A method, system, storage medium and device for predicting multiple complications of sepsis

Abstract

The application discloses a kind of sepsis multiple complications prediction method, system, storage medium and equipment, belong to intelligent medical technical field, method includes the following steps: S1. acquisition sepsis patient multidimensional time series physiological data;S2. medical perception feature vector is constructed to the time series physiological data;S3. sepsis multiple complications prediction model is constructed and introduced physiological logic space-time composite mask matrix;S4. the output multiple complications prediction probability using hierarchical expert hybrid network H-MoE;S5. the model is trained using dynamic weighted loss function based on task uncertainty;S6. risk output and feature contribution degree explanation;The model of the application is introduced by physiological logic mask and H-MoE architecture, realizes the accurate, synchronous prediction of sepsis complex complications, can effectively filter redundant features, and reveal the complex correlation therebetween, provide quantitative decision support for clinician, to realize early intervention and individualized treatment.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and in particular to a method, system, storage medium, and device for predicting multiple complications of sepsis. Background Technology

[0002] Sepsis is a common disease in the intensive care unit (ICU) with a high rate of unpredictable mortality. Sepsis patients frequently develop various complications, such as acute respiratory distress syndrome (ARDS), multiple organ dysfunction syndrome (MODS), disseminated intravascular coagulation (DIC), septic shock, and acute kidney injury (AKI). These complications may occur simultaneously or sequentially, significantly increasing the risk of death, prolonging hospital stays, and increasing the consumption of medical resources. Therefore, accurately predicting the risk of these serious complications is crucial for improving the prognosis of sepsis patients.

[0003] With the development of medical informatics and artificial intelligence technologies, extensive research has been conducted on using deep learning and machine learning models combined with intensive care databases for disease risk assessment. Currently, many studies are dedicated to predicting the occurrence of sepsis or its single complications (such as acute kidney injury or shock). However, these existing technological solutions have significant drawbacks and limitations:

[0004] First, most existing studies focus on predicting single disease outcomes or a limited number of complication combinations, lacking comprehensive predictive studies on the eight common complications of sepsis. However, sepsis complications are multidimensional and complex, often affecting multiple organ systems simultaneously. Predicting only a single complication cannot fully reflect the patient's true risk status. Current technology lacks models that can comprehensively predict multiple complications at the same time, which may lead to biased clinical decision-making.

[0005] Secondly, existing models generally lack sufficient consideration of the complex interrelationships among multiple complications. Due to limitations in their model architecture, they struggle to simultaneously model shared information and unique characteristics across multiple complication tasks, thus limiting their ability to assess the overall health status of patients and impacting the scientific rigor and effectiveness of clinical decision-making.

[0006] Finally, existing multi-task prediction architectures mostly employ a simple hard-sharing parameter model, and the loss function typically uses fixed weights for each task. Because different complications vary greatly in data distribution and prediction difficulty, traditional methods are prone to biasing the model towards easier tasks, neglecting high-risk but rare complications. Furthermore, the lack of "expert" processing mechanisms for different organ systems makes it difficult to simultaneously extract specific features from multiple complications.

[0007] In summary, there is an urgent need in this field for a comprehensive prediction scheme for multiple complications of sepsis that can deeply integrate medical logic, adaptively balance the differences between multiple tasks, and has pathological interpretability. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems existing in the prior art and provide a method, system, storage medium, and device for predicting multiple complications of sepsis. This invention achieves accurate and collaborative prediction of multiple complications of sepsis by introducing physiological logic masks and hierarchical expert hybrid networks, and possesses pathological interpretability.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] Firstly, a method for predicting multiple complications of sepsis is provided, including the following steps:

[0011] S1. Collect multidimensional time-series physiological data of sepsis patients;

[0012] S2. Construct a medical perception feature vector from the time-series physiological data, which includes at least:

[0013] Calculate the variability characteristics of time-series physiological data;

[0014] For lactate, creatinine and blood pressure indicators, a piecewise activation function was set for nonlinear mapping based on the sequential organ failure assessment medical threshold and / or sepsis-related clinical guideline thresholds.

[0015] S3. Construct a prediction model for multiple complications of sepsis, wherein the prediction model for multiple complications of sepsis includes:

[0016] A linear projection layer is used to map all features at each time step to the model feature dimension and add positional encoding;

[0017] A shared feature extraction layer, connected to the linear projection layer, employs a Transformer architecture containing only an encoder, for extracting sepsis-related feature representations from the output of the linear projection layer;

[0018] Multiple independent task-specific prediction layers are connected to the shared feature extraction layer, and each task-specific prediction layer corresponds to a sepsis complication task to be predicted.

[0019] A sepsis multi-complication prediction model is used to input the medical perception feature vector into the Transformer encoder of the shared feature extraction layer to obtain a shared temporal representation, and a physiological-logical spatiotemporal composite mask matrix M is introduced into the self-attention layer to constrain the attention weights; wherein, the physiological-logical spatiotemporal composite mask matrix M is composed of a temporal mask matrix. With spatial mask matrix It is obtained by combining elements one by one, and at least satisfies the following generation rules:

[0020] Temporal masking suppression: Preset average physiological response lag thresholds for different organ systems. When the interval between the historical time point and the current predicted time point satisfy When the attention score is set to negative infinity or the normalized attention weight is set to zero, the attention score is set to negative infinity. At that time, according to the exponential time decay function Reduce the corresponding attention score, among which The attenuation coefficient corresponds to the organ system;

[0021] Spatial masking suppression: Organ association weights are obtained based on edge weights and / or path relevance between organ nodes in a medical knowledge graph. ,when When the score is below a preset threshold, attention scores across organ feature dimensions are suppressed or masked; when... When the threshold is higher than the preset threshold, the attention score across organ feature dimensions will not be suppressed or the suppression of the attention score across organ feature dimensions will be weakened.

[0022] The physiological logic spatiotemporal composite masking matrix M is used to suppress or mask the self-attention score, so that the masked temporal position or feature dimension does not participate or contributes less in attention normalization, so as to obtain a shared temporal representation constrained by physiological logic.

[0023] S4. In the task-specific prediction layer, a hierarchical expert hybrid network H-MoE is used to output the prediction probability of multiple complications, wherein the H-MoE includes at least:

[0024] The organ system expert layer sets up multiple system expert modules for different organ systems such as respiratory, circulatory, urinary and metabolic systems. Each system expert module takes the shared temporal representation as input and outputs the intermediate risk representation or task feature representation of the corresponding organ system.

[0025] The gated aggregation layer includes a gated network, which uses the shared temporal representation as the activation weights of each system expert module as input and output. The activation weights are normalized by Softmax and then weighted and aggregated to evaluate the outputs of each system expert module.

[0026] S5. The sepsis multi-complication prediction model is trained using a dynamic weighted loss function based on task uncertainty, wherein a learnable variance parameter is set for different sepsis complication tasks, and the loss weight of each sepsis complication task is adaptively adjusted by the learnable variance parameter.

[0027] S6. Using the trained sepsis multi-complication prediction model, output the risk of multiple complications, including acute respiratory distress syndrome and septic shock, for the current sepsis patient, and output the feature contribution interpretation consistent with the constraints of the physiological-logical spatiotemporal composite mask matrix M.

[0028] In some embodiments, the multidimensional time-series physiological data includes demographic information, vital signs data upon admission, laboratory test results, past medical history, and medication records.

[0029] In some embodiments, the variability features in S2 include at least one or more of the following: coefficient of variation, slope within a time window, or difference magnitude.

[0030] In some embodiments, the nonlinear mapping method in S2 specifically involves constructing a piecewise activation function for key indicators such as lactate, creatinine, and mean arterial pressure, based on medical thresholds determined by the Sequential Organ Failure Assessment (SOFA) scoring system and sepsis-related clinical guidelines. This mapping transforms the original continuous numerical values ​​into medically perceived feature vectors reflecting the severity of organ damage, thus preserving numerical trends while highlighting the pathological significance of physiological indicators crossing clinical warning lines.

[0031] In some embodiments, the segmented activation function in S2 includes at least a segmented linear, segmented sigmoid, or segmented ReLU mapping based on a sequential organ failure assessment threshold and / or a sepsis-related clinical guideline threshold, for constructing segmented activation characterizations of lactate, creatinine, and blood pressure, respectively.

[0032] In some embodiments, the organ association weights in S3 The edge weights or path relevance between organ nodes in the medical knowledge graph are determined and used to suppress or filter attention scores across organ system feature dimensions.

[0033] In some embodiments, the construction of the physiological-logical spatiotemporal composite mask matrix M combines medical knowledge with temporal dependencies:

[0034] Time dimension: Based on the average physiological lag time (Lag time) of responses triggered by different organ involvement during the pathophysiological process of sepsis, an exponential time decay function is set. For lags exceeding a preset threshold... Historical temporal features are masked, and their attention weights are reset to zero or assigned to minimum values. For features within a threshold, their attention scores decay exponentially with increasing time intervals, thereby capturing the most clinically relevant temporal evolution patterns.

[0035] Spatial Dimension: Based on a pre-built medical knowledge graph, the path correlation between different organ nodes such as respiratory, circulatory, and renal is calculated. For organ feature dimensions with low logical correlation in medicine, spatial masks are applied for suppression to ensure that the model can focus on combinations of organ indicators with physiological synergistic effects when learning feature representations.

[0036] In some embodiments, S4 applies a hierarchical hybrid expert network (H-MoE) to the task-specific prediction layer. This network contains multiple modules targeting specific organ systems (such as respiratory system experts, circulatory system experts, etc.). By calculating the activation weights of shared representations across different expert modules using a gating network, feature differentiation extraction and aggregation for different complications such as acute respiratory distress syndrome (ARDS) and septic shock are achieved.

[0037] In some embodiments, the gating network in S4 is a lightweight fully connected network, and the output is normalized by Softmax to obtain the activation weights of each system expert module.

[0038] In some embodiments, the dynamic weighted loss function in S5 includes a weighting term for the loss of each task and a regularization term related to the learnable variance parameter, so that the weights of each task are adaptively updated with the learnable variance parameter during training.

[0039] Secondly, a system for predicting multiple complications of sepsis is provided, including:

[0040] The system comprises a data collection module, a preprocessing module, a model building module, a model training module, and a prediction module, among which:

[0041] Each module is configured to perform the steps of the method described in the first aspect, and the model building module includes at least a mask attention submodule for generating and applying a physiological-logical spatiotemporal composite mask matrix to participate in the Transformer self-attention normalization computation, and a task-specific prediction submodule for performing hierarchical expert hybrid network H-MoE gated weighted aggregation.

[0042] Thirdly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the method for predicting multiple complications of sepsis as described in the first aspect.

[0043] Fourthly, an electronic device is provided, including a memory and a processor, wherein the memory stores computer instructions executable by the processor, characterized in that the processor executes a method for predicting multiple complications of sepsis as described in the first aspect when executing the computer instructions.

[0044] It should be further noted that the technical features corresponding to the above-mentioned options and embodiments can be combined or substituted with each other to form new technical solutions without conflict.

[0045] This invention is the first to accurately predict eight common high-risk complications of sepsis and reveal the complex interrelationships between these complications. Compared with existing technologies, its advantages are as follows:

[0046] 1. This invention is the first to simultaneously predict eight common high-risk complications of sepsis, breaking the limitation of existing technologies that can only predict one or a few complications. It can provide clinicians with a panoramic view of the patient's future risk and achieve a comprehensive assessment of the patient's condition.

[0047] 2. Unlike existing general preprocessing, this invention uses physiological variability and threshold mapping based on SOFA scores to give input features clear pathological significance, capture early clinical warning signals that are easily ignored by conventional models, and achieve a deep integration of medical logic and data-driven approaches.

[0048] 3. By introducing a physiological logic mask matrix, the organ correlation and pathological evolution lag in the medical knowledge graph are embedded into the attention mechanism, which effectively suppresses temporal noise and enhances the ability to model the causal evolution among complications.

[0049] 4. A hierarchical hybrid expert network (H-MoE) is adopted to set up dedicated expert modules for different organ systems. Combined with an uncertainty dynamic weighted loss function, the prediction accuracy of high-risk, low-frequency complications is ensured, and the model is avoided from being dominated by high-frequency tasks.

[0050] 5. It can quantify the contribution of each physiological indicator to the prediction results, breaking the "black box" limitation of deep learning and enabling doctors to quickly identify the core triggers that lead to increased risk, thus possessing extremely high clinical application value.

[0051] 6. The model structure of this invention can automatically learn and utilize the complex interrelationships between different complications. The shared feature extraction layer captures general, high-level features related to the pathophysiology of sepsis, which imply potential links between complications and contribute to a deeper understanding of the disease mechanisms of sepsis.

[0052] 7. The method of this invention can be seamlessly integrated into existing clinical decision support systems, enabling real-time processing and risk prediction of new patient data. By outputting quantified and personalized complication probabilities, it provides doctors with intuitive and reliable decision-making basis, facilitating early warning and personalized intervention, and ultimately improving patient prognosis. Attached Figure Description

[0053] Figure 1 This is a simplified flowchart of a method for predicting multiple complications of sepsis according to the present invention;

[0054] Figure 2 This is a schematic diagram illustrating the specific process of a method for predicting multiple complications of sepsis according to the present invention;

[0055] Figure 3 This is a diagram of the multi-complication prediction model architecture based on the Transformer encoder of this invention.

[0056] Figure 4 This is a schematic diagram of the Pearson correlation coefficients among the eight sepsis complications of this invention. Detailed Implementation

[0057] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] It should be noted that the defects in the solutions in the prior art are all the results of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of this application in the following text should be the inventors' contributions to this application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0059] In view of the technical problems pointed out in the background art, the present invention provides the following embodiments:

[0060] Reference Figure 1 In one exemplary embodiment, a method for predicting multiple complications of sepsis is provided, comprising the following steps:

[0061] S1. Collect multidimensional time-series physiological data of sepsis patients;

[0062] S2. Construct a medical perception feature vector from the time-series physiological data;

[0063] S3. Construct a prediction model for multiple complications of sepsis, wherein the prediction model for multiple complications of sepsis includes:

[0064] A linear projection layer is used to map all features at each time step to the model feature dimension and add positional encoding;

[0065] A shared feature extraction layer, connected to the linear projection layer, employs a Transformer architecture containing only an encoder, for extracting sepsis-related feature representations from the output of the linear projection layer;

[0066] Multiple independent task-specific prediction layers are connected to the shared feature extraction layer, and each task-specific prediction layer corresponds to a sepsis complication task to be predicted.

[0067] The sepsis multi-complication prediction model is used to input the medical perception feature vector into the Transformer encoder of the shared feature extraction layer to obtain a shared temporal representation, and introduces a physiological logic spatiotemporal composite mask matrix M in the self-attention to constrain the attention weights.

[0068] S4. A hierarchical expert hybrid network H-MoE is used in the task-specific prediction layer to output the prediction probability of multiple complications. The H-MoE includes at least an organ system expert layer and a gated aggregation layer.

[0069] S5. The sepsis multi-complication prediction model is trained using a dynamic weighted loss function based on task uncertainty, wherein a learnable variance parameter is set for different complication tasks, and the loss weight of each task is adaptively adjusted by the learnable variance parameter.

[0070] S6. Using the trained sepsis multi-complication prediction model, output the risk of multiple complications, including acute respiratory distress syndrome and septic shock, for the current sepsis patient, and output the feature contribution interpretation consistent with the mask matrix constraints.

[0071] In S1, dynamic data within 48 hours of ICU admission are retrieved from the Electronic Health Records (EHRs) system. This data includes vital signs (heart rate, mean arterial pressure, respiratory rate, oxygen saturation, body temperature, etc.), laboratory tests (lactate, creatinine, white blood cell count, platelets, bilirubin, etc.), and patient static characteristics (age, sex, underlying diseases, etc.). Data sampling is set at one hour per time step. The data is presented as multivariate time-series structured data, with each data point corresponding to one sepsis patient, and the data is stored in a tabular format.

[0072] In S2, beyond routine cleaning, the core function lies in constructing medical perception feature vectors, including:

[0073] The variability feature is calculated by using a sliding window to calculate the coefficient of variation, slope, or difference magnitude of key indicators over the most recent 6 hours to reflect the instability of the physiological system, serving as an enhanced feature reflecting the instability of the circulatory system.

[0074] Medical threshold nonlinear mapping (segmented activation) is used to set segmented activation functions for key physiological indicators based on SOFA medical thresholds or clinical guidelines. For example, for creatinine (Cr), its clinical normal value and the boundaries of different stages of failure are known. The segmented mapping function f(Cr) is defined as follows: if Cr < If 110 ≤ Cr < 170, it is mapped to 0 (normal); if Cr < 170, it is mapped to 1 (Level 1 risk), and so on. This mapping transforms the original numerical values ​​into discrete pathological characteristics, improving the model's sensitivity to the deterioration threshold.

[0075] In step S3, a prediction model for multiple complications of sepsis is constructed. The structure of the constructed model is as follows: Figure 3 As shown, the medical perception feature vector is input into the Transformer encoder. To enable the model to learn associations that conform to medical logic, a mask matrix is ​​introduced into the self-attention mechanism. , This is an element-wise product.

[0076] Temporal mask suppression ( ): Preset average physiological response lag thresholds for different organ systems When the interval between the historical time point and the current predicted time point satisfy When the attention score is set to negative infinity or the normalized attention weight is set to zero, the attention score is set to negative infinity. At that time, according to the exponential time decay function Reduce the corresponding attention score, among which This is the attenuation coefficient corresponding to the organ system.

[0077] Spatial mask suppression ( Organ association weights are derived based on the edge weights or path relevance between organ nodes in the medical knowledge graph. When the association degree is lower than a preset threshold, attention across organ dimensions is disabled, constraining the model to focus on the feature space with pathological associations.

[0078] In S4, a hierarchical hybrid expert network (H-MoE) is used to output the predicted probabilities. In the task-specific prediction layer, H-MoE is used to process shared temporal representations:

[0079] Organ system expert layer: Separate expert modules are set up for systems such as "respiratory, circulatory, urinary, and metabolic", and each expert performs feature mapping for the intermediate risks of their corresponding system.

[0080] Gated Aggregation Layer: The gated network (lightweight fully connected layer) takes shared representations as input and outputs the activation weights of each expert module. After Softmax normalization, the expert outputs are weighted and aggregated. The final output includes the probabilities of multiple complications, including acute respiratory distress syndrome and septic shock.

[0081] In S5, a dynamically weighted loss function based on task uncertainty is used to train the model. This loss function introduces a learnable variance parameter. For complications with low incidence and difficult to predict, the loss function will automatically adjust... To balance its contribution to the overall gradient, the model is prevented from failing on "difficult-to-predict" high-risk tasks due to the dominance of "easy-to-predict" tasks such as AKI. The model continuously updates parameters through backpropagation and the Adam optimizer until the validation set loss stabilizes. This dynamically weighted loss function can be expressed as:

[0082]

[0083] Where N is the total number of predicted complications (in this implementation, N=8, corresponding to 8 predicted tasks: acute respiratory distress syndrome, mechanical ventilation requirement, multiple organ dysfunction syndrome, disseminated intravascular coagulation, septic shock, death, acute kidney injury, and renal replacement therapy requirement). The cross-entropy loss is for the i-th complication task. The variance parameter is a learnable parameter for this task, and the weights of each task are adaptively adjusted through training.

[0084] The trained model was then evaluated using an independent test set to verify its generalization ability and prediction accuracy on unseen data.

[0085] In S5, the model is deployed into a clinical decision support system to receive clinical data from new patients in real time and output the probability of their future occurrence of various complications, providing doctors with quantitative decision-making basis.

[0086] Furthermore, a specific prediction process is given based on the above method.

[0087] like Figure 2 As shown, the "Data Input" step involves batch retrieving raw data from the database. Next, the "Data Preprocessing" step cleans and prepares the raw data, including first cleaning the data and removing patient data with a mean blood pressure (MBP) feature loss rate exceeding 20%. Outliers are corrected in principle. Missing value imputation employs a feature-based strategy: demographic features are imputed with the overall mean, and categorical features are imputed with the most frequent value. Specifically, for continuous physiological indicators, this invention not only uses the median for imputation but also performs the following enhancement operations:

[0088] For variability characterization, the system uses a 6-hour sliding window to scan heart rate and blood pressure sequences in real time, and calculates the coefficient of variation, slope, or difference amplitude within each window to capture early fluctuations in circulatory system decompensation.

[0089] A medical threshold nonlinear mapping (segmented activation) is used for lactate, creatinine, and blood pressure indicators, setting segmented activation functions based on SOFA medical thresholds or clinical guidelines. Taking creatinine (Cr) as an example, if its concentration is within the normal range, the mapping is 0; the mapping value increases for each SOFA score level, thus transforming continuous values ​​into discrete activation representations with clinical and pathological significance. Feature normalization uses Z-score standardization, with the formula: Where X is the original value, μ is the global mean, and σ is the global standard deviation. After preprocessing, a medical perception feature vector is constructed.

[0090] Then, in the "Model Building" section, a multi-task deep learning model that deeply integrates medical logic is constructed. First, a linear projection layer maps the features at each time step to a 256-dimensional model space, and sine-cosine positional encoding is added.

[0091] ,

[0092] In the shared feature extraction layer, this invention introduces a physiological logic spatiotemporal composite mask matrix within the Transformer encoder block. This matrix suppresses long-term noise through an exponential time decay function and filters the feature space of irrelevant organs based on the organ system correlation score matrix A from the medical knowledge graph.

[0093] ,in, This is a lag factor in the response to damage to different organs. This mechanism makes the self-attention mechanism no longer an undifferentiated fully connected mechanism, but a dynamically constrained model that focuses on physiological features with pathological logical connections.

[0094] Following this is the Hierarchical Hybrid Expert Network (H-MoE) prediction layer. After processing through six encoder blocks, the feature vectors enter the H-MoE prediction layer. The gating network calculates the activation weights of the four expert modules—respiratory, circulatory, urinary, and metabolic—in real time based on the medical perception feature vectors. Each expert module performs deep feature mapping for its corresponding systemic disease. Finally, the outputs of each expert are aggregated through weighted summation. Finally, the aggregated features are fed into eight independent prediction heads, which simultaneously output the future risk probabilities of eight complications through the Sigmoid activation function: mechanical ventilation (MV), septic shock, acute kidney injury (AKI), renal replacement therapy (CRRT), acute respiratory distress syndrome (ARDS), multiple organ dysfunction syndrome (MODS), disseminated intravascular coagulation (DIC), and death.

[0095] Next, the constructed model is trained using a dynamically weighted loss function based on task uncertainty. The model not only learns to predict probabilities but also learns the learnable variance parameters for each complication task. During the iteration process, if the prediction uncertainty of a certain complication task is large, the model will automatically adjust its weight in the total loss to prevent the training process from being dominated by tasks with large sample sizes and easy convergence, and to ensure that all eight tasks can achieve optimal performance.

[0096] Then, during training, the model automatically adjusts the weights of each task using the backpropagation algorithm: if a certain complication (with high prediction uncertainty) is detected, the model will automatically adjust the corresponding weights. To prevent the model from neglecting high-risk, low-frequency tasks in pursuit of accuracy on easy-to-learn tasks, the Adam optimizer updates parameters based on gradients. After each training epoch, the model's overall loss is calculated using the validation set. If the validation loss does not decrease significantly for 10 consecutive epochs, an early stopping mechanism is triggered, stopping training and saving the current model parameters as the final optimal model.

[0097] Finally, the model was validated using a series of indicators to ensure its stability and reliability. After this, it can be deployed to receive clinical data from new patients in real time and output the probability values ​​of various future complications, assisting clinicians in early clinical intervention. Specifically, to comprehensively evaluate the performance of the model, the area under the patient operating characteristic curve (AUROC) was used as the primary evaluation indicator. This indicator was chosen because it effectively quantifies the model's discriminative ability, reflecting the model's ability to differentiate between different outcome categories at different classification thresholds. This robust indicator comprehensively reflects the model's overall performance at all possible classification thresholds. Experiments showed that the model demonstrated robust performance in predicting all eight sepsis complication outcomes. The AUROC values ​​for all outcomes consistently reached a high level, with an average of 0.769. Specifically, the AUROC predicted by mechanical ventilation (MV) was 0.886 (95% confidence interval: 0.878–0.912), the AUROC predicted by continuous renal replacement therapy (CRRT) was 0.854 (95% confidence interval: 0.555–0.981), and the lowest AUROC predicted by septic shock was 0.704 (95% confidence interval: 0.670–0.779). To further improve the comprehensive evaluation of the overall model performance, this invention also introduces auxiliary indicators such as accuracy (ACC), precision, recall, and F1 score, which complement the main indicators to achieve a comprehensive evaluation of model performance.

[0098] Furthermore, to delve deeper into the clinical decision-making logic of the model, the experiment analyzed the top 15 most influential features and their corresponding weight distributions when predicting all eight complications. The results showed that there were significant differences in the key features of different outcomes, indicating that the model can identify differentiated prediction logic for a single complication and provide more targeted information for clinical decision support.

[0099] Furthermore, such as Figure 4As shown, the relationships among eight sepsis complications were quantified by calculating the Pearson correlation coefficient. The meta-analysis revealed varying degrees of association among these complications. Most notably, acute respiratory distress syndrome (ARDS) showed the most significant positive correlation with multiple organ dysfunction syndrome (MODS) (correlation coefficient r = 0.483). Statistically significant positive correlations were also found between death and MODS (r = 0.274), acute kidney injury (AKI) and MODS (r = 0.264), MODS and disseminated intravascular coagulation (DIC) (r = 0.259), and ARDS and death (r = 0.258). These results collectively demonstrate that complex and statistically significant associations exist among these key complications during the course of sepsis. This quantitative evidence further validates the robustness of the Transformer-based model constructed in this invention, proving its ability to effectively and accurately characterize these complex disease associations.

[0100] In another exemplary embodiment, based on the same inventive concept as the method, a method for predicting multiple complications of sepsis is provided, comprising:

[0101] The data collection module is used to collect multidimensional time-series physiological data of sepsis patients;

[0102] A preprocessing module is used to construct a medical perception feature vector from the time-series physiological data, which includes at least:

[0103] Calculate the variability characteristics of time-series physiological data;

[0104] For lactate, creatinine and blood pressure indicators, a piecewise activation function was set for nonlinear mapping based on the sequential organ failure assessment medical threshold and / or sepsis-related clinical guideline thresholds.

[0105] The model building module is used to build a prediction model for multiple complications of sepsis. The prediction model for multiple complications of sepsis includes:

[0106] A linear projection layer is used to map all features at each time step to the model feature dimension and add positional encoding;

[0107] A shared feature extraction layer, connected to the linear projection layer, employs a Transformer architecture containing only an encoder, for extracting sepsis-related feature representations from the output of the linear projection layer;

[0108] Multiple independent task-specific prediction layers are connected to the shared feature extraction layer, and each task-specific prediction layer corresponds to a sepsis complication task to be predicted.

[0109] A sepsis multi-complication prediction model is used to input the medical perception feature vector into the Transformer encoder of the shared feature extraction layer to obtain a shared temporal representation, and a physiological-logical spatiotemporal composite mask matrix M is introduced into the self-attention layer to constrain the attention weights; wherein, the physiological-logical spatiotemporal composite mask matrix M is composed of a temporal mask matrix. With spatial mask matrix It is obtained by combining elements one by one, and at least satisfies the following generation rules:

[0110] Temporal masking suppression: Preset average physiological response lag thresholds for different organ systems. When the interval between the historical time point and the current predicted time point satisfy When the attention score is set to negative infinity or the normalized attention weight is set to zero, the attention score is set to negative infinity. At that time, according to the exponential time decay function Reduce the corresponding attention score, among which The attenuation coefficient corresponds to the organ system;

[0111] Spatial masking suppression: Organ association weights are obtained based on edge weights and / or path relevance between organ nodes in a medical knowledge graph. ,when When the score is below a preset threshold, attention scores across organ feature dimensions are suppressed or masked; when... When the threshold is higher than the preset threshold, the attention score across organ feature dimensions will not be suppressed or the suppression of the attention score across organ feature dimensions will be weakened.

[0112] The physiological logic spatiotemporal composite masking matrix M is used to suppress or mask the self-attention score, so that the masked temporal position or feature dimension does not participate or contributes less in attention normalization, so as to obtain a shared temporal representation constrained by physiological logic.

[0113] The task-specific prediction layer employs a hierarchical expert hybrid network (H-MoE) to output the prediction probability of multiple complications, wherein the H-MoE includes at least:

[0114] The organ system expert layer sets up multiple system expert modules for different organ systems such as respiratory, circulatory, urinary and metabolic systems. Each system expert module takes the shared temporal representation as input and outputs the intermediate risk representation or task feature representation of the corresponding organ system.

[0115] The gated aggregation layer includes a gated network, which uses the shared temporal representation as the activation weights of each system expert module for input and output. The activation weights are normalized by Softmax and then used to perform weighted aggregation of the outputs of each system expert module.

[0116] The model training module is used to train the sepsis multi-complication prediction model using a dynamic weighted loss function based on task uncertainty. Learnable variance parameters are set for different sepsis complication tasks, and the loss weights for each sepsis complication task are adaptively adjusted by the learnable variance parameters.

[0117] The model prediction module is used to output the risk of multiple complications, including acute respiratory distress syndrome and septic shock, of the current sepsis patient using the trained sepsis multi-complication prediction model, and output the feature contribution interpretation consistent with the constraints of the physiological logical spatiotemporal composite mask matrix M.

[0118] The model building module includes at least a mask attention submodule for generating and applying a physiological-logical spatiotemporal composite mask matrix M to participate in the Transformer self-attention normalization calculation, and a task-specific prediction submodule for performing hierarchical expert hybrid network H-MoE gated weighted aggregation. This system can be integrated into a clinical decision support system (CDSS) to assist physicians in early intervention.

[0119] In another exemplary embodiment, based on the same inventive concept as the method, a computer-readable storage medium is provided, which stores a computer program. When executed by a processor, the computer program implements a method for predicting multiple complications of sepsis provided in this embodiment of the invention. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0120] In another exemplary embodiment, based on the same inventive concept as the method, an electronic device is provided, including a memory and a processor. The memory stores computer instructions that can be executed on the processor. When the processor executes the computer instructions, it performs a sepsis multiple complication prediction method based on a deep learning model provided in this embodiment of the invention.

[0121] The processor may be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present invention.

[0122] The embodiments of the subject matter and functional operation described in this specification can be implemented in: tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing device or for controlling the operation of a data processing device. Alternatively or additionally, the program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing device.

[0123] The processing and logic flow described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flow can also be executed by dedicated logic circuitry—such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits), and the device can also be implemented as dedicated logic circuitry.

[0124] Suitable processors for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. The basic components of a computer include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0125] It should be understood that each block in a flowchart or block diagram can represent a module, segment, or portion of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the figures. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0126] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A method for predicting multiple complications of sepsis, characterized in that, Includes the following steps: S1. Collect multidimensional time-series physiological data of sepsis patients; S2. Construct a medical perception feature vector from the time-series physiological data, which includes at least: Calculate the variability characteristics of time-series physiological data; For lactate, creatinine and blood pressure indicators, a piecewise activation function was set for nonlinear mapping based on the sequential organ failure assessment medical threshold and / or sepsis-related clinical guideline thresholds. S3. Construct a prediction model for multiple complications of sepsis, wherein the prediction model for multiple complications of sepsis includes: A linear projection layer is used to map all features at each time step to the model feature dimension and add positional encoding; A shared feature extraction layer, connected to the linear projection layer, employs a Transformer architecture containing only an encoder, for extracting sepsis-related feature representations from the output of the linear projection layer; Multiple independent task-specific prediction layers are connected to the shared feature extraction layer, and each task-specific prediction layer corresponds to a sepsis complication task to be predicted. A sepsis multi-complication prediction model is used to input the medical perception feature vector into the Transformer encoder of the shared feature extraction layer to obtain a shared temporal representation, and a physiological-logical spatiotemporal composite mask matrix M is introduced into the self-attention layer to constrain the attention weights; wherein, the physiological-logical spatiotemporal composite mask matrix M is composed of a temporal mask matrix. With spatial mask matrix It is obtained by combining elements one by one, and at least satisfies the following generation rules: Temporal masking suppression: Preset average physiological response lag thresholds for different organ systems. When the interval between the historical time point and the current predicted time point satisfy When the attention score is set to negative infinity or the normalized attention weight is set to zero, the attention score is set to negative infinity. At that time, according to the exponential time decay function Reduce the corresponding attention score, among which The attenuation coefficient corresponds to the organ system; Spatial masking suppression: Organ association weights are obtained based on edge weights and / or path relevance between organ nodes in a medical knowledge graph. ,when When the score is below a preset threshold, attention scores across organ feature dimensions are suppressed or masked; when... When the threshold is higher than the preset threshold, the attention score across organ feature dimensions will not be suppressed or the suppression of the attention score across organ feature dimensions will be weakened. The physiological logic spatiotemporal composite masking matrix M is used to suppress or mask the self-attention score, so that the masked temporal position or feature dimension does not participate or contributes less in attention normalization, so as to obtain a shared temporal representation constrained by physiological logic. S4. In the task-specific prediction layer, a hierarchical expert hybrid network H-MoE is used to output the prediction probability of multiple complications, wherein the H-MoE includes at least: The organ system expert layer sets up multiple system expert modules for different organ systems, including respiratory, circulatory, urinary and metabolic systems. Each system expert module takes the shared temporal representation as input and outputs the intermediate risk representation or task feature representation of the corresponding organ system. The gated aggregation layer includes a gated network, which uses the shared temporal representation as the activation weights of each system expert module as input and output. The activation weights are normalized by Softmax and then weighted and aggregated to evaluate the outputs of each system expert module. S5. The sepsis multi-complication prediction model is trained using a dynamic weighted loss function based on task uncertainty, wherein a learnable variance parameter is set for different sepsis complication tasks, and the loss weight of each sepsis complication task is adaptively adjusted by the learnable variance parameter. S6. Using the trained sepsis multi-complication prediction model, output the risk of multiple complications, including acute respiratory distress syndrome and septic shock, for the current sepsis patient, and output the feature contribution interpretation consistent with the constraints of the physiological-logical spatiotemporal composite mask matrix M.

2. The method according to claim 1, characterized in that, The variability features in S2 include at least one or more of the following: coefficient of variation, slope within the time window, or difference magnitude.

3. The method according to claim 1, characterized in that, The multidimensional time-series physiological data includes demographic information, vital signs data upon admission, laboratory test results, past medical history, and medication records.

4. The method according to claim 1, characterized in that, The segmented activation function in S2 includes at least a segmented linear, segmented sigmoid, or segmented ReLU mapping based on a sequential organ failure assessment threshold and / or a sepsis-related clinical guideline threshold, used to construct segmented activation characterizations for lactate, creatinine, and blood pressure, respectively.

5. The method according to claim 1, characterized in that, The organ association weight in S3 The edge weights or path relevance between organ nodes in the medical knowledge graph are determined and used to suppress or filter attention scores across organ system feature dimensions.

6. The method according to claim 1, characterized in that, The gated network in S4 is a lightweight fully connected network, and the output is normalized by Softmax to obtain the activation weights of each system expert module.

7. The method according to claim 1, characterized in that, The dynamic weighted loss function in S5 includes a weighting term for the loss of each task and a regularization term related to the learnable variance parameter, so that the weights of each task are adaptively updated with the learnable variance parameter during training.

8. A method system for predicting multiple complications of sepsis, characterized in that, include: The data collection module is used to collect multidimensional time-series physiological data of sepsis patients; A preprocessing module is used to construct a medical perception feature vector from the time-series physiological data, which includes at least: Calculate the variability characteristics of time-series physiological data; For lactate, creatinine and blood pressure indicators, a piecewise activation function was set for nonlinear mapping based on the sequential organ failure assessment medical threshold and / or sepsis-related clinical guideline thresholds. The model building module is used to build a prediction model for multiple complications of sepsis, wherein the prediction model for multiple complications of sepsis includes: A linear projection layer is used to map all features at each time step to the model feature dimension and add positional encoding; A shared feature extraction layer, connected to the linear projection layer, employs a Transformer architecture containing only an encoder, for extracting sepsis-related feature representations from the output of the linear projection layer; Multiple independent task-specific prediction layers are connected to the shared feature extraction layer, and each task-specific prediction layer corresponds to a sepsis complication task to be predicted. A sepsis multi-complication prediction model is used to input the medical perception feature vector into the Transformer encoder of the shared feature extraction layer to obtain a shared temporal representation, and a physiological-logical spatiotemporal composite mask matrix M is introduced into the self-attention layer to constrain the attention weights; wherein, the physiological-logical spatiotemporal composite mask matrix M is composed of a temporal mask matrix. With spatial mask matrix It is obtained by combining elements one by one, and at least satisfies the following generation rules: Temporal masking suppression: Preset average physiological response lag thresholds for different organ systems. When the interval between the historical time point and the current predicted time point satisfy When the attention score is set to negative infinity or the normalized attention weight is set to zero, the attention score is set to negative infinity. At that time, according to the exponential time decay function Reduce the corresponding attention score, among which The attenuation coefficient corresponds to the organ system; Spatial masking suppression: Organ association weights are obtained based on edge weights and / or path relevance between organ nodes in a medical knowledge graph. ,when When the score is below a preset threshold, attention scores across organ feature dimensions are suppressed or masked; when... When the threshold is higher than the preset threshold, the attention score across organ feature dimensions will not be suppressed or the suppression of the attention score across organ feature dimensions will be weakened. The physiological logic spatiotemporal composite masking matrix M is used to suppress or mask the self-attention score, so that the masked temporal position or feature dimension does not participate or contributes less in attention normalization, so as to obtain a shared temporal representation constrained by physiological logic. The task-specific prediction layer employs a hierarchical expert hybrid network (H-MoE) to output the prediction probability of multiple complications, wherein the H-MoE includes at least: The organ system expert layer sets up multiple system expert modules for different organ systems, including respiratory, circulatory, urinary and metabolic systems. Each system expert module takes the shared temporal representation as input and outputs the intermediate risk representation or task feature representation of the corresponding organ system. The gated aggregation layer includes a gated network, which uses the shared temporal representation as the activation weights of each system expert module as input and output. The activation weights are normalized by Softmax and then weighted and aggregated to evaluate the outputs of each system expert module. The model training module is used to train the sepsis multi-complication prediction model using a dynamic weighted loss function based on task uncertainty. Learnable variance parameters are set for different sepsis complication tasks, and the loss weights for each sepsis complication task are adaptively adjusted by the learnable variance parameters. The model prediction module is used to output the risk of multiple complications, including acute respiratory distress syndrome and septic shock, of the current sepsis patient using a trained sepsis multi-complication prediction model, and output the feature contribution interpretation consistent with the constraints of the physiological-logical spatiotemporal composite mask matrix M.

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

10. An electronic device comprising a memory and a processor, the memory storing computer instructions, characterized in that, When the processor executes the instructions, it implements the method described in any one of claims 1 to 7.

Images