk warning is triggered, including: The model interpretability analysis method is used to interpret the model's single prediction and obtain the preset number of features with the highest contribution. A high-level risk warning is triggered when the predicted risk probability is greater than the threshold and the contributing feature includes at least one of the core inflammatory indicators.

8. A monitoring system for postoperative embolism syndrome in patients with liver cancer, characterized in that, Includes the following modules: The generation module is used to acquire multidimensional time-series clinical data of patients, interpolate missing values ​​in the data using a probabilistic regression model, and generate uncertainty estimates for each interpolation point. The calculation module is used to construct derived features based on the interpolated data. The derived features include: calculating the time series rate of change gradient of at least one core biochemical indicator in the clinical data with the reciprocal of the uncertainty estimate as the weight; and calculating the time decay weighted fluctuation value of at least one vital sign indicator in the clinical data. An integration module is used to train a tree-structured ensemble learning early warning model using a feature set containing the derived features, through a cost-sensitive loss function, which incorporates a class imbalance penalty factor and sample-specific weights determined by the time-series rate of change gradient of positive samples. The triggering module is used to input the data of patients to be warned into the trained warning model and output the risk probability. When the risk probability exceeds the threshold and the model interpretability analysis shows that the key contributing features include core inflammatory indicators, a high-level risk warning is triggered.

9. The system according to claim 8, characterized in that, The step of using a probabilistic regression model to interpolate missing values ​​in the data and generating uncertainty estimates for each interpolation point includes: A Gaussian process regression model is used, and the time series of each indicator is modeled using radial basis function kernels. The predicted mean of the model output at the missing time points is used as the interpolation, and the predicted variance is used as the uncertainty estimate.

10. The system according to claim 8, characterized in that, The step of calculating the time-series rate gradient of at least one core biochemical indicator in the clinical data, using the reciprocal of the uncertainty estimate as a weight, includes: C-reactive protein, procalcitonin, and white blood cell count were selected as core biochemical indicators. Within a preset time window, the values ​​of each index are weighted by the reciprocal of the uncertainty estimate, and the weighted time series data points are fitted with linear regression. The slope of the fitted line is then used as the gradient of the time series rate of change.

Images