A method for constructing a comprehensive intelligent early warning system for trauma patients

Abstract

The present invention provides a method for constructing a comprehensive intelligent early warning system for trauma patients, which relates to the field of medical artificial intelligence technology, including processing multimodal data through a feature engineering framework, extracting multi-scale time series features and correlation features; using a deep learning prediction framework to identify abnormal states and predict complication risks, and performing feature fusion through an attention mechanism; constructing a medical knowledge graph to obtain a vector representation of domain rule knowledge; and generating prediction results through integrated verification through a bidirectional mapping mechanism. The present invention can achieve accurate early warning of the progression of the condition of trauma patients, improve the accuracy and interpretability of early warnings, and provide effective support for clinical decision-making.

Description

Technical Field

[0001] The present invention relates to medical artificial intelligence technology, and in particular to a method for constructing a comprehensive intelligent early warning system for trauma patients. Background Art

[0002] Trauma patients often face the risk of sudden changes in their condition and complications during treatment, and require timely warning and intervention. Existing warning systems are mainly based on the analysis of a single type of vital sign data or clinical indicators, and trigger warnings by setting fixed thresholds. There are also the following major problems:

[0003] Existing early warning systems often only focus on data anomalies at a single point in time, ignoring the dynamic trends and interactions between indicators, and are unable to effectively capture early warning signals during the evolution of the disease. This approach is prone to cause early warning delays and false alarms.

[0004] The current early warning models generally adopt a unified scoring standard and fixed threshold, which fails to fully consider the individual differences of patients and disease characteristics, making it difficult to achieve personalized and accurate early warning. For different types of trauma patients, their risk manifestations and early warning needs vary significantly.

[0005] The existing system lacks deep integration of medical expertise, and the early warning results lack interpretability and clinical guidance. Relying solely on data-driven prediction models cannot effectively combine professional experience and treatment standards accumulated in clinical practice, affecting the reliability and practicality of early warning results. Summary of the invention

[0006] The embodiment of the present invention provides a method for constructing a comprehensive intelligent early warning system for trauma patients, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] A method for constructing a comprehensive intelligent early warning system for trauma patients is provided, comprising:

[0009] Through the feature engineering framework, multimodal data processing is performed on vital signs data, clinical test data, imaging examination data and electronic medical record data to extract multi-scale time series features of vital signs fluctuation trends and clinical indicator change rates, and a feature interaction network is constructed to obtain the correlation features of multi-dimensional indicator interaction relationships and comorbidity correlations.

[0010] A deep learning prediction framework is constructed, and the multi-scale time series features are used to identify abnormal states using a recurrent neural network and a dynamic baseline calculation method, and the multi-scale time series features and associated features are used to predict complication risks using a graph neural network; the abnormal state identification results and complication risk prediction results are adaptively fused through an attention mechanism, and the attention mechanism constructs a personalized weight matrix based on individual patient characteristics, and a dynamic threshold mechanism is used to determine the warning level;

[0011] Construct a medical knowledge graph to convert the domain rule knowledge of drug interaction networks and disease complication association networks into low-dimensional vector representations;

[0012] The warning level is integrated and verified with the medical knowledge graph through a bidirectional mapping mechanism to generate prediction results including risk level, warning reasons and intervention suggestions.

[0013] In an optional embodiment,

[0014] Through the feature engineering framework, multimodal data processing is performed on vital signs data, clinical test data, imaging examination data and electronic medical record data to extract multi-scale time series features of vital signs fluctuation trends and clinical indicator change rates. The feature interaction network is constructed to obtain the correlation features of multi-dimensional indicator interaction relationships and comorbidity correlations, including:

[0015] Preprocess vital sign data, clinical test data, and electronic medical record data, use spline interpolation to process unequally spaced sampling data, perform data synchronization alignment based on time windows, and obtain preprocessed standardized medical data;

[0016] Extract multi-scale time series features based on the standardized medical data, calculate short-term time series features including instantaneous change rate and fluctuation amplitude, extract mid-term time series features including trend slope and periodicity features based on time series analysis, and obtain long-term time series features to calculate cumulative change trends using time decay weights;

[0017] A feature interaction network is constructed based on the standardized medical data, the correlation strength between indicators is calculated based on the Pearson correlation coefficient matrix and the dynamic correlation pattern between indicators is captured through the time-lag correlation analysis function, a weighted undirected graph is constructed using the medical indicators as the nodes of the graph and the correlation strength as the weight of the edge, the importance of the node is determined according to the neighbor set of the node and the edge weight to obtain the multi-dimensional indicator interaction relationship, the comorbidity risk score is calculated based on the clinical indicator weight and the risk mapping function, the comorbidity network is constructed based on the co-occurrence frequency and time-series dependency characteristics of the indicators, and the association characteristics of the comorbidity correlation are obtained.

[0018] In an optional embodiment,

[0019] Using the multi-scale time series features, using a recurrent neural network and a dynamic baseline calculation method to identify abnormal conditions includes:

[0020] Perform attention calculation on multi-scale temporal features to obtain weighted feature representation;

[0021] Based on the weighted feature representation, enhanced features are obtained through a hierarchical combination of a bidirectional long short-term memory network and a gated recurrent unit regulated by a dynamic gating coefficient;

[0022] Based on the enhanced features, a state perception baseline is constructed, the medical intervention data is scored for intensity and the intervention impact time window is determined, the event score is calculated based on the rate of change of vital signs and the amount of change of test indicators, and the importance of the event is determined in combination with the degree of coordinated change of multiple indicators, the monitoring interval is divided into a steady-state period and a conversion period according to the intervention impact time window, the steady-state period baseline value is calculated using a first smoothing factor, the conversion period initial value is calculated using a second smoothing factor adjusted according to the importance of the event, the initial baseline value and the target baseline value of the conversion period are weighted by exponential decay to obtain a dynamic baseline value of the conversion period, the deviation tolerance of the steady-state period and the conversion period is determined based on the historical fluctuation standard deviation and the importance of the event to construct a baseline interval, and the deviation of the observed value from the corresponding baseline interval and the state duration are weighted to obtain an abnormal state judgment result;

[0023] Based on the weighted feature representation, a single indicator anomaly score and a combined anomaly assessment score are calculated respectively, the single indicator anomaly scores are weightedly accumulated to obtain an indicator fusion score, the indicator fusion score is adaptively fused with the combined anomaly assessment score, and then weightedly fused with the abnormal state judgment result to obtain an anomaly recognition result.

[0024] In an optional embodiment,

[0025] Based on the weighted feature representation, the enhanced features obtained by the hierarchical combination of bidirectional long short-term memory network and gated recurrent unit regulated by dynamic gating coefficient include:

[0026] Using the weighted feature representation, a hierarchical recurrent neural network is constructed, including a bidirectional long short-term memory network base layer and a gated recurrent unit enhancement layer, the weighted feature representation is mapped into a query matrix, a key matrix and a value matrix, the similarity between the query matrix and the key matrix is ​​calculated to obtain an attention weight, and the attention weight is multiplied by the value matrix to obtain a context feature;

[0027] The bidirectional long short-term memory network base layer performs forward propagation and back propagation on the context features to obtain bidirectional features, and concatenates the bidirectional features to form fusion features;

[0028] The gated recurrent unit enhancement layer concatenates the fused features with the historical hidden states and the current input features to obtain combined features, performs nonlinear transformation on the combined features to obtain dynamic gating coefficients, modulates the dynamic gating coefficients with the outputs of the update gate and the reset gate respectively to obtain modulated update gate outputs and reset gate outputs, resets the historical hidden states based on the modulated reset gate outputs to obtain reset states, transforms the reset states with the current input features to obtain candidate states, selectively updates the candidate states and historical states using the modulated update gate outputs, introduces jump connections to maintain the original feature information, and adaptively fuses the updated states with the original feature information to obtain enhanced features.

[0029] In an optional embodiment,

[0030] Predicting the risk of complications using a graph neural network using the multi-scale time series features and association features includes:

[0031] Dividing the multi-scale time series features into fine-grained, medium-grained and coarse-grained time slices according to the time granularity, encoding the time slices by time position to obtain multi-granularity time series features; obtaining a time attention feature based on the association feature and the multi-granularity time series feature;

[0032] A local spatial graph is constructed based on the interaction relationship of the multidimensional indicators, a global spatial graph is constructed based on the correlation of the comorbidities, a feature vector of the node pairs in the local spatial graph and the global spatial graph is nonlinearly transformed to obtain a spatial attention weight, and the temporal attention feature is fused with the historical feature modulated by the spatial attention weight to obtain a spatiotemporal fusion feature;

[0033] Performing multi-head attention calculation on the spatiotemporal fusion feature to obtain an enhanced feature, and then performing adaptive selection through a gating unit after splicing the enhanced feature with the historical feature to obtain a gated feature;

[0034] Calculating the dynamic contribution of the gated feature to obtain a feature importance score, and weighting the gated feature based on the feature importance score to obtain a final feature representation;

[0035] The time series difference of the prediction results at adjacent moments is calculated through the weighted time decay function to obtain the time series consistency constraint, and the safety interval constraint is constructed based on the relative size relationship of the node risk values ​​to obtain the risk propagation constraint. According to the time series consistency constraint, risk propagation constraint and classification loss, the final feature representation is jointly optimized with multi-task to generate the complication risk prediction result.

[0036] In an optional embodiment,

[0037] The abnormal state recognition results and complication risk prediction results are adaptively fused through the attention mechanism. The attention mechanism builds a personalized weight matrix based on the individual characteristics of the patient and uses a dynamic threshold mechanism to determine the warning level, including:

[0038] Introducing a time attenuation factor into the abnormal state recognition result to weight it to obtain a time-series weighted abnormal state representation, and converting the complication risk prediction result into a risk degree vector;

[0039] A personalized feature vector for the patient is constructed based on the patient's trauma severity score, trauma site and number, degree of organ function damage, post-traumatic time, and intensity of clinical intervention measures; an attention score is calculated using the personalized feature vector for the patient, a personalized weight matrix is ​​generated, and the personalized weight matrix is ​​adaptively fused with the temporal weighted abnormal state representation and the risk degree vector to obtain a warning score;

[0040] An adaptive factor is constructed based on the historical warning accuracy, the basic threshold vector is dynamically updated using the adaptive factor to obtain a current threshold vector, and the warning level is determined according to the warning score and the current threshold vector.

[0041] In an optional embodiment,

[0042] The warning level is integrated with the medical knowledge graph to generate prediction results including risk level, warning reason and intervention suggestions, including:

[0043] Converting the warning level into a warning feature vector, wherein the warning feature vector includes a warning level probability distribution and a warning trigger feature set, wherein the warning trigger feature set is composed of features whose feature importance scores are greater than a preset threshold;

[0044] A bidirectional mapping mechanism is used for integrated verification. In the forward verification, the semantic similarity between the warning feature vector and the knowledge graph path representation is calculated. The paths with semantic similarity greater than the similarity threshold are constructed as a verification path set. In the reverse verification, the expected risk level is calculated based on the path credibility of the verification path set. The initial verification result is obtained by comparing the expected risk level with the warning level predicted by deep learning.

[0045] Construct a multi-level verification framework to deeply verify the initial verification results, calculate the path matching rate of the verification path set with the medical knowledge graph through the medical logic verification layer; calculate the medical verification score of the early warning time series chain in the continuous time window through the time series integrity verification layer, and the medical verification score is the weighted average of the medical logic verification results of adjacent time points in the early warning time series chain; calculate the compatibility coefficient between intervention measures through the clinical safety verification layer, and the compatibility coefficient is a standardized score calculated based on the interaction strength between intervention measures;

[0046] A comprehensive verification score is calculated based on the path matching rate, medical verification score and compatibility coefficient. When the comprehensive verification score is greater than the verification threshold, the verified risk level is adopted, an early warning cause explanation chain is generated based on the verification path set, and intervention suggestions are generated from the intervention measures that have passed the safety verification.

[0047] In an optional embodiment,

[0048] The multi-level verification framework includes:

[0049] The path matching rate between the verification path set and the medical knowledge graph is calculated through the medical logic verification layer. The path matching rate includes a direct matching score and an indirect matching score. The direct matching score is calculated based on the explicit path. The indirect matching score is calculated by building a reasoning chain to calculate the reliability of the implicit association. Different types of medical associations are weighted using learnable dynamic weights.

[0050] The medical verification score of the early warning time series chain in the multi-scale time window is calculated through the time series integrity verification layer. The multi-scale time window includes a short-term window for verifying acute changes, a medium-term window for verifying disease progression, and a long-term window for verifying treatment effects. The time series attention mechanism is introduced to adaptively weight the verification results at different time points.

[0051] The compatibility coefficient between interventions was calculated by the clinical safety validation layer, and the compatibility coefficient included a weighted combination of the basic safety score, the synergy score, and the temporal plausibility score, which was used to calculate a standardized score based on the strength of the interaction between the interventions.

[0052] The present invention processes multi-source heterogeneous data through a feature engineering framework, extracts multi-scale time series features and correlation features, can comprehensively capture the trend of patient status changes and the interaction relationship between indicators, and improves the integrity and accuracy of feature expression.

[0053] The present invention adopts a deep learning prediction framework, combines recurrent neural networks and graph neural networks, realizes feature adaptive fusion through the attention mechanism, and constructs a personalized weight matrix based on the individual characteristics of patients. It can accurately identify abnormal conditions and predict the risk of complications, thereby improving the accuracy and personalization of early warning.

[0054] The present invention constructs a medical knowledge graph and integrates and verifies it with the early warning results through a two-way mapping mechanism, converting the domain expert knowledge into a computable vector representation, achieving the interpretability of the early warning results, and at the same time providing intervention suggestions with clinical guidance significance, thereby improving the practicality and reliability of the early warning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic diagram of a process for constructing a comprehensive intelligent early warning system for trauma patients according to an embodiment of the present invention;

[0056] Figure 2 This is a diagram of the hierarchical recurrent neural network structure of the present invention;

[0057] Figure 3 It is a ROC curve performance comparison chart. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Combine the following Figure 1-Figure 3 The technical solution of the present invention is described in detail with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1 FIG. 1 is a flow chart of a method for constructing a comprehensive intelligent early warning system for trauma patients according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] Through the feature engineering framework, multimodal data processing is performed on vital signs data, clinical test data, imaging examination data and electronic medical record data to extract multi-scale time series features of vital signs fluctuation trends and clinical indicator change rates, and a feature interaction network is constructed to obtain the correlation features of multi-dimensional indicator interaction relationships and comorbidity correlations.

[0062] A deep learning prediction framework is constructed, and the multi-scale time series features are used to identify abnormal states using a recurrent neural network and a dynamic baseline calculation method, and the multi-scale time series features and associated features are used to predict complication risks using a graph neural network; the abnormal state identification results and complication risk prediction results are adaptively fused through an attention mechanism, and the attention mechanism constructs a personalized weight matrix based on individual patient characteristics, and a dynamic threshold mechanism is used to determine the warning level;

[0063] Construct a medical knowledge graph to convert the domain rule knowledge of drug interaction networks and disease complication association networks into low-dimensional vector representations;

[0064] The warning level is integrated and verified with the medical knowledge graph through a bidirectional mapping mechanism to generate prediction results including risk level, warning reasons and intervention suggestions.

[0065] For example, standard preprocessing is performed on vital signs data, clinical test data, and electronic medical record data. Vital signs data include continuous monitoring indicators such as heart rate, blood pressure, and respiratory rate, with a sampling frequency of once every 5 minutes; clinical test data include interval sampling indicators such as blood routine and biochemistry, with an interval of 4-24 hours; electronic medical record data include unstructured text data such as basic patient information and medical records.

[0066] For data of different frequencies, the cubic spline interpolation method is used for alignment, and all indicators are unified into a time series with a 5-minute interval. Multi-scale time series features are extracted through a sliding time window. Short-term features reflect acute changes within 0-24 hours, such as a heart rate increase of more than 20 beats / minute within 30 minutes; medium-term features capture gradual changes in 24-72 hours, such as an increase in the fluctuation range of blood pressure within 48 hours; long-term features represent cumulative effects over 72 hours, such as a weight loss of more than 5% within a week. Practice has shown that compared with single time scale features, multi-scale features increase the abnormality detection rate by about 15%.

[0067] The characteristic interaction network was constructed by time series-based correlation analysis. First, the time series correlation coefficients between indicators were calculated, and the indicator pairs with absolute values ​​of correlation coefficients greater than 0.6 were selected to construct network edges. Then, the importance of nodes was determined based on the co-occurrence frequency of indicators, and the importance score was between 0 and 1. The highly correlated indicator groups were identified through the network community discovery algorithm as the complication correlation characteristics.

[0068] Medical knowledge graph construction: The knowledge graph contains about 2,000 drug nodes and about 1,000 disease nodes. The relationship types include 10 types such as drug interaction, treatment relationship, and concurrent relationship. The nodes are mapped to 64-dimensional vector representations through a random walk algorithm. The strength of drug interactions is divided into 5 levels, and the strength of complication associations is divided into 3 levels.

[0069] In an optional embodiment,

[0070] Through the feature engineering framework, multimodal data processing is performed on vital signs data, clinical test data, imaging examination data and electronic medical record data to extract multi-scale time series features of vital signs fluctuation trends and clinical indicator change rates. The feature interaction network is constructed to obtain the correlation features of multi-dimensional indicator interaction relationships and comorbidity correlations, including:

[0071] Preprocess vital sign data, clinical test data, and electronic medical record data, use spline interpolation to process unequally spaced sampling data, perform data synchronization alignment based on time windows, and obtain preprocessed standardized medical data;

[0072] Extract multi-scale time series features based on the standardized medical data, calculate short-term time series features including instantaneous change rate and fluctuation amplitude, extract mid-term time series features including trend slope and periodicity features based on time series analysis, and obtain long-term time series features to calculate cumulative change trends using time decay weights;

[0073] A feature interaction network is constructed based on the standardized medical data, the correlation strength between indicators is calculated based on the Pearson correlation coefficient matrix and the dynamic correlation pattern between indicators is captured through the time-lag correlation analysis function, a weighted undirected graph is constructed using the medical indicators as the nodes of the graph and the correlation strength as the weight of the edge, the importance of the node is determined according to the neighbor set of the node and the edge weight to obtain the multi-dimensional indicator interaction relationship, the comorbidity risk score is calculated based on the clinical indicator weight and the risk mapping function, the comorbidity network is constructed based on the co-occurrence frequency and time-series dependency characteristics of the indicators, and the association characteristics of the comorbidity correlation are obtained.

[0074] Exemplarily, preprocessing is performed on vital signs data, clinical test data, imaging examination data and electronic medical record data. Vital signs data include indicators such as blood pressure, heart rate, and body temperature. Clinical test data include blood routine, biochemical and other test results. Imaging examination data include structured description text and diagnostic conclusions of radiological images (X-rays, CT, MRI, etc.). Electronic medical record data include diagnosis, medication and other information. For data collected at different frequencies, the cubic spline interpolation method is used to deal with missing values ​​and unequal interval sampling problems. Taking blood pressure data as an example, the original sampling interval is 4 hours, which is converted into one data point per hour through interpolation. For multi-source heterogeneous data, alignment is performed based on a 24-hour time window to obtain a standardized medical data matrix.

[0075] Extract multi-scale time series features from standardized medical data. Short-term feature calculation includes: calculating the difference between adjacent time points to obtain the instantaneous rate of change, and calculating the difference between the maximum and minimum values ​​to obtain the fluctuation range. Taking heart rate as an example, calculate the heart rate change per hour, and calculate the maximum fluctuation range within 24 hours. Medium-term feature extraction includes: fitting the trend line based on the sliding time window to obtain the slope, and obtaining periodic features through autocorrelation analysis. Long-term features introduce time decay weights, perform weighted averaging on historical data, and obtain cumulative change trends.

[0076] Construct a feature interaction network. First, calculate the correlation strength matrix between medical indicators. Take blood pressure and heart rate as examples to calculate their Pearson correlation coefficients. Capture dynamic associations through time-lag correlation analysis, such as changes in blood pressure leading changes in heart rate by 2 hours. Use medical indicators as network nodes and correlation strength as edge weights to construct a weighted undirected graph. Calculate node importance based on node degree centrality and edge weights to identify key medical indicators. Analyze comorbidity correlation. Calculate comorbidity risk scores based on the degree of abnormality of clinical indicators and weights set by expert experience. Construct a comorbidity network based on the co-occurrence frequency and time-series dependency characteristics of statistical indicators, capture the evolutionary relationship between diseases, and obtain the association characteristics of comorbidity correlation. For example, hypertensive patients are prone to coronary heart disease, and the two disease indicators have a strong time-series correlation.

[0077] Through multimodal data preprocessing and feature engineering, the present invention realizes standardized processing and time series alignment of medical data, laying the foundation for subsequent analysis; based on multi-scale time series feature extraction and feature interaction network construction, it comprehensively depicts the dynamic change law and interaction relationship of medical indicators, and improves the feature expression ability; through comorbidity correlation analysis, it reveals the law of disease development and evolution, provides a basis for clinical diagnosis and treatment decisions and disease prediction and early warning, and has important clinical application value.

[0078] In an optional embodiment,

[0079] Using the multi-scale time series features, using a recurrent neural network and a dynamic baseline calculation method to identify abnormal conditions includes:

[0080] Perform attention calculation on multi-scale temporal features to obtain weighted feature representation;

[0081] Based on the weighted feature representation, enhanced features are obtained through a hierarchical combination of a bidirectional long short-term memory network and a gated recurrent unit regulated by a dynamic gating coefficient;

[0082] Based on the enhanced features, a state perception baseline is constructed, the medical intervention data is scored for intensity and the intervention impact time window is determined, the event score is calculated based on the rate of change of vital signs and the amount of change of test indicators, and the importance of the event is determined in combination with the degree of coordinated change of multiple indicators, the monitoring interval is divided into a steady-state period and a conversion period according to the intervention impact time window, the steady-state period baseline value is calculated using a first smoothing factor, the conversion period initial value is calculated using a second smoothing factor adjusted according to the importance of the event, the initial baseline value and the target baseline value of the conversion period are weighted by exponential decay to obtain a dynamic baseline value of the conversion period, the deviation tolerance of the steady-state period and the conversion period is determined based on the historical fluctuation standard deviation and the importance of the event to construct a baseline interval, and the deviation of the observed value from the corresponding baseline interval and the state duration are weighted to obtain an abnormal state judgment result;

[0083] Based on the weighted feature representation, a single indicator anomaly score and a combined anomaly assessment score are calculated respectively, the single indicator anomaly scores are weightedly accumulated to obtain an indicator fusion score, the indicator fusion score is adaptively fused with the combined anomaly assessment score, and then weightedly fused with the abnormal state judgment result to obtain an anomaly recognition result.

[0084] Exemplarily, dynamic attention calculation is performed on multi-scale time series features, a query matrix is ​​constructed based on the feature vector at the current moment, a key-value matrix is ​​constructed based on the historical feature sequence, the query matrix is ​​multiplied by the key-value matrix and softmax normalized to obtain an attention score, and the attention score is multiplied by the key-value matrix to obtain a weighted feature representation.

[0085] Based on the weighted feature representation, a hierarchical combination of bidirectional long short-term memory network and gated recurrent unit is constructed. The bidirectional long short-term memory network contains two hidden layers, forward and backward, which capture the forward and backward dependencies of temporal features respectively. The gated recurrent unit regulates the information flow through dynamic gating coefficients, which are determined by the current input and historical state. The two networks are combined in series to output enhanced feature representation.

[0086] Construct a state-aware baseline. Score the medical intervention data. For example, the larger the drug dose and the more complex the operation, the higher the score. Determine the impact time window based on the intervention intensity. For example, the impact period of high-dose medication is 4 hours. Calculate the change rate and amount of each indicator to obtain the event score. For example, a blood pressure change of more than 20% will result in a higher score. Determine the importance of the event by combining the degree of coordinated change of multiple indicators.

[0087] Based on the intervention impact time window, the monitoring interval was divided into a steady-state period and a transition period. The baseline value was calculated using a smoothing factor of 0.8 for the steady-state period. For the transition period, the smoothing factor was adjusted according to the importance of the event. For example, when the importance was 0.9, the smoothing factor was 0.5. The starting baseline value and the target baseline value of the transition period were weighted by exponential decay, with an attenuation coefficient of 0.85, to obtain the dynamic baseline of the transition period.

[0088] The tolerance for the steady-state period deviation is determined based on the historical fluctuation standard deviation, such as plus or minus 2 times the standard deviation. The tolerance for the transition period is determined based on the importance of the event, such as plus or minus 3 times the standard deviation when the importance is 0.9. The deviation between the observed value and the baseline interval is weighted in combination with the state duration. The longer the duration, the greater the weight, and the abnormal state judgment result is obtained.

[0089] Finally, anomaly fusion recognition is performed. Based on weighted features, the single indicator anomaly score and the combined anomaly assessment score are calculated respectively. The single indicator anomaly score is weighted and accumulated, and the weight is determined by the indicator importance. The indicator fusion score and the combined anomaly assessment score are adaptively fused, and the fusion coefficient is determined by the historical accuracy. Then, the final recognition result is obtained by weighted fusion with the abnormal state judgment result.

[0090] This application introduces a multi-scale feature extraction mechanism, while paying attention to the feature changes in the short-term, medium-term, and long-term time scales. By comprehensively analyzing the indicator change patterns at different time scales, the detection rate of abnormal states has been significantly improved. In terms of baseline construction, existing methods mostly use fixed smoothing factors, which are slow to respond to medical interventions. This application proposes an adaptive smoothing mechanism driven by event importance, so that the baseline can respond quickly to medical interventions. In response to the problem that existing technologies often evaluate the abnormal state of each indicator separately, this application constructs a complete abnormal fusion architecture that includes single indicator evaluation, combined indicator evaluation, and state judgment.

[0091] In an optional embodiment,

[0092] Based on the weighted feature representation, the enhanced features obtained by the hierarchical combination of bidirectional long short-term memory network and gated recurrent unit regulated by dynamic gating coefficient include:

[0093] Using the weighted feature representation, a hierarchical recurrent neural network is constructed, including a bidirectional long short-term memory network base layer and a gated recurrent unit enhancement layer, the weighted feature representation is mapped into a query matrix, a key matrix and a value matrix, the similarity between the query matrix and the key matrix is ​​calculated to obtain an attention weight, and the attention weight is multiplied by the value matrix to obtain a context feature;

[0094] The bidirectional long short-term memory network base layer performs forward propagation and back propagation on the context features to obtain bidirectional features, and concatenates the bidirectional features to form fusion features;

[0095] The gated recurrent unit enhancement layer concatenates the fused features with the historical hidden states and the current input features to obtain combined features, performs nonlinear transformation on the combined features to obtain dynamic gating coefficients, modulates the dynamic gating coefficients with the outputs of the update gate and the reset gate respectively to obtain modulated update gate outputs and reset gate outputs, resets the historical hidden states based on the modulated reset gate outputs to obtain reset states, transforms the reset states with the current input features to obtain candidate states, selectively updates the candidate states and historical states using the modulated update gate outputs, introduces jump connections to maintain the original feature information, and adaptively fuses the updated states with the original feature information to obtain enhanced features.

[0096] Exemplarily, a hierarchical recurrent neural network is constructed based on weighted feature representation. First, the weighted feature representation is mapped into a query matrix, a key matrix, and a value matrix through a linear transformation. The query matrix and the key matrix are calculated to obtain the attention weight by similarity, which is specifically implemented through a dot product operation, and the dot product result is normalized to obtain the weight coefficient. The attention weight is multiplied by the value matrix to obtain the context feature, realizing the dynamic weighted combination of features. For example, if the input feature dimension is 4096, three 4096-dimensional matrices are obtained through linear mapping, and the calculated attention weight is a 4096-dimensional vector.

[0097] The basic layer of the bidirectional long short-term memory network includes two directions: forward and reverse. In forward propagation, context features are processed in chronological order, and the hidden state at each moment depends on the current input and the state at the previous moment. Backward propagation processes features in the opposite order. The hidden state dimensions in both directions are 512, which are concatenated to obtain a 1024-dimensional fusion feature. The information flow is controlled by the forget gate, input gate, and output gate to achieve long-term dependency modeling.

[0098] The gated recurrent unit enhancement layer first concatenates the 1024-dimensional fusion features with the 512-dimensional historical hidden state and the 512-dimensional current input features to obtain a 2048-dimensional combined feature. The combined feature is transformed through a fully connected layer and an activation function to obtain a dynamic gating coefficient with a dimension of 512. This coefficient is multiplied by the sigmoid output of the update gate and the reset gate to obtain a modulated gating signal. Based on the reset gate signal, the historical hidden state is selectively forgotten to obtain a reset state, which is transformed with the current input feature to obtain a candidate state. The modulated update gate signal is used to selectively update the candidate state and the historical state. At the same time, a jump connection is introduced to maintain the original feature information, and the updated 512-dimensional state is adaptively fused with the 1024-dimensional original feature through a gating mechanism to obtain a 1024-dimensional enhanced feature.

[0099] This application designs an innovative combination of a bidirectional long short-term memory network base layer and a gated recurrent unit enhancement layer. The base layer achieves comprehensive modeling of long-term dependencies through forward and reverse bidirectional processing. The enhancement layer introduces dynamic gating coefficients so that the gating mechanism can be adaptively adjusted according to data characteristics.

[0100] This application also innovatively introduces a skip connection mechanism to maintain the original feature information and balance the original features and deep features through an adaptive fusion strategy.

[0101] This application effectively solves the key technical problems in medical time series data processing through an innovative hierarchical feature enhancement framework, providing more reliable technical support for clinical early warning decision-making. This solution not only improves the accuracy of early warning, but also has good real-time and interpretability, and has important clinical application value.

[0102] Figure 3 This is a ROC curve performance comparison chart. Figure 3 The ROC (Receiver Operating Characteristic) curve comparison results of the three methods are shown: Traditional fixed threshold method (dashed line): AUC value is 0.72, the curve curvature is small, indicating limited detection performance; Simple moving average method (dash-dotted line): AUC value is 0.83, which is an improvement over the traditional method; This solution (thick solid line): AUC value reaches 0.92, and the curve is significantly better than the other two methods. The horizontal axis of the ROC curve is the false positive rate, and the vertical axis is the true positive rate. The closer the curve is to the upper left corner, the better the detection performance. The AUC value represents the area under the ROC curve. The closer its value is to 1, the higher the detection accuracy of the method. The AUC value of this scheme (0.92) is 27.8% higher than that of the traditional fixed threshold method (0.72), and the AUC value of this scheme is 10.8% higher than that of the simple moving average method (0.83). The ROC curve is closer to the ideal point in the upper left corner, indicating that the false positive rate can be significantly reduced while maintaining a high true positive rate. Compared with the traditional fixed threshold method, the false positive rate is greatly reduced at the same true positive rate, reducing the time for medical staff to deal with false alarms, and improving the reliability and practicality of the early warning system. The figure intuitively shows the significant advantages of this scheme in anomaly detection accuracy, verifying the practical value and technical innovation of this scheme in clinical applications. By providing more accurate anomaly detection results, this scheme can effectively improve the quality of monitoring of critically ill patients.

[0103] In an optional embodiment, predicting the risk of complications using a graph neural network using the multi-scale time series features and association features includes:

[0104] Dividing the multi-scale time series features into fine-grained, medium-grained and coarse-grained time slices according to the time granularity, encoding the time slices by time position to obtain multi-granularity time series features; obtaining a time attention feature based on the association feature and the multi-granularity time series feature;

[0105] A local spatial graph is constructed based on the interaction relationship of the multidimensional indicators, a global spatial graph is constructed based on the correlation of the comorbidities, a feature vector of the node pairs in the local spatial graph and the global spatial graph is nonlinearly transformed to obtain a spatial attention weight, and the temporal attention feature is fused with the historical feature modulated by the spatial attention weight to obtain a spatiotemporal fusion feature;

[0106] Performing multi-head attention calculation on the spatiotemporal fusion feature to obtain an enhanced feature, and then performing adaptive selection through a gating unit after splicing the enhanced feature with the historical feature to obtain a gated feature;

[0107] Calculating the dynamic contribution of the gated feature to obtain a feature importance score, and weighting the gated feature based on the feature importance score to obtain a final feature representation;

[0108] The time series difference of the prediction results at adjacent moments is calculated through the weighted time decay function to obtain the time series consistency constraint, and the safety interval constraint is constructed based on the relative size relationship of the node risk values ​​to obtain the risk propagation constraint. According to the time series consistency constraint, risk propagation constraint and classification loss, the final feature representation is jointly optimized with multi-task to generate the complication risk prediction result.

[0109] Exemplarily, a complication risk prediction method based on multi-scale time series features and correlation features first preprocesses the electronic medical record data to extract multi-dimensional information such as the patient's physiological indicators, medication records, and test examinations.

[0110] The time series features are divided into different scales according to the time granularity. For example, the fine granularity is the hourly indicator changes, such as blood pressure, heart rate, etc.; the medium granularity is the daily indicator changes, such as medication dosage, test values, etc.; the coarse granularity is the weekly or monthly indicator change trend. Position coding information is added to time slices of different granularities, and the coding method uses sine and cosine functions, so that the model can capture the time series position information.

[0111] Construct a temporal attention mechanism and calculate the importance weights of features at different time points. For each time point, calculate the similarity between its feature vector and the features of all historical time points to obtain an attention score. Perform a weighted summation of historical features based on the score to obtain a temporal attention feature containing long-term dependencies.

[0112] Based on the correlation between multidimensional indicators, a local spatial graph is constructed. The nodes in the graph represent different indicators, and the edges represent the degree of correlation between the indicators. At the same time, a global spatial graph is constructed based on the co-occurrence relationship between complications to capture the law of disease development. The feature vectors of the node pairs in the graph are transformed nonlinearly to obtain the spatial attention weights, which are used to modulate the historical features. The temporal attention features are fused with the modulated historical features to obtain the spatiotemporal fusion features.

[0113] Multi-head attention calculation is performed on the spatiotemporal fusion features, and different attention heads focus on different feature subspaces to enhance the feature expression ability. The enhanced features are spliced ​​with the original historical features, and feature selection is performed through the gating unit to retain important information. Specifically, the gating unit contains an update gate and a reset gate, which respectively control the retention and forgetting of historical information.

[0114] Calculate the contribution of the gated features to the prediction results, and use a gradient-based method to obtain the feature importance score. Weight the features according to the scores to highlight the impact of key features. At the same time, introduce temporal consistency constraints and risk propagation constraints for multi-task optimization. In the specific implementation, an alternating optimization strategy based on the Adam optimizer can be used to achieve multi-task joint training by dynamically adjusting the weight coefficients of different tasks. Temporal consistency constraints ensure smooth changes in the prediction results at adjacent moments, and risk propagation constraints ensure the impact of high-risk nodes on neighboring nodes.

[0115] The present invention adopts a fusion modeling method of multi-scale temporal features and multi-level spatial features, which can comprehensively capture the evolution law of patient status and improve prediction accuracy; introduces temporal attention and spatial attention mechanisms, adaptively learns the importance of features, highlights the impact of key information, and enhances model interpretability; through multi-task optimization of temporal consistency constraints and risk propagation constraints, the temporal smoothness and spatial consistency of the prediction results are guaranteed, and the prediction stability is improved.

[0116] In an optional embodiment,

[0117] The abnormal state recognition results and complication risk prediction results are adaptively fused through the attention mechanism. The attention mechanism builds a personalized weight matrix based on the individual characteristics of the patient and uses a dynamic threshold mechanism to determine the warning level, including:

[0118] Introducing a time attenuation factor into the abnormal state recognition result to weight it to obtain a time-series weighted abnormal state representation, and converting the complication risk prediction result into a risk degree vector;

[0119] A personalized feature vector for the patient is constructed based on the patient's trauma severity score, trauma site and number, degree of organ function damage, post-traumatic time, and intensity of clinical intervention measures; an attention score is calculated using the personalized feature vector for the patient, a personalized weight matrix is ​​generated, and the personalized weight matrix is ​​adaptively fused with the temporal weighted abnormal state representation and the risk degree vector to obtain a warning score;

[0120] An adaptive factor is constructed based on the historical warning accuracy, the basic threshold vector is dynamically updated using the adaptive factor to obtain a current threshold vector, and the warning level is determined according to the warning score and the current threshold vector.

[0121] Exemplarily, the specific technical implementation process of realizing adaptive feature fusion of abnormal state recognition and complication risk prediction results through attention mechanism is as follows:

[0122] The time decay factor is introduced into the abnormal state recognition result for weighting. For the abnormal state recognition result at each time point, the time decay factor is calculated using an exponential decay function according to the interval from the current time. The larger the time interval, the smaller the decay factor. Multiply the abnormal state recognition result with the corresponding time decay factor to obtain the abnormal state representation after time series weighting. For example, the abnormal state recognition results of a patient at three consecutive time points are severe, moderate, and mild, respectively, and the corresponding time decay factors are 0.3, 0.6, and 0.9, then the weighted abnormal state representation is 0.9, 0.6, and 0.3.

[0123] At the same time, the complication risk prediction results are converted into risk degree vectors. According to the predicted probability of each type of complication, it is mapped to the interval from zero to one to represent the risk degree. For example, the predicted probabilities of lung infection, deep vein thrombosis, and stress ulcer are 80%, 60%, and 40%, respectively, and the corresponding risk degree vectors are 0.8, 0.6, and 0.4.

[0124] Then, a personalized feature vector for the patient was constructed. Based on the patient's physiological criticality score (such as APACHE II score of 25 points, normalized to 0.5), number of injured parts (2 injuries, normalized to 0.4), degree of damage to major organ functions (moderate injury, normalized to 0.6), time after trauma (48 hours, normalized to 0.3), and intensity of clinical intervention measures (accepting invasive mechanical ventilation, corresponding intensity is 0.8), these normalized feature values ​​were concatenated to form a feature vector [0.5, 0.4, 0.6, 0.3, 0.8]. The intensity of clinical intervention measures was divided into 5 levels: ECMO or CRRT treatment was 1.0, invasive mechanical ventilation was 0.8, non-invasive ventilator or vasoactive drugs was 0.6, high-flow oxygen inhalation was 0.4, and conventional oxygen inhalation or infusion was 0.2.

[0125] The attention score is calculated using the patient's personalized feature vector. The feature vector is mapped to the interval between zero and one through nonlinear transformation to represent the attention score. A personalized weight matrix is ​​generated based on the attention score to perform weighted fusion of the abnormal state representation and the risk degree vector. The weight matrix is ​​multiplied with the time-series weighted abnormal state representation and the risk degree vector to obtain the warning score.

[0126] Finally, an adaptive factor is constructed based on the historical warning accuracy. The accuracy of the warning results in the past period of time is counted. When the accuracy is high, the adaptive factor is increased, and vice versa. The adaptive factor is multiplied by the preset basic threshold vector to obtain the current threshold vector. The warning score is compared with the current threshold vector to determine the final warning level.

[0127] The present invention introduces a time decay factor to weight the abnormal state recognition results, fully considering the importance of time series information, making the early warning results more accurate and reliable. By dynamically adjusting the size of the time decay factor, the influence of historical information can be flexibly controlled; an attention mechanism is constructed based on the personalized characteristics of patients to achieve adaptive fusion of abnormal states and complication risks. The attention score can dynamically adjust the importance weights of different features according to the specific circumstances of the patient, thereby improving the personalization of the early warning results. The dynamic threshold mechanism is used to determine the early warning level, avoiding the limitations brought by the fixed threshold. By introducing an adaptive factor to dynamically adjust the threshold, the early warning system can be continuously optimized and improved according to the actual effect, thereby improving the accuracy and reliability of the early warning.

[0128] In an optional embodiment,

[0129] The warning level is integrated with the medical knowledge graph to generate prediction results including risk level, warning reason and intervention suggestions, including:

[0130] Converting the warning level into a warning feature vector, wherein the warning feature vector includes a warning level probability distribution and a warning trigger feature set, wherein the warning trigger feature set is composed of features whose feature importance scores are greater than a preset threshold;

[0131] A bidirectional mapping mechanism is used for integrated verification. In the forward verification, the semantic similarity between the warning feature vector and the knowledge graph path representation is calculated. The paths with semantic similarity greater than the similarity threshold are constructed as a verification path set. In the reverse verification, the expected risk level is calculated based on the path credibility of the verification path set. The initial verification result is obtained by comparing the expected risk level with the warning level predicted by deep learning.

[0132] Construct a multi-level verification framework to deeply verify the initial verification results, calculate the path matching rate of the verification path set with the medical knowledge graph through the medical logic verification layer; calculate the medical verification score of the early warning time series chain in the continuous time window through the time series integrity verification layer, and the medical verification score is the weighted average of the medical logic verification results of adjacent time points in the early warning time series chain; calculate the compatibility coefficient between intervention measures through the clinical safety verification layer, and the compatibility coefficient is a standardized score calculated based on the interaction strength between intervention measures;

[0133] A comprehensive verification score is calculated based on the path matching rate, medical verification score and compatibility coefficient. When the comprehensive verification score is greater than the verification threshold, the verified risk level is adopted, an early warning cause explanation chain is generated based on the verification path set, and intervention suggestions are generated from the intervention measures that have passed the safety verification.

[0134] For example, the feature vector of the warning level is transformed. Taking the heart failure warning of a patient as an example, the warning level is "high risk", and the corresponding warning probability distribution is low risk 0.15, medium risk 0.25, and high risk 0.60. Through feature importance analysis, the importance scores of vital signs such as blood pressure, heart rate, and respiration are 0.82, 0.75, and 0.68 respectively, which are all higher than the preset threshold of 0.5, forming a warning trigger feature set.

[0135] Then, bidirectional mapping verification is performed. In the forward verification, the semantic similarity between the warning feature vector and the paths such as "heart failure-symptoms-complications" in the knowledge graph is calculated, and the paths with similarity greater than 0.8 are selected into the verification path set. In the reverse verification, the expected risk level calculated based on the path credibility is "high risk", which is consistent with the deep learning prediction result, and the preliminary verification is passed.

[0136] Then, a multi-level verification framework was constructed. In the medical logic verification layer, the matching rate between the verification path and the knowledge graph was calculated to be 0.85; in the temporal integrity verification layer, the warning time chain of the past 7 days was analyzed, and the medical verification score was calculated to be 0.78; in the clinical safety verification layer, the compatibility coefficient of intervention measures such as diuretics and vasodilators was evaluated to be 0.92.

[0137] Finally, the comprehensive verification score calculated based on the above three verification indicators was 0.85, which was higher than the verification threshold of 0.8. Therefore, the verified high-risk warning level was used to generate a warning cause explanation chain including "increased blood pressure-heart failure-exacerbated heart failure", and a diuretic and vasodilator treatment plan that passed safety verification was recommended.

[0138] The present invention can effectively improve the interpretability and credibility of warning results and reduce the false alarm rate through a bidirectional mapping verification mechanism between warning feature vectors and knowledge graphs.

[0139] In an optional implementation, the multi-level verification framework includes:

[0140] The path matching rate between the verification path set and the medical knowledge graph is calculated through the medical logic verification layer. The path matching rate includes a direct matching score and an indirect matching score. The direct matching score is calculated based on the explicit path. The indirect matching score is calculated by building a reasoning chain to calculate the reliability of the implicit association. Different types of medical associations are weighted using learnable dynamic weights.

[0141] The medical verification score of the early warning time series chain in the multi-scale time window is calculated through the time series integrity verification layer. The multi-scale time window includes a short-term window for verifying acute changes, a medium-term window for verifying disease progression, and a long-term window for verifying treatment effects. The time series attention mechanism is introduced to adaptively weight the verification results at different time points.

[0142] The compatibility coefficient between interventions was calculated by the clinical safety validation layer, and the compatibility coefficient included a weighted combination of the basic safety score, the synergy score, and the temporal plausibility score, which was used to calculate a standardized score based on the strength of the interaction between the interventions.

[0143] Exemplarily, the multi-level verification framework first calculates the path matching rate through the medical logic verification layer. For direct matching scores, the system extracts explicit paths from the medical knowledge graph, including disease-symptom, disease-examination, disease-treatment and other association paths. For example, for the association "hypertension-headache", the system searches for a direct association path in the knowledge graph, and calculates the matching score if it exists. For indirect matching scores, the system constructs an inference chain to calculate implicit associations, such as multi-hop paths such as "hypertension-vasoconstriction-cerebral vasospasm-headache". The system sets dynamic weights for different types of medical associations, such as symptom association weight 0.3, examination association weight 0.4, and treatment association weight 0.3. The weights can be dynamically adjusted according to the actual application scenario.

[0144] The temporal integrity verification layer calculates the verification score based on multi-scale time windows. The short-term window is set to 24 hours to verify acute changes in indicators such as blood pressure and blood sugar; the medium-term window is set to 7 days to verify the progression of the disease; the long-term window is set to 30 days to verify the treatment effect. The system introduces a temporal attention mechanism to weight the verification results at different time points. For example, for blood pressure monitoring of patients with hypertension, if the recent measured values ​​fluctuate greatly, the weight of the corresponding time point will be increased accordingly.

[0145] The clinical safety verification layer calculates the compatibility coefficient between interventions. The basic safety score considers factors such as drug incompatibility and dosage range; the synergistic effect score evaluates the combined effect of multiple interventions; and the temporal rationality score verifies the rationality of the dosing sequence. For example, for patients who use antihypertensive drugs and diuretics at the same time, the system calculates the intensity of the interaction between the drugs and generates a standardized score. If the two drugs have a synergistic effect, the synergistic effect score is higher; if the dosing interval is unreasonable, the temporal rationality score is lower.

[0146] The present invention realizes the deep integration of early warning results and medical knowledge through a bidirectional mapping mechanism, which not only verifies the rationality of the early warning results, but also provides an explainable verification path; constructs a multi-level verification framework including medical logic, temporal integrity and clinical safety to realize all-round verification of early warning results; finally, improves the adaptability of the verification process through dynamic weights and attention mechanisms, so that the verification results are more in line with actual clinical needs.

[0147] The present invention expands the traditional one-way verification into a two-way mapping verification to improve the reliability of verification; introduces a multi-level verification framework to perform comprehensive verification from three dimensions: medical logic, temporal integrity, and clinical safety; and adopts dynamic weights and attention mechanisms to enhance the adaptability and flexibility of the verification process. Through the above improvements, the present invention significantly improves the reliability and interpretability of the early warning results. Experimental results show that compared with the prior art, the early warning accuracy of this method is increased by 15%, and the false alarm rate is reduced by 20%. At the same time, it can provide more reliable verification basis and intervention suggestions for clinical decision-making. These improvements are of great significance to improving the practicality and credibility of the medical early warning system.

[0148] According to a second aspect of the embodiments of the present invention,

[0149] An electronic device is provided, comprising:

[0150] processor;

[0151] a memory for storing processor-executable instructions;

[0152] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0153] According to a third aspect of the embodiments of the present invention,

[0154] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0155] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a comprehensive intelligent early warning system for trauma patients, characterized in that: include: Through the feature engineering framework, multimodal data processing is performed on vital signs data, clinical test data, imaging examination data and electronic medical record data to extract multi-scale time series features of vital signs fluctuation trends and clinical indicator change rates, and a feature interaction network is constructed to obtain the correlation features of multi-dimensional indicator interaction relationships and comorbidity correlations. A deep learning prediction framework is constructed, and the multi-scale time series features are used to identify abnormal states using a recurrent neural network and a dynamic baseline calculation method. The multi-scale time series features and the association features are used to predict the risk of complications using a graph neural network, including: dividing the multi-scale time series features into fine-grained, medium-grained, and coarse-grained time slices according to the time granularity, and encoding the time slices by time position to obtain multi-granular time series features; obtaining time attention features based on the association features and the multi-granular time series features; constructing a local spatial graph based on the interactive relationship of the multi-dimensional indicators, and constructing a global spatial graph based on the comorbidity correlation, performing a nonlinear transformation on the feature vectors of the node pairs in the local spatial graph and the global spatial graph to obtain spatial attention weights, and fusing the time attention features with the historical features modulated by the spatial attention weights to obtain spatiotemporal fusion features; performing multi-headed Attention calculation obtains enhanced features, and the enhanced features are spliced ​​with historical features and adaptively selected through the gating unit to obtain gated features; the dynamic contribution of the gated features is calculated to obtain the feature importance score, and the gated features are weighted based on the feature importance score to obtain the final feature representation; the time series difference of the prediction results at adjacent moments is calculated through the weighted time decay function to obtain the time series consistency constraint, and the safety interval constraint is constructed based on the relative size relationship of the node risk value to obtain the risk propagation constraint. According to the time series consistency constraint and the risk propagation constraint and the classification loss, the final feature representation is multi-task jointly optimized to generate the complication risk prediction result; the abnormal state recognition result and the complication risk prediction result are adaptively fused through the attention mechanism, and the attention mechanism constructs a personalized weight matrix based on the individual characteristics of the patient, and the dynamic threshold mechanism is used to determine the warning level; Construct a medical knowledge graph to convert the domain rule knowledge of drug interaction networks and disease complication association networks into low-dimensional vector representations; The warning level is integrated and verified with the medical knowledge graph through a bidirectional mapping mechanism to generate prediction results including risk level, warning reasons and intervention suggestions.

2. The method according to claim 1, characterized in that Through the feature engineering framework, multimodal data processing is performed on vital signs data, clinical test data, imaging examination data and electronic medical record data to extract multi-scale time series features of vital signs fluctuation trends and clinical indicator change rates. The feature interaction network is constructed to obtain the correlation features of multi-dimensional indicator interaction relationships and comorbidity correlations, including: Preprocess vital sign data, clinical test data, and electronic medical record data, use spline interpolation to process unequally spaced sampling data, perform data synchronization alignment based on time windows, and obtain preprocessed standardized medical data; Extract multi-scale time series features based on the standardized medical data, calculate short-term time series features including instantaneous change rate and fluctuation amplitude, extract mid-term time series features including trend slope and periodicity features based on time series analysis, and obtain long-term time series features to calculate cumulative change trends using time decay weights; A feature interaction network is constructed based on the standardized medical data, the correlation strength between indicators is calculated based on the Pearson correlation coefficient matrix and the dynamic correlation pattern between indicators is captured through the time-lag correlation analysis function, a weighted undirected graph is constructed using the medical indicators as the nodes of the graph and the correlation strength as the weight of the edge, the importance of the node is determined according to the neighbor set of the node and the edge weight to obtain the multi-dimensional indicator interaction relationship, the comorbidity risk score is calculated based on the clinical indicator weight and the risk mapping function, the comorbidity network is constructed based on the co-occurrence frequency and time-series dependency characteristics of the indicators, and the association characteristics of the comorbidity correlation are obtained.

3. The method according to claim 1, characterized in that Using the multi-scale time series features, using a recurrent neural network and a dynamic baseline calculation method to identify abnormal conditions includes: Perform attention calculation on multi-scale temporal features to obtain weighted feature representation; Based on the weighted feature representation, enhanced features are obtained through a hierarchical combination of a bidirectional long short-term memory network and a gated recurrent unit regulated by a dynamic gating coefficient; Based on the enhanced features, a state perception baseline is constructed, the medical intervention data is scored for intensity and the intervention impact time window is determined, the event score is calculated based on the rate of change of vital signs and the amount of change of test indicators, and the importance of the event is determined in combination with the degree of coordinated change of multiple indicators, the monitoring interval is divided into a steady-state period and a conversion period according to the intervention impact time window, the steady-state period baseline value is calculated using a first smoothing factor, the conversion period initial value is calculated using a second smoothing factor adjusted according to the importance of the event, the initial baseline value and the target baseline value of the conversion period are weighted by exponential decay to obtain a dynamic baseline value of the conversion period, the deviation tolerance of the steady-state period and the conversion period is determined based on the historical fluctuation standard deviation and the importance of the event to construct a baseline interval, and the deviation of the observed value from the corresponding baseline interval and the state duration are weighted to obtain an abnormal state judgment result; Based on the weighted feature representation, a single indicator anomaly score and a combined anomaly assessment score are calculated respectively, the single indicator anomaly scores are weightedly accumulated to obtain an indicator fusion score, the indicator fusion score is adaptively fused with the combined anomaly assessment score, and then weightedly fused with the abnormal state judgment result to obtain an anomaly recognition result.

4. The method according to claim 3, characterized in that: Based on the weighted feature representation, the enhanced features obtained by the hierarchical combination of bidirectional long short-term memory network and gated recurrent unit regulated by dynamic gating coefficient include: Using the weighted feature representation, a hierarchical recurrent neural network is constructed, including a bidirectional long short-term memory network base layer and a gated recurrent unit enhancement layer, the weighted feature representation is mapped into a query matrix, a key matrix and a value matrix, the similarity between the query matrix and the key matrix is ​​calculated to obtain an attention weight, and the attention weight is multiplied by the value matrix to obtain a context feature; The bidirectional long short-term memory network base layer performs forward propagation and back propagation on the context features to obtain bidirectional features, and concatenates the bidirectional features to form fusion features; The gated recurrent unit enhancement layer concatenates the fused features with the historical hidden states and the current input features to obtain combined features, performs nonlinear transformation on the combined features to obtain dynamic gating coefficients, modulates the dynamic gating coefficients with the outputs of the update gate and the reset gate respectively to obtain modulated update gate outputs and reset gate outputs, resets the historical hidden states based on the modulated reset gate outputs to obtain reset states, transforms the reset states with the current input features to obtain candidate states, selectively updates the candidate states and historical states using the modulated update gate outputs, introduces jump connections to maintain the original feature information, and adaptively fuses the updated states with the original feature information to obtain enhanced features.

5. The method according to claim 1, characterized in that: The abnormal state recognition results and complication risk prediction results are adaptively fused through the attention mechanism. The attention mechanism builds a personalized weight matrix based on the individual characteristics of the patient and uses a dynamic threshold mechanism to determine the warning level, including: Introducing a time attenuation factor into the abnormal state recognition result to weight it to obtain a time-series weighted abnormal state representation, and converting the complication risk prediction result into a risk degree vector; A personalized feature vector for the patient is constructed based on the patient's trauma severity score, trauma site and number, degree of organ function damage, post-traumatic time, and intensity of clinical intervention measures; an attention score is calculated using the personalized feature vector for the patient, a personalized weight matrix is ​​generated, and the personalized weight matrix is ​​adaptively fused with the temporal weighted abnormal state representation and the risk degree vector to obtain a warning score; An adaptive factor is constructed based on the historical warning accuracy, the basic threshold vector is dynamically updated using the adaptive factor to obtain a current threshold vector, and the warning level is determined according to the warning score and the current threshold vector.

6. The method according to claim 1, characterized in that The warning level is integrated with the medical knowledge graph to generate prediction results including risk level, warning reason and intervention suggestions, including: Converting the warning level into a warning feature vector, wherein the warning feature vector includes a warning level probability distribution and a warning trigger feature set, wherein the warning trigger feature set is composed of features whose feature importance scores are greater than a preset threshold; A bidirectional mapping mechanism is used for integrated verification. In the forward verification, the semantic similarity between the warning feature vector and the knowledge graph path representation is calculated. The paths with semantic similarity greater than the similarity threshold are constructed as a verification path set. In the reverse verification, the expected risk level is calculated based on the path credibility of the verification path set. The initial verification result is obtained by comparing the expected risk level with the warning level predicted by deep learning. Construct a multi-level verification framework to deeply verify the initial verification results, calculate the path matching rate of the verification path set with the medical knowledge graph through the medical logic verification layer; calculate the medical verification score of the early warning time series chain in the continuous time window through the time series integrity verification layer, and the medical verification score is the weighted average of the medical logic verification results of adjacent time points in the early warning time series chain; calculate the compatibility coefficient between intervention measures through the clinical safety verification layer, and the compatibility coefficient is a standardized score calculated based on the interaction strength between intervention measures; A comprehensive verification score is calculated based on the path matching rate, medical verification score and compatibility coefficient. When the comprehensive verification score is greater than the verification threshold, the verified risk level is adopted, an early warning cause explanation chain is generated based on the verification path set, and intervention suggestions are generated from the intervention measures that have passed the safety verification.

7. The method according to claim 6, characterized in that The multi-level verification framework includes: The path matching rate between the verification path set and the medical knowledge graph is calculated through the medical logic verification layer. The path matching rate includes a direct matching score and an indirect matching score. The direct matching score is calculated based on the explicit path. The indirect matching score is calculated by building a reasoning chain to calculate the reliability of the implicit association. Different types of medical associations are weighted using learnable dynamic weights. The medical verification score of the early warning time series chain in the multi-scale time window is calculated through the time series integrity verification layer. The multi-scale time window includes a short-term window for verifying acute changes, a medium-term window for verifying disease progression, and a long-term window for verifying treatment effects. The time series attention mechanism is introduced to adaptively weight the verification results at different time points. The compatibility coefficient between interventions was calculated by the clinical safety validation layer, and the compatibility coefficient included a weighted combination of the basic safety score, the synergy score, and the temporal plausibility score, which was used to calculate a standardized score based on the strength of the interaction between the interventions.

Images