Neurology nursing risk dynamic early warning and intelligent decision support system

Abstract

The present application relates to the field of intelligent medical multi-modal data analysis and nursing auxiliary decision-making, in particular to a neurology nursing risk dynamic early warning and intelligent decision support system. The system comprises a data acquisition module, a feature extraction module, a causal reasoning module, a decision module and a feedback module. The data acquisition module collects physiological time series and nursing logs to obtain a multi-modal original data set. The feature extraction module performs nonlinear dynamics analysis on physiological data to generate a state manifold vector and performs semantic embedding on logs to generate an intervention state tensor containing time and space features. The causal reasoning module constructs a dynamic causal graph, maps vectors and tensors as nodes, and uses tensors to adjust transmission weights to suppress non-pathological fluctuations. The decision module is based on the graph counterfactual reasoning, calculates the dynamic risk and difference, generates a graded early warning and the optimal nursing path, and updates the graph parameters according to the treatment results. The system effectively solves the high-frequency false alarm problem caused by daily nursing actions of traditional monitoring equipment, reduces the alarm fatigue of medical staff, and ensures the focus on real critical events.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare multimodal data analysis and nursing decision support technology, specifically a dynamic early warning and intelligent decision support system for neurological nursing risks. Background Technology

[0002] Critical care in neurology is a crucial link in ensuring patient safety and improving prognosis. The timeliness and accuracy of risk warnings directly determine the effectiveness of medical interventions. Nursing monitoring systems mainly consist of bedside monitors and electronic nursing record terminals, which monitor patient status by collecting physiological time-series data and nursing logs. However, the condition of neurological patients during the physiological compensation period is often hidden and difficult to detect. There is a core contradiction between the high sensitivity setting of the monitoring system and frequent nursing operations, leading to a large number of false alarms. Existing technologies mostly use conventional mean analysis or linear statistical methods, which are difficult to retain the high-frequency nonlinear dynamic characteristics of physiological signals and cannot effectively integrate qualitative unstructured nursing logs with quantitative physiological indicators, resulting in a significant data heterogeneity gap. In addition, traditional monitoring lacks causal reasoning and counterfactual deduction capabilities, making it difficult to distinguish between signal artifacts caused by iatrogenic operations and real pathological deterioration, and unable to accurately assess the potential benefits of intervention measures. This results in poor specificity of warning signals and difficulty in providing accurate dynamic decision support in complex clinical environments. Therefore, there is an urgent need for a solution to address the problems existing in current technologies.

[0003] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention discloses a dynamic early warning and intelligent decision support system for neurological nursing risks. Specifically, the technical solution of this invention is as follows:

[0005] The data acquisition module is configured to collect monitoring data of the target object, including physiological time series data and unstructured nursing log data, and obtain a multimodal raw dataset;

[0006] The multimodal feature extraction module is configured to perform heterogeneous feature parsing on the multimodal raw dataset, perform nonlinear dynamic analysis on the physiological time series data to generate a state manifold vector representing the system complexity, and perform semantic embedding on the unstructured nursing log data to generate an intervention state tensor containing operational spatiotemporal features.

[0007] The causal reasoning and gating module is configured to construct a dynamic causal graph, map the state manifold vector to physiological state nodes in the graph, map the intervention state tensor to operation event nodes in the graph, and use the intervention state tensor as a control signal to adaptively gating adjust the transmission weight of the state manifold vector in the graph to suppress non-pathological signal fluctuations.

[0008] The decision support module is configured to perform counterfactual inference based on the dynamic causal graph, calculate the dynamic risk probability and counterfactual risk difference in the current state, and generate a graded early warning signal and the optimal nursing path based on the counterfactual risk difference.

[0009] The feedback optimization module is configured to update the conditional probability parameters of the dynamic causal graph based on the handling results of the graded early warning signals.

[0010] Preferably, the multimodal feature extraction module includes:

[0011] An entropy manifold construction unit is used to process the physiological time series data, retain the high-frequency noise features in the physiological time series data, and use a multi-scale sample entropy algorithm and a detrended fluctuation analysis method to calculate the complexity features of the physiological time series data at preset different time scales, and combine the complexity features to construct the state manifold vector, wherein the state manifold vector is used to characterize the stability of the target object's autonomic nervous system.

[0012] The semantic vectorization unit is used to process the unstructured nursing log data. It uses a pre-trained medical semantic model to extract semantic features of nursing operation events and drug administration events, and combines them with a preset time decay function to generate the intervention state tensor. The intervention state tensor includes operation type, operation intensity and time decay weight information.

[0013] Preferably, the causal reasoning and gating module includes:

[0014] The graph construction unit is configured to build the dynamic causal graph, which includes physiological state nodes, operational event nodes, and potential complication nodes, and defines the directed edges between nodes as conditional transition probabilities.

[0015] An adaptive gating unit is configured to perform hybrid inference, inputting the intervention state tensor as a gating signal into the gating structure of a neural network; if the intervention state tensor indicates the existence of a preset interfering nursing operation, the input weight of the gating structure to the state manifold vector is reduced, attributing the fluctuation of the physiological time series data to iatrogenic artifacts; if the intervention state tensor indicates the absence of the interfering nursing operation, the input weight is maintained or increased, attributing the fluctuation to pathological deterioration.

[0016] Preferably, the decision support module includes:

[0017] The risk calculation unit is configured to calculate, based on the dynamic causal graph, a first risk probability under the condition of maintaining the current state without intervention, and a second risk probability under the condition of implementing a preset standard care plan.

[0018] The difference assessment unit is configured to calculate the difference between the first risk probability and the second risk probability to obtain the counterfactual risk difference.

[0019] The tiered early warning unit is configured to compare the counterfactual risk difference with a preset intervention threshold: when the counterfactual risk difference is greater than the intervention threshold, a high-priority alarm is triggered and a targeted nursing intervention suggestion is output; when the counterfactual risk difference is less than or equal to the intervention threshold, silent monitoring is maintained or a low-priority prompt is output.

[0020] Preferably, the entropy manifold building unit is configured to perform the following operations when processing the physiological time-series data:

[0021] Receive the raw physiological time-series data, which includes heart rate, blood pressure, and blood oxygen saturation data;

[0022] The high-frequency fluctuation characteristics in the physiological time series data are preserved, and low-pass filtering or smoothing and denoising processing is not performed to maintain the fractal structure of the signal.

[0023] Calculate the sample entropy values ​​of the physiological time series data under a preset series of continuous time scale factors to form an entropy value sequence;

[0024] The entropy sequence is mapped to a high-dimensional feature space to generate the state manifold vector, wherein a significant decrease in the entropy in the state manifold vector is configured to characterize the depletion of physiological compensatory capacity.

[0025] Preferably, the semantic vectorization unit is configured to perform the following operations when generating the intervention state tensor:

[0026] Identify key operational entities in the unstructured nursing log data, including suctioning, turning over, back percussion, and administration of sedatives;

[0027] Assign a preset interference intensity benchmark value to each identified key operational entity;

[0028] Based on the time difference between the current moment and the moment of occurrence of the critical operation entity, the current residual interference impact value is calculated using the time decay function;

[0029] The remaining interference impact value is numerically fused with the semantic features to generate the intervention state tensor.

[0030] Preferably, the decision support module further includes:

[0031] The attribution explanation unit is used to generate a causal explanation for the graded early warning signal. When the high-priority alarm is triggered, it traces back the high-weight path in the dynamic causal graph, identifies the core physiological state node or missing operational event node that leads to the increased risk, and converts the identification result into an attribution description in natural language form, which is output synchronously with the graded early warning signal.

[0032] Preferably, the feedback optimization module includes:

[0033] The result tracking unit is used to record the actual nursing execution record and the physiological response data of the target object after the decision support module outputs the optimal nursing path;

[0034] The parameter correction unit is used to use the actual nursing execution record and the physiological response data as correction samples to calculate the deviation between the predicted results and the actual results in the dynamic causal graph, and to use the deviation to backpropagate and update the node connection weights and conditional probability tables in the dynamic causal graph.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. This system creatively utilizes the intervention state tensor generated from unstructured nursing logs as a gating signal through causal reasoning and gating modules. The system can intelligently distinguish whether fluctuations in physiological data are caused by iatrogenic artifacts due to nursing operations such as suctioning and turning over, or by pathological deterioration of the patient's condition. If it is determined to be interference from nursing operations, the system will automatically reduce the input weight of the fluctuation; otherwise, it will maintain sensitivity. This mechanism effectively solves the problem of high-frequency false alarms caused by routine nursing actions in traditional monitoring equipment, reduces alarm fatigue of medical staff, and ensures focus on real critical events.

[0037] 2. The entropy manifold construction unit of this system deliberately preserves the high-frequency fluctuation characteristics in physiological time series data; through multi-scale sample entropy and detrended fluctuation analysis, the system can capture the changes in the stability of the autonomic nervous system implied in the signal fractal structure; this method can keenly detect the depletion of the patient's physiological compensatory ability before the macroscopic vital signs collapse significantly, thus providing clinicians with a valuable golden window for early intervention.

[0038] 3. The decision support module of this system breaks through the single threshold alarm mode and introduces a counterfactual inference mechanism based on dynamic causal graphs. The system can calculate the risk probability under two assumptions: maintaining the status quo without intervention and implementing standard nursing care. It can trigger graded warnings based on the difference in counterfactual risk between the two. This mechanism not only informs people of the existence of risks, but also generates the optimal nursing path by quantifying the difference in benefits before and after intervention. It helps medical staff make the most cost-effective diagnosis and treatment decisions in complex cases and avoids over-treatment or under-intervention.

[0039] 4. This system addresses the pain point of traditional equipment's difficulty in jointly analyzing physiological parameters and nursing records. Through multimodal feature extraction, the system transforms unstructured logs into semantic tensors containing spatiotemporal features and time decay weights, achieving numerical fusion with physiological data. Furthermore, the attribution explanation unit can trace back high-weight paths in the dynamic causal graph, automatically identify core nodes leading to increased risk, and convert algorithm results into natural language descriptions, enabling medical staff to quickly understand the reasons for alarms and improving the system's clinical transparency and credibility. Attached Figure Description

[0040] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0041] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0043] Example 1:

[0044] Please see Figure 1 A dynamic early warning and intelligent decision support system for neurological nursing risks, comprising:

[0045] The data acquisition module is configured to collect monitoring data of the target object. The monitoring data includes physiological time series data and unstructured nursing log data, and obtain a multimodal raw dataset.

[0046] The multimodal feature extraction module is configured to perform heterogeneous feature parsing on the original multimodal dataset, perform nonlinear dynamic analysis on physiological time series data to generate state manifold vectors representing system complexity, and perform semantic embedding on unstructured nursing log data to generate intervention state tensors containing operational spatiotemporal features.

[0047] The causal reasoning and gating module is configured to construct a dynamic causal graph, map the state manifold vector to physiological state nodes in the graph, map the intervention state tensor to operation event nodes in the graph, and use the intervention state tensor as a control signal to adaptively gating the transmission weight of the state manifold vector in the graph to suppress non-pathological signal fluctuations.

[0048] The decision support module is configured to perform counterfactual inference based on dynamic causal graphs, calculate the dynamic risk probability and counterfactual risk difference in the current state, and generate graded early warning signals and optimal nursing pathways based on the counterfactual risk difference.

[0049] The feedback optimization module is configured to update the conditional probability parameters of the dynamic causal graph based on the handling results of the graded early warning signals.

[0050] This embodiment details the architecture of a dynamic early warning and intelligent decision support system for neurological nursing risks. This system aims to address the core technical challenges of detecting the concealed symptoms in critically ill neurological patients during their physiological compensation phase, and the numerous false alarms caused by nursing procedures in highly sensitive monitoring equipment. The data acquisition module performs the function of the sensing front-end, acquiring physiological time-series data through the high-frequency sampling interface of the bedside monitor. Derived from hardware settings, its physical meaning is the sampling rate for capturing microscopic dynamic characteristics, with the unit being Hz. In this example, it is set to be no less than 125Hz.

[0051] Meanwhile, this module captures unstructured nursing log data through the electronic medical record system (EMR) or mobile nursing terminal (PDA), and aligns the two on a unified system timeline to obtain a multimodal raw dataset; the multimodal feature extraction module performs heterogeneous feature parsing on the above dataset;

[0052] For physiological time-series data, the system does not use conventional mean analysis, but instead performs nonlinear dynamic analysis, specifically by introducing a scaling factor. The raw physiological time-series data were coarsely processed, dividing the original sequences into segments of length [missing information]. Non-overlapping windows are used and averaged to construct component sequences at different time scales. Subsequently, the entropy values ​​of each scale sequence are calculated based on the multi-scale sample entropy algorithm to quantify the dynamic complexity of the autonomic nervous system in different spatial phases. Simultaneously, detrended fluctuation analysis is used to calculate the scaling exponent. This is used to evaluate the long-range correlation of signals, thereby characterizing the fractal features of the autonomic nervous system; finally, the entropy values ​​and scaling exponents at each scale are concatenated into vectors and mapped to the state manifold vector. This vector aims to reflect the regulatory capacity of the patient's autonomic nervous system; for the nursing log, the system performs semantic embedding processing to generate an intervention state tensor containing spatiotemporal features of the operation. To quantify the potential interference of operations on physiological systems;

[0053] The causal reasoning and gating module constructs a dynamic causal graph, which transforms the state manifold vector... Mapping to physiological state nodes, the intervention state tensor The mapping is used as an operation event node; based on this, the system uses the intervention state tensor as a control signal to adaptively gating the transmission weight of physiological nodes in the graph.

[0054] The specific adjustment logic is as follows: calculate the gating coefficient. ,in For the gated weight matrix, For bias vectors, The Sigmoid activation function is used; and this coefficient is applied to the weight matrix of the graph edges, i.e. In response to the detection of strong interference, i.e. The gating mechanism suppresses the weights of the corresponding physiological nodes, attributing signal fluctuations to iatrogenic artifacts; in response to signal fluctuations without operational interference, i.e. The gate remains open, transmitting the signal downstream;

[0055] The decision support module performs counterfactual inference based on the graph, calculates the dynamic risk probability and counterfactual risk difference in the current state, and generates graded early warning signals and optimal nursing pathways accordingly; the feedback optimization module updates the conditional probability parameters of the graph in reverse according to the treatment results, and completes closed-loop learning.

[0056] Example 2:

[0057] The multimodal feature extraction module includes: an entropy manifold construction unit, which is used to process physiological time series data, retain high-frequency noise features in the physiological time series data, and use a multi-scale sample entropy algorithm and detrended fluctuation analysis method to calculate the complexity features of physiological time series data at different preset time scales, and combine the complexity features to construct a state manifold vector, wherein the state manifold vector is used to characterize the stability of the target object's autonomic nervous system;

[0058] The semantic vectorization unit is used to process unstructured nursing log data. It uses a pre-trained medical semantic model to extract semantic features of nursing operation events and medication events, and combines them with a preset time decay function to generate an intervention state tensor. The intervention state tensor contains operation type, operation intensity, and time decay weight information. The pre-trained medical semantic model refers to a deep learning language model pre-trained on a large medical corpus, such as Med-BERT or a pre-trained Transformer model with an equivalent architecture. It encodes the natural language text in the nursing log through a multi-layer self-attention mechanism, transforming the text description into a high-dimensional dense vector containing medical logical semantics.

[0059] When processing physiological time-series data, the entropy manifold building unit is configured to perform the following operations: receive raw physiological time-series data, including heart rate, blood pressure, and blood oxygen saturation data; retain the high-frequency fluctuation characteristics in the physiological time-series data, and do not perform low-pass filtering or smoothing and denoising processing to maintain the fractal structure of the signal;

[0060] The entropy values ​​of physiological time-series data under a series of preset continuous time scale factors are calculated to form an entropy value sequence; the entropy value sequence is mapped to a high-dimensional feature space to generate a state manifold vector, wherein a significant decrease in the entropy value in the state manifold vector is configured to characterize the depletion of physiological compensatory capacity;

[0061] This embodiment further specifies the entropy manifold construction unit and semantic vectorization unit in the multimodal feature extraction module; the entropy manifold construction unit receives raw physiological time-series data, including heart rate, blood pressure and blood oxygen saturation;

[0062] In this step, the unit strictly adheres to a high-frequency feature preservation strategy, prohibiting low-pass filtering or smoothing / denoising processes to maintain the fractal structure of the signal. To achieve timing alignment of heterogeneous data, the system sets a uniform time sliding window parameter: window length. Minutes, sliding step The unit is calculated using the Multi-Scale Sample Entropy (MSE) algorithm and the Detrended Fluctuation Analysis (DFA) method, with the calculation object being the current sliding window. Entropy sequence generated from extracted physiological time-series data fragments. and state manifold vector Marked as the time corresponding to the current sliding step size State characteristics;

[0063] Similarly, semantic vectorization units statistically analyze the same window of time range. The nursing log data within the system is used to generate corresponding intervention state tensors, thereby constructing a time-resolution tensor. Seconds of synchronized multimodal sequence samples; because healthy neural control signals typically manifest as Pink noise, this complex wave structure, is a manifestation of physiological adaptation; the unit is calculated using the multi-scale sample entropy algorithm (MSE) and the detrended fluctuation analysis method (DFA);

[0064] The system sets a series of continuous time scale factors. The value originates from a preset parameter set and physically represents the time granularity of the observed physiological signal. It is dimensionless and, in this example, takes the integer sequence from 1 to 20. For each scale factor, the system calculates the sample entropy value. This forms an entropy sequence. This sequence, as a discretized representation of microdynamics, is subsequently mapped to a high-dimensional state manifold vector. The system takes basic feature inputs as inputs; based on this, it further calculates the scaling index for detrended volatility analysis. , used to quantify the long-range correlation of signals;

[0065] The system performs feature fusion and mapping operations: scaling exponents spliced ​​into the entropy sequence The tail of the vector forms an augmented feature vector. The physical meaning of vector concatenation here lies in achieving complementarity between local and global features, that is, utilizing scaling exponents. For entropy sequence The evolutionary trend is constrained globally by long-range correlation to prevent overfitting of local features; among which It describes the local microscopic complexity of the system at different time granularities, while The overall long-range correlation trend of the signal is constrained, and the combination of the two constitutes a multi-view feature input, preventing information loss caused by single-dimensional analysis. Through the nonlinear mapping of the fully connected layer, the augmented feature is projected from the original feature space to the manifold space, establishing a deep mapping relationship between the entropy distribution and the stability state of the autonomic nervous system.

[0066] Will The input is fed into a fully connected neural network layer, through the formula The generation dimension is D, and in this embodiment, D is preferably set to a state manifold vector of 64 or 128, where, This is the feature mapping weight matrix of the fully connected layer. This is the bias vector of the fully connected layer. is the activation function of the linear rectifier unit; this vector is used to characterize the stability of the target object's autonomic nervous system; in response to a significant decrease in the entropy value in the vector, the system determines that its physiological compensatory capacity is exhausted.

[0067] Simultaneously, the semantic vectorization unit processes unstructured nursing log data to generate intervention state tensors. To accurately represent the complex characteristics of the operation, the tensor is constructed as a high-dimensional hybrid feature vector. Specifically, the unit utilizes a pre-trained medical semantic model to extract semantic embedding vectors of the operation events, corresponding to the operation type. It calculates the remaining interference impact value at the current moment by combining it with a preset time decay function, and integrates the intensity With decay information;

[0068] The unit concatenates the semantic embedding vector with the residual interference impact value to generate the intervention state tensor, which can simultaneously carry the semantic attributes of the operation and the physical impact intensity information, thereby eliminating the heterogeneous gap between traditional qualitative recording and quantitative analysis.

[0069] Example 3:

[0070] The causal reasoning and gating module includes: a graph construction unit, configured to build a dynamic causal graph, which contains physiological state nodes, operational event nodes, and potential complication nodes, and defines the directed edges between nodes as conditional transition probabilities;

[0071] An adaptive gating unit is configured to perform hybrid inference, inputting the intervention state tensor as a gating signal into the gating structure of the neural network;

[0072] If the intervention state tensor indicates the presence of preset interfering nursing operations, the input weights of the gating structure to the state manifold vector are reduced, attributing fluctuations in physiological time-series data to iatrogenic artifacts. Preset interfering nursing operations include clinical actions that have a transient impact on the patient's physiological signs, such as suctioning causing drastic fluctuations in heart rate and oxygen saturation, turning or patting the back causing motion artifacts, and intravenous sedation causing pharmacological relief of signs. The system determines the source of fluctuations by comparing the current operation sequence with the preset list of interfering operations. Each preset interfering nursing operation corresponds to a preset interference intensity benchmark value. ;

[0073] If the intervention state tensor indicates that there are no interfering nursing operations, then maintain or increase the input weights and attribute the fluctuations to pathological deterioration.

[0074] When generating the intervention state tensor, the semantic vectorization unit is configured to perform the following operations: identify key operational entities in unstructured nursing log data, including suctioning, turning over, back percussion, and administration of sedatives; and assign a preset interference intensity baseline value to each identified key operational entity.

[0075] Based on the time difference between the current moment and the moment of occurrence of the key operation entity, the remaining interference impact value is calculated using the time decay function; the remaining interference impact value is numerically fused with semantic features to generate the intervention state tensor.

[0076] This embodiment further specifies the generation process of causal reasoning and gating modules and intervention state tensors, and clarifies the mathematical calculation path; the semantic vectorization unit performs entity recognition steps, and uses named entity recognition technology, such as BiLSTM-CRF, to extract key operation entities from the log, such as suctioning, turning over, back percussion and sedative administration.

[0077] The unit assigns a reference value for the interference intensity to each entity. This value comes from a pre-defined lookup table in the expert knowledge base. For example: suctioning = 0.9, turning over = 0.6, medication administration = 0.3. The physical meaning is the instantaneous impact intensity of a specific operation on physiological indicators, normalized to the [0,1] interval; the unit is based on the current time. With the time of operation Time difference ,in All physical units were uniformly converted to minutes, and the remaining interference impact value was calculated using the time decay function. The calculation follows an exponential decay model. ,in The decay rate parameter for the operation type, in units of For example, suctioning procedures The setting is to reduce the effect to 10% over 15 minutes;

[0078] The system will semantic embedding vector of the operation Numerical fusion is performed, and a splicing operation is used to generate the intervention state tensor. At this point, the concatenated tensor structurally exhibits an added intensity dimension at the end of the semantic feature dimension, resulting in a tensor with the following dimensions: ,in, For the dimension of the semantic embedding vector, this The dimensional hybrid vector structure ensures that subsequent networks can simultaneously process the quality and quantity of operations within the same feature space.

[0079] Building upon this, the graph construction unit establishes a dynamic causal graph, which is instantiated as a graph neural network. The directed edges between physiological state nodes, operational event nodes, and potential complication nodes are defined as learnable weight matrices. The numerical distribution of this matrix strictly corresponds to the topological structure of the causal graph: zero elements in the matrix indicate that there is no direct causal path between nodes, and the magnitude and sign of non-zero elements represent the strength and direction of the causal interaction, respectively; where... elements in The numerical value directly corresponds to the node in the causal graph. For nodes The strength of the direct causal effect, with positive values ​​representing promotion / activation and negative values ​​representing inhibition / antagonism; The sparse structure directly reflects the topological connectivity of the causal graph, that is, when When this occurs, it indicates that there is no direct causal path between nodes; the adaptive gating unit performs hybrid inference;

[0080] This unit will intervene in the state tensor. The signal is input into the neural network as a gating signal; in order to achieve differentiated regulation of different physiological nodes, the system does not calculate a single scalar, but rather calculates the relationship between the number of physiological nodes and the input signal. Equal gate vectors The specific gating formula is as follows: ,in , For the Sigmoid function, ensure the output is in The system uses this vector to perform row-by-row broadcast correction on the adjacency matrix: ; In response to the intervention state tensor indicating the presence of highly disruptive nursing procedures, i.e. The system specifically inhibits the first Each physiological node's output weights to other nodes are used to attribute the node's fluctuations to iatrogenic artifacts; in response to If the input weights are maintained, the fluctuations will be attributed to pathological deterioration.

[0081] Example 4:

[0082] The decision support module includes a risk calculation unit, configured to calculate the first risk probability under the condition of maintaining the current state without intervention, and the second risk probability under the condition of implementing a preset standard care plan, based on a dynamic causal graph.

[0083] The difference assessment unit is configured to calculate the difference between the first risk probability and the second risk probability to obtain the counterfactual risk difference.

[0084] The tiered early warning unit is configured to compare the counterfactual risk difference with a preset intervention threshold: when the counterfactual risk difference is greater than the intervention threshold, a high-priority alarm is triggered and targeted nursing intervention suggestions are output; when the counterfactual risk difference is less than or equal to the intervention threshold, silent monitoring is maintained or a low-priority prompt is output.

[0085] The decision support module also includes: an attribution explanation unit, which generates causal explanations for graded early warning signals. When a high-priority alarm is triggered, it traces back the high-weighted paths in the dynamic causal graph, identifies the core physiological state nodes or missing operational event nodes that lead to increased risk, and converts the identification results into attribution descriptions in natural language form, which are output synchronously with the graded early warning signals.

[0086] This embodiment further specifies the counterfactual reasoning mechanism and attribution explanation function in the decision support module, and clarifies the specific operators; the risk calculation unit calculates two key probabilities in parallel based on the dynamic causal graph, and adopts the do-calculus operator in the structural causal model;

[0087] First risk probability ,in For complication node variables, For nursing operation node variables, Let be the current physiological state node variable, physically representing the incidence of complications without additional intervention. The calculation of this probability involves the following structural intervention steps: in the adjacency matrix... In the middle, all pointers to the operation node The column vectors are set to zero, that is... This is done to sever the upstream dependency path of the node; in the input feature matrix In the middle, the operation node The corresponding feature vector is forced to be zero vector; the forward propagation of the graph convolutional network is performed based on the modified adjacency matrix and feature matrix, and the calculation method is the Sigmoid activation value after the forward propagation of the graph network;

[0088] The specific forward propagation process follows the hierarchical update rules of graph convolutional networks: ,in, This represents an adjacency matrix with added self-loops, used to preserve the node's own characteristic information during message passing; for The degree matrix, here This constitutes the symmetric normalized Laplace operator, used to prevent the explosion or disappearance of characteristic values ​​during multi-level propagation. For the first The node feature matrix of the layer represents the latent state distribution after aggregating neighborhood information, i.e., the distribution of the network in the 1st layer. The high-dimensional representation of nodes abstracted from layers;

[0089] For the first The trainable weight matrix of a layer plays the role of feature transformation and dimension mapping, and is used to extract node features and pass effective information between layers. It is a non-linear activation function. Initialized as a physiological state node and operation event nodes The characteristic matrix formed; after After layer propagation, extract complication nodes. Corresponding feature vector The output probability is achieved through a fully connected layer: ;

[0090] Second risk probability The physical meaning is the forced implementation of a preset standard care plan; during calculation, the direction is also cut off. The connection weights, and the feature matrix Middle operation node The feature vectors are forcibly replaced with semantic embedding vectors of preset standard nursing procedures. Perform forward propagation;

[0091] The difference assessment unit calculates the difference between the two to obtain the counterfactual risk difference. This parameter quantifies the potential benefits of intervention; the tiered early warning unit will... With the preset intervention threshold For example, 0.3, for comparison;

[0092] In response to If the system determines that the consequences of not intervening are severe and the benefits of intervention are significant, it will trigger a high-priority alert and output nursing recommendations; in response to The system maintains silent monitoring or outputs low-priority prompts;

[0093] Meanwhile, the attribution explanation unit is activated when a high-priority alarm is triggered; this unit uses an attention backtracking mechanism to calculate the gradient contribution of each node in the graph to the output node. ,in For the first in the atlas Each node; system identification The largest front Each node serves as a core physiological state node, such as a sharp drop in entropy or a missing operational event node; the unit converts the recognition results into an attribution description in natural language form through template filling, and outputs it synchronously with the alarm.

[0094] Example 5:

[0095] The feedback optimization module includes: a result tracking unit, which records the actual nursing execution records and the physiological response data of the target object after the decision support module outputs the optimal nursing path; and a parameter correction unit, which uses the actual nursing execution records and physiological response data as correction samples to calculate the deviation between the predicted results and the actual results in the dynamic causal graph, and uses the deviation to backpropagate and update the node connection weights and conditional probability tables in the dynamic causal graph.

[0096] This embodiment further specifies the self-evolution mechanism of the feedback optimization module; the result tracking unit continues to run after the system outputs the optimal nursing path, recording the actual nursing execution record, including whether the nurse performed the operation and the execution time, as well as the physiological response data of the target object, i.e., the entropy manifold vector after the operation. Changes;

[0097] The parameter correction unit constructs an online learning closed loop; this unit uses the above records as correction samples to calculate the prediction bias. Given that the state manifold vector is a high-dimensional feature, the system uses Euclidean distance, i.e., the L2 norm, to measure the bias. The formula is as follows: ,in For the predicted state vector, This is the actual observed state vector;

[0098] The system uses a multi-task joint loss function to measure deviation; this loss function consists of two parts: the first part is the state reconstruction loss. The first part is used to constrain the physical consistency of physiological state evolution; the second part is risk prediction loss. That is, binary cross-entropy loss, which only exists in electronic medical records with clear complication diagnosis records. Activated at time; the total loss of the system is ,in The hyperparameters for balancing the weights;

[0099] The system utilizes this deviation The backpropagation algorithm is executed to update the node connection weights and conditional probability tables in the dynamic causal graph; the updated causal graph is used for risk assessment in the next time step, realizing the dynamic iteration of model parameters.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic early warning and intelligent decision support system for neurological nursing risks, characterized in that, include: The data acquisition module is configured to collect monitoring data of the target object, including physiological time series data and unstructured nursing log data, and obtain a multimodal raw dataset; The multimodal feature extraction module is configured to perform heterogeneous feature parsing on the multimodal raw dataset, perform nonlinear dynamic analysis on the physiological time series data to generate a state manifold vector representing the system complexity, and perform semantic embedding on the unstructured nursing log data to generate an intervention state tensor containing operational spatiotemporal features. The causal reasoning and gating module is configured to construct a dynamic causal graph, map the state manifold vector to physiological state nodes in the graph, map the intervention state tensor to operation event nodes in the graph, and use the intervention state tensor as a control signal to adaptively gating adjust the transmission weight of the state manifold vector in the graph to suppress non-pathological signal fluctuations. The decision support module is configured to perform counterfactual inference based on the dynamic causal graph, calculate the dynamic risk probability and counterfactual risk difference in the current state, and generate a graded early warning signal and the optimal nursing path based on the counterfactual risk difference. The feedback optimization module is configured to update the conditional probability parameters of the dynamic causal graph based on the handling results of the graded early warning signals. The multimodal feature extraction module includes: An entropy manifold construction unit is used to process the physiological time series data, retain the high-frequency noise features in the physiological time series data, and use a multi-scale sample entropy algorithm and a detrended fluctuation analysis method to calculate the complexity features of the physiological time series data at preset different time scales, and combine the complexity features to construct the state manifold vector, wherein the state manifold vector is used to characterize the stability of the target object's autonomic nervous system. The semantic vectorization unit is used to process the unstructured nursing log data. It uses a pre-trained medical semantic model to extract semantic features of nursing operation events and drug administration events, and combines them with a preset time decay function to generate the intervention state tensor. The intervention state tensor includes operation type, operation intensity and time decay weight information. The causal reasoning and gating module includes: The graph construction unit is configured to build the dynamic causal graph, which includes physiological state nodes, operational event nodes, and potential complication nodes, and defines the directed edges between nodes as conditional transition probabilities. An adaptive gating unit is configured to perform hybrid inference, inputting the intervention state tensor as a gating signal into the gating structure of a neural network; if the intervention state tensor indicates the existence of a preset interfering nursing operation, the input weight of the gating structure to the state manifold vector is reduced, attributing the fluctuation of the physiological time series data to iatrogenic artifacts; if the intervention state tensor indicates the absence of the interfering nursing operation, the input weight is maintained or increased, attributing the fluctuation to pathological deterioration.

2. The neurology nursing risk dynamic early warning and intelligent decision support system according to claim 1, characterized in that, The decision support module includes: The risk calculation unit is configured to calculate, based on the dynamic causal graph, a first risk probability under the condition of maintaining the current state without intervention, and a second risk probability under the condition of implementing a preset standard care plan. The difference assessment unit is configured to calculate the difference between the first risk probability and the second risk probability to obtain the counterfactual risk difference. The tiered early warning unit is configured to compare the counterfactual risk difference with a preset intervention threshold: when the counterfactual risk difference is greater than the intervention threshold, a high-priority alarm is triggered and a targeted nursing intervention suggestion is output; when the counterfactual risk difference is less than or equal to the intervention threshold, silent monitoring is maintained or a low-priority prompt is output.

3. The neurological nursing risk dynamic early warning and intelligent decision support system according to claim 1, characterized in that, When processing the physiological time-series data, the entropy manifold building unit is configured to perform the following operations: Receive the raw physiological time-series data, which includes heart rate, blood pressure, and blood oxygen saturation data; The high-frequency fluctuation characteristics in the physiological time series data are preserved, and low-pass filtering or smoothing and denoising processing is not performed to maintain the fractal structure of the signal. Calculate the sample entropy values ​​of the physiological time series data under a preset series of continuous time scale factors to form an entropy value sequence; The entropy sequence is mapped to a high-dimensional feature space to generate the state manifold vector, wherein a significant decrease in the entropy in the state manifold vector is configured to characterize the depletion of physiological compensatory capacity.

4. The neurological nursing risk dynamic early warning and intelligent decision support system according to claim 1, characterized in that, When generating the intervention state tensor, the semantic vectorization unit is configured to perform the following operations: Identify key operational entities in the unstructured nursing log data, including suctioning, turning over, back percussion, and administration of sedatives; Assign a preset interference intensity benchmark value to each identified key operational entity; Based on the time difference between the current moment and the moment of occurrence of the critical operation entity, the current residual interference impact value is calculated using the time decay function; The remaining interference impact value is numerically fused with the semantic features to generate the intervention state tensor.

5. The neurology nursing risk dynamic early warning and intelligent decision support system according to claim 1, characterized in that: The decision support module also includes: The attribution explanation unit is used to generate a causal explanation for the graded early warning signal. When a high-priority alarm is triggered, it traces back the high-weight path in the dynamic causal graph, identifies the core physiological state node or missing operational event node that leads to the increased risk, and converts the identification result into an attribution description in natural language form, which is output synchronously with the graded early warning signal.

6. The neurology nursing risk dynamic early warning and intelligent decision support system according to claim 1, characterized in that: The feedback optimization module includes: The result tracking unit is used to record the actual nursing execution record and the physiological response data of the target object after the decision support module outputs the optimal nursing path; The parameter correction unit is used to use the actual nursing execution record and the physiological response data as correction samples to calculate the deviation between the predicted results and the actual results in the dynamic causal graph, and to use the deviation to backpropagate and update the node connection weights and conditional probability tables in the dynamic causal graph.

Images