A pressure injury intelligent quantification risk assessment and intervention guidance system

Abstract

This invention discloses an intelligent quantitative risk assessment and intervention guidance system for pressure injuries, relating to the field of medical and nursing technology. It replaces traditional statistical association analysis with a structural causal model, constructing a two-layer directional acyclic causal graph consistent with the pathological mechanism of pressure injuries. Combined with a backdoor adjustment criterion, it effectively eliminates confounding bias, enabling risk assessment to distinguish between pathogenic factors and spurious correlations. Through counterfactual reasoning, it quantifies the marginal causal effect of each interventionable factor, clarifying the risk reduction magnitude of different intervention measures, providing a prioritization basis for clinical nursing, and improving the efficiency of nursing resource utilization. It uses a visual format of causal graph overlaid with a heatmap to display the risk transmission path and the degree of factor contribution, enhancing decision interpretability and improving clinical acceptance. Furthermore, it generates individualized targeted intervention plans based on real-time multimodal patient data, effectively reducing the incidence of pressure ulcers in hospitalized patients.

Description

Technical Field

[0001] This invention relates to the field of medical and nursing technology, specifically to an intelligent quantitative risk assessment and intervention guidance system for stress injuries. Background Technology

[0002] Pressure injuries, also known as pressure ulcers or bedsores, are a common complication in patients who are bedridden for a long time and have limited mobility. Their occurrence is closely related to multiple factors such as prolonged pressure on local tissues, microcirculatory disorders, damage to the skin barrier, and abnormal systemic metabolic state. They not only prolong the patient's hospitalization period and increase medical and nursing costs, but in severe cases, they can lead to secondary infections and even endanger the patient's life.

[0003] Current mainstream clinical pressure ulcer risk assessment tools rely heavily on manual assessment scales such as the Braden and Norton scales. These tools depend heavily on the clinical experience and subjective judgment of nursing staff, resulting in inherent limitations such as lengthy assessment times, poor consistency among assessors, and the inability to provide real-time updates. While some intelligent assessment systems utilize machine learning algorithms to build risk prediction models based on historical case data, improving assessment efficiency to some extent, these models are essentially statistical association analyses. They can only uncover the correlation between variables and outcomes, failing to distinguish between true causative factors and spurious associations caused by confounding factors. They are susceptible to assessment bias due to confounding variables and cannot reveal the underlying pathological transmission mechanisms of pressure ulcers.

[0004] In addition, existing technologies cannot quantify the causal contribution of a single risk factor to the overall risk, nor can they provide a quantitative estimate of the risk reduction corresponding to different intervention measures. This results in a lack of clear priority criteria for clinical interventions, leading to the adoption of homogeneous and universal nursing protocols, and an inability to achieve targeted interventions based on individual patient conditions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent quantitative risk assessment and intervention guidance system for pressure injuries. This system constructs a two-layer directional acyclic causal graph containing latent variable layers and observed variable layers, and combines prior clinical knowledge with historical case data to fit the causal path structure and path coefficients. Through counterfactual reasoning, it calculates the marginal causal effect of each interventionable risk factor on the pressure ulcer risk score, quantifies the causal contribution of each factor, and ranks them by proportion. The causal graph visually displays the risk transmission path and the degree of factor contribution, and generates individualized intervention plans with quantitative execution parameters based on the weighted ranking of causal effects.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a smart quantitative risk assessment and intervention guidance system for pressure injuries, the system comprising:

[0007] The multimodal data acquisition module is used to access multi-source heterogeneous pressure-related damage data and output standardized observation data after preprocessing the data.

[0008] The structural causal model construction module is used to construct a two-layer directed acyclic causal graph containing a latent variable layer and an observed variable layer. It uses tissue vulnerability, microenvironmental conditions, mechanical load, metabolic state, and nursing compliance as latent variables, and combines clinical prior knowledge with historical case data to fit and determine the causal path structure and path coefficients.

[0009] The counterfactual reasoning calculation module is used to map real-time standardized observation data to a structural causal model, identify backdoor paths through depth-first traversal, perform counterfactual reasoning based on the backdoor adjustment criterion, and calculate the marginal causal effect of each interventionable risk factor on the pressure ulcer risk score.

[0010] The risk quantification and assessment module is used to integrate the results of causal reasoning, output a comprehensive risk score for stress injury and the corresponding risk level, and calculate the causal contribution of each risk factor and sort them by proportion.

[0011] The causal visualization output module is used to generate a directional acyclic causal graph with superimposed thermodynamic mapping;

[0012] The individualized intervention guidance module is used to screen clinically modifiable factors and rank them by causal effect weighting to generate individualized intervention plans.

[0013] Furthermore, the data sources accessed by the multimodal data acquisition module include body surface pressure monitoring devices, skin microenvironment sensors, hospital laboratory information systems, nursing documentation systems, and bedside vital signs monitors; the data preprocessing process performs time-series resampling alignment at a fixed time granularity, identifies and removes outliers using the 3σ criterion, fills missing data using the multiple interpolation chain equation method, and performs Z-score standardization on continuous observation variables to eliminate dimensional differences.

[0014] Furthermore, the latent variable layer includes exogenous latent variables and endogenous outcome variables, while the observed variable layer consists of directly collectable clinical indicators. Each latent variable corresponds to multiple directly collectable observed variables. Confirmatory factor analysis is used to characterize the loading relationship between latent variables and observed variables, and observed variables with an absolute value of loading coefficient greater than 0.6 are retained.

[0015] The exogenous latent variables include

[0016] Tissue vulnerability latent variables: Observed variables corresponding to skin elasticity value, serum albumin concentration, peripheral transcutaneous oxygen partial pressure, patient age and skin integrity score, reflecting the inherent ability of the patient's skin and subcutaneous tissue to withstand pressure damage;

[0017] Latent variables of microenvironmental conditions: These are observed variables corresponding to local skin temperature, relative skin humidity, skin surface pH, and sweat secretion rate, reflecting the influence of the local environment of the skin on tissue tolerance.

[0018] Mechanical load latent variables: Observed variables corresponding to the peak pressure at the sacrococcygeal interface, pressure duration, estimated local shear force, and body position change interval, reflecting the level of external force load acting on local tissues;

[0019] Metabolic status latent variables: corresponding to fasting blood glucose level, hemoglobin concentration, body mass index and C-reactive protein level, reflecting the impact of the patient's overall metabolic and nutritional status on tissue repair capacity;

[0020] Latent variables of nursing compliance: These are observed variables corresponding to the rate of turning over, the completion rate of skin cleaning and care, the implementation rate of nutritional support, and the usage rate of decompression devices, reflecting the degree of implementation of clinical nursing measures.

[0021] Furthermore, the causal paths of the directional acyclic causal graph include: a direct path of mechanical load to pressure injury risk, a direct path of tissue vulnerability to pressure injury risk, an indirect mediating path of metabolic state to tissue vulnerability, an indirect mediating path of microenvironmental conditions to tissue vulnerability, an indirect regulatory path of nursing compliance to mechanical load, and an indirect regulatory path of nursing compliance to microenvironmental conditions.

[0022] Furthermore, the structural causal model construction module employs a PC algorithm with prior constraints for causal structure learning, uses maximum likelihood estimation to fit the standardized path coefficients of all directed edges, verifies the model's goodness of fit by comparing the fit index, Tucker-Lewis index, and root mean square index of approximation error, and uses bootstrap sampling to calculate the 95% confidence interval of the path coefficients, excluding insignificant paths whose confidence intervals contain 0.

[0023] Furthermore, the counterfactual reasoning calculation module substitutes the preprocessed real-time observed variables into the fitted structural causal model, calculates the current actual values ​​of each latent variable using the regression factor scoring method, and simultaneously calculates the baseline pressure ulcer risk score under the current state. For each variable to be intervened, a depth-first search algorithm is used to traverse the directed acyclic graph to identify all backdoor paths from the intervention variable to the outcome variable, i.e., non-causal paths with the intervention variable as the starting point, the outcome variable as the ending point, and edges pointing to the intervention variable. Confounding variables on each backdoor path are extracted to form a confounding variable set for the corresponding intervention variable. Based on the backdoor adjustment criterion, the distribution of the confounding variables is integrated to block all backdoor paths, and the expected outcome when the intervention variable is set to the target optimal value is calculated.

[0024] Furthermore, the formula for calculating the expected outcome is as follows: ,in, Indicates the intervention variable Perform intervention operations Afterwards, the outcome variable The conditional expectation, i.e., the expected value of the pressure ulcer risk score after adjusting for the corresponding risk factors to the target value. This indicates the modifiable risk variables to be assessed, corresponding to modifiable latent variables such as mechanical load, microenvironmental conditions, and nursing compliance; Indicating intervention variables The target optimal value is the standardized value of the latent variable under the ideal clinical recommendation. Indicating intervention variables The corresponding set of promiscuous variables, i.e., all those located in arrive Variables on the backdoor path; Represents a set of mixed variables The space of all possible values; Represents a set of mixed variables A specific combination of values; Indicates when Take target value And the mixed variable set At that time, the outcome variable The expected value of the condition; Represents a set of mixed variables Values The marginal probability.

[0025] Furthermore, the counterfactual reasoning calculation module subtracts the expected outcome after intervention from the current baseline risk score to obtain the marginal causal effect size. The calculation formula is: ,in, Indicating intervention variables The average marginal causal effect is the change in pressure ulcer risk score after the factor is adjusted to the target value; a negative value indicates a reduction in risk after intervention, and the larger the absolute value, the more significant the risk reduction effect of the intervention on this factor. This represents the baseline risk score for pressure ulcers in the current state, i.e., the patient's real-time risk score when no intervention has been performed.

[0026] Furthermore, the risk quantification assessment module will use the posterior probability of pressure ulcer occurrence output by the structural causal model. By mapping the score to the 0-100 range using a linear transformation, a comprehensive causal risk score is obtained. The calculation formula is: ,in, The comprehensive causal risk score for pressure injury ranges from 0 to 100, with a higher score indicating a higher probability of pressure injury. This represents the posterior probability of a patient experiencing stress injury, as output by the structural causal model, and its value ranges from 0 to 1.

[0027] The risk quantification assessment module is divided into four risk levels according to the scoring range:

[0028] Low risk Medium risk is High risk is Extremely high risk Different levels correspond to different nursing response levels.

[0029] Furthermore, the risk quantification assessment module calculates the proportion of the absolute value of the total causal effect of each risk factor to the sum of the absolute values ​​of the total causal effects of all factors, as the causal contribution of that factor, i.e.: ,in, Indicates the first The causal contribution of each risk factor reflects the degree to which that factor contributes to the current overall risk. Indicates the first The marginal causal effect of each risk factor; This indicates the total number of risk factors involved in the contribution calculation; Indicates the first The marginal causal effect of each risk factor.

[0030] Furthermore, the causal visualization output module uses the Sugiyama hierarchical layout algorithm to automatically arrange the directional acyclic causal graph; the node color level changes linearly from dark blue to dark red, corresponding to a causal contribution of 0% to 100%; the line width of the directed edge is positively correlated with the absolute value of the standardized path coefficient of the corresponding path.

[0031] Furthermore, the individualized intervention guidance module categorizes risk factors into short-term interventionable, medium- to long-term interventionable, and non-interventionable categories, prioritizing the generation of intervention plans for short-term interventionable categories.

[0032] Furthermore, the individualized intervention guidance module is ordered according to causal effect weighting, and its priority calculation formula is as follows: ,in, Indicates the first The priority score of each intervention factor is assigned; the higher the score, the higher the intervention priority. Indicates the first The clinical feasibility weight of each intervention factor is assigned, with a value ranging from 0 to 1.

[0033] Compared with existing technologies, this intelligent quantitative risk assessment and intervention guidance system for pressure injuries has the following beneficial effects:

[0034] This invention replaces traditional statistical association analysis with a structural causal model. By constructing a two-layer directional acyclic causal graph that conforms to the pathological mechanism of pressure injury, and combining it with the backdoor adjustment criterion, it effectively eliminates confounding bias, enabling risk assessment to distinguish between pathogenic factors and spurious correlations. Through counterfactual reasoning, it quantifies the marginal causal effect of each interventionable factor, clarifies the risk reduction magnitude of different intervention measures, provides a basis for prioritizing clinical nursing, and improves the efficiency of nursing resource utilization. It uses a visual representation of the risk transmission path and the degree of factor contribution by overlaying the causal graph with a heatmap, enhancing the interpretability of decisions and improving clinical acceptance. Furthermore, it generates individualized targeted intervention plans based on real-time multimodal patient data, effectively reducing the incidence of pressure ulcers in hospitalized patients.

[0035] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0037] Figure 1 This is a structural block diagram of an intelligent quantitative risk assessment and intervention guidance system for pressure injuries;

[0038] Figure 2 A flowchart illustrating a constrained PC algorithm used in an intelligent quantitative risk assessment and intervention guidance system for pressure injuries;

[0039] Figure 3 The flowchart shows the maximum likelihood estimation method used in a smart quantitative risk assessment and intervention guidance system for pressure injury. Detailed Implementation

[0040] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0041] This embodiment provides an intelligent quantitative risk assessment and intervention guidance system for pressure injuries, which can be deployed on hospital nursing information platforms, bedside monitoring terminals, or cloud-based nursing decision servers. It is used to achieve dynamic quantitative assessment and individualized intervention guidance for the risk of pressure injuries in hospitalized patients. Figure 1 As shown, the system includes a multimodal data acquisition module, a structural causal model construction module, a counterfactual reasoning calculation module, a risk quantification assessment module, a causal visualization output module, and an individualized intervention guidance module. Each module is connected sequentially according to the data flow direction, and works together to complete the entire process from multi-source data access to individualized intervention plan output.

[0042] The multimodal data acquisition module is used to access multi-source heterogeneous pressure-related damage data, and outputs standardized observation data after preprocessing the data, providing input data in a unified format for subsequent causal calculations.

[0043] The multimodal data acquisition module supports standard medical data interfaces such as HL7 and FHIR. The data sources it can access include interface pressure time-series data output by body surface pressure monitoring devices, local skin physiological parameters collected by skin microenvironment sensors, laboratory test results stored in the hospital's laboratory information system, nursing execution logs recorded in the nursing documentation system, and real-time vital signs data collected by bedside vital signs monitors. All data sources are aligned with a unified system timestamp to ensure the time consistency of multi-source data.

[0044] The data preprocessing process involves time-series resampling and alignment at a fixed time granularity, outlier identification and removal using the 3σ criterion, missing data filling using the multiple interpolation chain equation method, and Z-score standardization for continuous observation variables to eliminate dimensional differences.

[0045] For time series resampling alignment: linear interpolation is used to perform fixed-granularity resampling on all time series data. Under normal monitoring conditions, the time granularity is set to 1 hour, which can be dynamically adjusted to 15 minutes for extremely high-risk patients to ensure that the values ​​of each dimension of data correspond one-to-one at the same time node, forming a synchronous multi-dimensional observation sequence.

[0046] For outlier removal: The 3σ criterion is used to identify and remove outlier data: Based on a 24-hour sliding time window, the mean and standard deviation of all valid data within the window are calculated for each variable. Data that exceed the mean ± 3 times the standard deviation are marked as outliers and removed to eliminate interference data introduced by factors such as equipment failure, operational errors, and human input errors.

[0047] For missing value imputation: The multiple imputation chain equation method is used to imput missing data segments. The variable with missing values ​​is used as the dependent variable and the other clinically relevant observation variables are used as independent variables to construct a regression prediction equation. The number of iterations is set to 20. After generating 5 sets of reasonable imputation values, the mean is taken as the final imputation result to ensure the integrity and continuity of the data time series.

[0048] For standardization, Z-score standardization is performed on all continuous observation variables, calculated using the following formula: ,in, For the original values ​​of the observed variable, This represents the population mean of the variable on a large sample historical case dataset. The population standard deviation of this variable is 0. After standardization, the mean of all observed variables is 0 and the standard deviation is 1, which can eliminate the difference in dimensions between different indicators and ensure the comparability of subsequent causal coefficient calculations.

[0049] The structural causal model construction module is used to construct a two-layer directed acyclic causal graph containing a latent variable layer and an observed variable layer. Tissue vulnerability, microenvironmental conditions, mechanical load, metabolic state, and nursing compliance are used as latent variables. The causal path structure and path coefficients are determined by fitting clinical prior knowledge and historical case data to form a structural causal model that can be used for quantitative reasoning.

[0050] Specifically, the latent variable layer is divided into two categories: exogenous latent variables and endogenous outcome variables. The observed variable layer consists of directly collectable clinical indicators. In this embodiment, a total of 5 categories of exogenous latent variables and 1 category of endogenous outcome variables are preset. Each category of exogenous latent variables is pre-matched with multiple corresponding observed variables. Confirmatory factor analysis (CFA) is used to test the construct validity of the latent variables and screen effective observed variables. Specifically, a measurement model is constructed, and each observed variable is set to be represented by a linear combination of the corresponding latent variable and the measurement error term, expressed as: In the formula, Let i be the standardized value of the i-th observed variable. For the corresponding latent variables, The factor loading coefficient for this observed variable. Let's assume the measurement error term is... Follows a mean of 0 and a variance of The latent variables are normally distributed, and the error terms of each observed variable are independent. Maximum likelihood estimation is used to fit the measurement model, and the standardized factor loading coefficient of each observed variable is calculated. The absolute value of the loading coefficient represents the degree to which the observed variable explains the variation of the latent variable. Then, a construct validity test is performed. When the comparative fit index (CFI) of the measurement model is greater than 0.9, the Tucker-Lewis index (TLI) is greater than 0.9, and the root mean square error (RMSEA) of approximation is less than 0.08, the construct validity of the latent variables is deemed to meet the requirements. Based on this, observed variables with an absolute value of loading coefficient greater than 0.6 are selected and retained in the final model, while observed variables with insufficient loading are removed to ensure that the latent variables have good convergent validity.

[0051] After screening, the final observed variables and their clinical significance for each latent variable are as follows:

[0052] Tissue vulnerability latent variables: corresponding to five observed variables, namely skin elasticity value, serum albumin concentration, peripheral transcutaneous oxygen partial pressure, patient age, and skin integrity score, which comprehensively reflect the inherent ability of the patient's skin and subcutaneous tissue to withstand pressure damage;

[0053] Latent variables of microenvironment conditions: These correspond to four observed variables: local skin temperature, relative skin humidity, skin surface pH, and sweat secretion rate, reflecting the degree of influence of the local environment on tissue tolerance.

[0054] Mechanical load latent variables: corresponding to four observed variables, namely, peak pressure at the sacrococcygeal interface, pressure duration, estimated local shear force, and body position change interval, which reflect the comprehensive level of external force load acting on local tissues;

[0055] Metabolic status latent variables: corresponding to four observed variables, namely fasting blood glucose level, hemoglobin concentration, body mass index, and C-reactive protein level, which reflect the impact of the patient's overall metabolic and nutritional status on tissue repair capacity;

[0056] Latent variables of nursing compliance: corresponding to four observation variables, namely, the rate of turning over, the completion rate of skin cleaning and care, the implementation rate of nutritional support, and the usage rate of decompression devices, which reflect the actual implementation degree of various clinical nursing measures.

[0057] The endogenous outcome variable is the stress injury risk score, which corresponds to the posterior probability of stress injury and is the final output variable of the model.

[0058] The structural causal model construction module is based on the clinical pathological mechanism of pressure injury and pre-applies two types of constraints to the causal graph structure: The first type is a mandatory existence path constraint, which pre-sets the core causal paths that must be retained and cannot be deleted during algorithm execution. Specifically, this includes six paths: mechanical load pointing to pressure injury risk, tissue vulnerability pointing to pressure injury risk, metabolic state pointing to tissue vulnerability, microenvironmental conditions pointing to tissue vulnerability, nursing compliance pointing to mechanical load, and nursing compliance pointing to microenvironmental conditions. The second type is a directional constraint, which limits the causal direction between latent variables and prohibits reverse paths that contradict the pathological mechanism. For example, it prohibits tissue vulnerability from pointing to metabolic state and prohibits pressure injury risk from pointing to various exogenous latent variables.

[0059] Based on prior constraints, a constrained PC algorithm is used to learn the causal structure of the de-identified historical case dataset: specifically, such as... Figure 2 As shown, a fully connected undirected graph is constructed with all latent variables as nodes, and an undirected edge exists between any two nodes. For each pair of adjacent nodes, the independence of the two variables under the zero-order condition is tested using the Pearson partial correlation test, with a significance level of α=0.05. If the p-value corresponding to the partial correlation coefficient between the two variables is greater than α, and the edge is not a mandatory retention path, then the undirected edge between the two nodes is deleted. The order of the condition set is increased sequentially, from order 1 to... For each pair of still adjacent nodes, traverse all adjacent node subsets of the corresponding order as a condition set, and test the conditional independence of the two variables under this condition set. If there exists any condition set that makes the two variables conditionally independent, and the edge is not a forced retention path, then delete the undirected edge between the two nodes, until all orders have been tested.

[0060] Based on V-structure recognition and prior direction constraints, all retained undirected edges are oriented, all collision structures (V-structures) are identified, and the direction of the corresponding edges is determined. Combined with pre-set direction constraints, the remaining undirected edges are oriented. At the same time, loop detection is used to ensure that there are no directed loops in the final graph structure, thus obtaining the initial structure of the oriented acyclic causal graph.

[0061] After determining the causal structure, the maximum likelihood estimation (MLE) method is used to fit the standardized path coefficients of all directed edges to obtain the quantitative effect strength of each causal path. Specifically, for example... Figure 3 As shown, the sample covariance matrix of all latent variables is calculated based on the historical case dataset. Based on the structure of the directed acyclic causal graph, the implicit covariance matrix of the model is derived. ,in For the set of path coefficient parameters to be solved, construct the maximum likelihood fitting function: ,in, tr represents the total number of latent variables. The trace of the matrix, Given the logarithm of the matrix determinant, the iterative generalized least squares method is used to iteratively solve the fitted function. When the change in the fitted function value before and after iteration is less than... When convergence is determined, the unstandardized path coefficients of all directed edges are obtained. To achieve horizontal comparability of the influence strength of different paths, all path coefficients are standardized. Based on the sample standard deviation of latent variables, the unstandardized coefficients are converted into standardized path coefficients. The absolute value of the standardized path coefficients ranges from 0 to 1. The larger the absolute value, the stronger the influence of the causal path.

[0062] After the fitting is completed, a goodness-of-fit test is performed using the Comparative Fit Index (CFI), the Tucker-Lewis Index (TLI), and the Root Mean Square Error (RMSEA). When CFI ≥ 0.9, TLI ≥ 0.9, and RMSEA ≤ 0.08, the overall model fitting effect is considered to meet the requirements for clinical application.

[0063] The significance of path coefficients was tested using a bootstrap method: repeated sampling with replacement was performed on the historical case dataset, with a total of 1000 samplings. In each sampling, the coefficients of all paths were refitted and calculated to obtain the empirical sampling distribution of the coefficients of each path. Based on the sampling distribution, the 95% confidence intervals of the path coefficients were calculated. If the 95% confidence interval of a path contained the value 0, the path was determined to be statistically insignificant and was removed. Finally, all significant paths were retained to obtain the optimized structural causal model.

[0064] The counterfactual reasoning calculation module is used to map real-time standardized observation data to a structural causal model, identify backdoor paths through depth-first traversal, perform counterfactual reasoning based on the backdoor adjustment criterion, calculate the marginal causal effect of each interventionable risk factor on the pressure ulcer risk score, and quantify the risk reduction after intervention of each factor.

[0065] Specifically, the preprocessed real-time observed variables are substituted into the fitted structural causal model, and the regression factor scoring method is used to calculate the current actual values ​​of each latent variable. Then, the values ​​of each latent variable are substituted into the structural equation, and the posterior probability of pressure ulcer occurrence in the current state is calculated through the logistic link function, which is further converted into a basic pressure ulcer risk score. This refers to the patient's real-time risk value when no intervention is performed.

[0066] For three types of potentially modifiable latent variables—mechanical load, microenvironmental conditions, and nursing compliance—backdoor path identification is performed sequentially. Specifically, the intervention variable is set as the starting node, and the outcome variable is set as the ending node. The set of visited nodes and the current path stack are initialized. Starting from the starting node, a depth-first search is performed along all edges in both directions of the graph. Each visited node is added to the visited set to avoid repeated visits and loops. When the search path reaches the ending node, the path is determined to be a backdoor path: if there is at least one edge pointing to the starting node in the path, and there is no V-structure with the starting node as the collision point, then the path is determined to be a backdoor path. After traversing all reachable paths, all backdoor paths are summarized, and all nodes other than the intervention and outcome variables are extracted from each backdoor path. After deduplication, a set of confounding variables corresponding to the intervention variable is formed.

[0067] Based on the backdoor adjustment criterion, by integrating the distribution of confounding variables, all backdoor paths are blocked, eliminating bias caused by confounding factors. The expected outcome when the intervention variable is set to the target optimal value is calculated using the following formula: In the formula, This indicates that an intervention operation is performed on the intervention variable X. Then, the conditional expectation of the outcome variable Y is the expected value of the pressure ulcer risk score after the corresponding risk factors are adjusted to the target value; X is the modifiable risk variable to be evaluated, corresponding to modifiable latent variables such as mechanical load, microenvironmental conditions, and nursing compliance. Z represents the target optimal value of the intervention variable X, which is the standardized value of the latent variable under the clinically recommended ideal state, and is uniformly labeled based on clinical nursing guidelines and large sample healthy population data; Z is the set of confounding variables corresponding to the intervention variable X, that is, all variables located on the back door path from X to Y; Let Z be the space of all discrete values ​​of the mixed variable set; For a specific combination of values ​​of the mixed variable set Z; This indicates that when X takes the target value Furthermore, when the confounding variable set is z, the conditional expectation of the outcome variable Y is directly calculated from the fitted structural causal model; The marginal probability of the confounding variable set Z taking the value z is obtained from the variable distribution statistics based on the historical case dataset.

[0068] The average marginal causal effect size of the corresponding intervention variable is obtained by subtracting the expected outcome after intervention from the current baseline risk score. The calculation formula is: In the formula, The mean marginal causal effect of intervention variable X represents the change in pressure ulcer risk score after adjusting the factor to the target value; a negative value indicates a reduction in risk after intervention, and the larger the absolute value, the more significant the risk reduction effect of the intervention; a positive value indicates an increase in risk after intervention and has no clinical intervention value.

[0069] The risk quantification assessment module is used to integrate causal reasoning results, output a comprehensive risk score for stress injury and corresponding risk level, and calculate the causal contribution of each risk factor and sort them by proportion to identify the patient's core risk sources.

[0070] Specifically, the posterior probability p of pressure ulcer occurrence output by the structural causal model is mapped to the 0-100 score range through a linear transformation to obtain the comprehensive causal risk score S, calculated as follows: In the formula, S is the comprehensive causal risk score for pressure injury, ranging from 0 to 100. A higher score indicates a higher probability of pressure injury. p is the posterior probability of pressure injury in patients output by the structural causal model, ranging from 0 to 1.

[0071] The risk level is divided into four levels according to the scoring range, and different levels correspond to different nursing response levels: The risk level is low, and the corresponding routine nursing care response should be provided, with an assessment once daily. The risk level is medium, so we should increase rounds and nursing care, and assess patients every 12 hours. As this is a high-risk area, it requires targeted prevention and control care, with assessments conducted every 4 hours. For extremely high risk, corresponding to special level nursing care and prevention, assessment is conducted every hour. The higher the risk level, the shorter the interval between data collection and nursing rounds, achieving dynamic matching between risk level and nursing resources. The risk level threshold is determined based on the receiver operating characteristic curve analysis of pressure ulcer incidence and risk score from 10,000 historical cases. The threshold corresponding to the maximum point of Youden's index is selected and finally determined after adjustment by clinical expert consensus.

[0072] The proportion of the absolute value of the total causal effect of each risk factor to the sum of the absolute values ​​of the total causal effects of all factors is used as the causal contribution of that factor. The formula is as follows: ,in, Indicates the first The causal contribution of each risk factor reflects the degree to which that factor contributes to the current overall risk. Indicates the first The marginal causal effect of each risk factor; This indicates the total number of risk factors involved in the contribution calculation; Indicates the first The marginal causal effect of each risk factor is used to rank all risk factors from highest to lowest causal contribution, thus obtaining the current patient's core risk factor ranking, which provides a priority basis for the generation of subsequent intervention plans.

[0073] The causal visualization output module is used to generate a directional acyclic causal graph with superimposed thermodynamic mapping, which displays the risk transmission path and the contribution of each factor in an intuitive graphical form, thereby improving the interpretability of clinical decision-making.

[0074] The causal visualization output module uses the Sugiyama hierarchical layout algorithm to automatically arrange the directional acyclic causal graph: all variables are divided into three levels according to the causal transmission direction: input layer, intermediate layer and output layer. Variables in the same layer are arranged horizontally and evenly, and different layers are connected by directed edges; the algorithm automatically adjusts the horizontal position of nodes to minimize path intersections and ensure that the graph is clear and easy to read.

[0075] The node display uses a heatmap method, and the color gradation is achieved using RGB linear interpolation. 0% contribution corresponds to dark blue (RGB: 0, 0, 128), and 100% contribution corresponds to dark red (RGB: 220, 20, 60). The higher the contribution, the more red the node's color leans, enabling nurses to quickly identify core risk factors. The width of the directed edge is positively correlated with the absolute value of the standardized path coefficient of the corresponding path. The larger the path coefficient, the wider the edge, intuitively reflecting the influence intensity of different causal paths. The final causal graph can be output to clinical nursing workstations and mobile nursing terminals, supporting node clicks to view detailed path coefficients, contribution values, and other information, assisting nurses in understanding the patient's risk formation mechanism and transmission path.

[0076] The individualized intervention guidance module is used to screen clinically modifiable factors and rank them by weighted causal effect to generate individualized intervention plans containing quantitative execution parameters, providing clear execution basis for clinical nursing.

[0077] Specifically, the individualized intervention guidance module categorizes all risk factors into three types:

[0078] Short-term interventions include mechanical stress, microenvironmental conditions, and nursing compliance, which can be adjusted and improved within hours to days through clinical nursing procedures;

[0079] Medium- to long-term interventions, including metabolic status, require gradual improvement through nutritional support and treatment of underlying diseases.

[0080] Non-interventional factors include age, inherent skin condition, and other factors that cannot be changed through clinical intervention.

[0081] Intervention plans are prioritized for short-term, manageable factors. Medium- to long-term intervention factors are included in the plans as long-term management recommendations. Non-manageable factors are only given risk warnings and are not included in intervention measures.

[0082] For the identified interventionable factors, the intervention priority is calculated by weighting the causal effect and clinical feasibility. The priority calculation formula is as follows: In the formula, The priority score is the score for the i-th intervention factor; the higher the score, the higher the intervention priority. The clinical feasibility weight of the i-th intervention factor is 0-1. It is pre-calibrated based on the clinical implementation difficulty and resource consumption of each intervention measure and stored in the system configuration file. It can be adjusted by the medical institution according to its own conditions during deployment.

[0083] After sorting the intervention factors by priority scores from highest to lowest, corresponding clinical intervention measures and quantitative execution parameters are matched for each intervention factor:

[0084] In response to mechanical load factors, the frequency of body position changes, the selection and usage parameters of decompression devices, and the operating procedures for local decompression should be matched.

[0085] In response to microenvironmental factors, we tailor skin cleansing solutions, local temperature and humidity control parameters, and guidelines for the use of skin protectants.

[0086] To address nursing compliance factors, we matched the frequency of nursing performance checks, key points of staff training, and quality control measures.

[0087] Each intervention measure is labeled with the corresponding expected risk reduction, allowing nurses to intuitively understand the intervention benefits. Ultimately, an individualized intervention plan is generated, which includes intervention priority, specific operational measures, execution parameters, and expected risk reduction. This plan is then output to the nursing execution end and can be simultaneously linked to the nursing documentation system to automatically generate nursing plans, reducing the workload of nurses in document entry.

[0088] Furthermore, this system adopts an operation mode that separates offline model building from online inference calculation: In the offline state, the structural causal model building module completes model building, verification and updating based on accumulated historical case data, and the update cycle can be set to monthly or quarterly; in the online state, the other modules perform inference calculations based on real-time collected patient data and output dynamic risk assessment results and intervention plans.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart quantitative risk assessment and intervention guidance system for pressure injuries, characterized in that, The system includes: The multimodal data acquisition module is used to access multi-source heterogeneous pressure-related damage data and output standardized observation data after preprocessing the data. The structural causal model construction module is used to construct a two-layer directed acyclic causal graph containing a latent variable layer and an observed variable layer. It uses tissue vulnerability, microenvironmental conditions, mechanical load, metabolic state, and nursing compliance as latent variables, and combines clinical prior knowledge with historical case data to fit and determine the causal path structure and path coefficients. The counterfactual reasoning calculation module is used to map real-time standardized observation data to a structural causal model, identify backdoor paths through depth-first traversal, perform counterfactual reasoning based on the backdoor adjustment criterion, and calculate the marginal causal effect of each interventionable risk factor on the pressure ulcer risk score. The risk quantification and assessment module is used to integrate the results of causal reasoning, output a comprehensive risk score for stress injury and the corresponding risk level, and calculate the causal contribution of each risk factor and sort them by proportion. The causal visualization output module is used to generate a directional acyclic causal graph with superimposed thermodynamic mapping; The individualized intervention guidance module is used to screen clinically modifiable factors and rank them by causal effect weighting to generate individualized intervention plans.

2. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 1, characterized in that, The latent variable layer includes exogenous latent variables and endogenous outcome variables, while the observed variable layer consists of directly collectable clinical indicators. The exogenous latent variables include tissue vulnerability latent variables, microenvironmental condition latent variables, mechanical load latent variables, metabolic state latent variables, and nursing compliance latent variables. The endogenous outcome variable is the stress injury risk score.

3. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 1, characterized in that, The causal pathways of the directional acyclic causal graph include: the direct pathway of mechanical load to pressure injury risk, the direct pathway of tissue vulnerability to pressure injury risk, the indirect mediating pathway of metabolic state to tissue vulnerability, the indirect mediating pathway of microenvironmental conditions to tissue vulnerability, the indirect regulatory pathway of nursing compliance to mechanical load, and the indirect regulatory pathway of nursing compliance to microenvironmental conditions.

4. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 1, characterized in that, The structural causal model construction module uses the PC algorithm with prior constraints to learn the causal structure, and uses the maximum likelihood estimation method to fit the standardized path coefficients of all directed edges. The goodness of fit of the model is verified by comparing the fit index, the Tucker-Lewis index, and the root mean square index of the approximation error. The bootstrap method is used to sample and calculate the 95% confidence interval of the path coefficients, and insignificant paths with confidence intervals containing 0 are excluded.

5. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 1, characterized in that, The counterfactual reasoning module substitutes the preprocessed real-time observed variables into the fitted structural causal model, calculates the current actual values ​​of each latent variable using the regression factor scoring method, and simultaneously calculates the baseline pressure ulcer risk score under the current state. For each variable to be intervened, a depth-first search algorithm is used to traverse the directed acyclic graph to identify all backdoor paths from the intervention variable to the outcome variable, i.e., non-causal paths with the intervention variable as the starting point, the outcome variable as the ending point, and edges pointing to the intervention variable. Confounding variables on each backdoor path are extracted to form a confounding variable set for the corresponding intervention variable. Based on the backdoor adjustment criterion, the distribution of the confounding variables is integrated to block all backdoor paths, and the expected outcome when the intervention variable is set to the target optimal value is calculated.

6. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 5, characterized in that, The formula for calculating the expected outcome is as follows: ,in, Indicates the intervention variable Perform intervention operations Afterwards, the outcome variable The conditional expectation, i.e., the expected value of the pressure ulcer risk score after adjusting for the corresponding risk factors to the target value. This represents the modifiable risk variable to be evaluated; Indicating intervention variables The target optimal value is the standardized value of the latent variable under the ideal clinical recommendation. Indicating intervention variables The corresponding set of promiscuous variables, i.e., all those located in arrive Variables on the backdoor path; Represents a set of mixed variables The space of all possible values; Represents a set of mixed variables A specific combination of values; Indicates when Take target value And the mixed variable set At that time, the outcome variable The expected value of the condition; Represents a set of mixed variables Values The marginal probability.

7. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 6, characterized in that, The counterfactual reasoning calculation module subtracts the expected outcome after intervention from the current baseline risk score to obtain the marginal causal effect size. The calculation formula is: ,in, Indicating intervention variables The average marginal causal effect is the change in pressure ulcer risk score after the factor is adjusted to the target value; a negative value indicates a reduction in risk after intervention, and the larger the absolute value, the more significant the risk reduction effect of the intervention on this factor. This represents the baseline risk score for pressure ulcers in the current state, i.e., the patient's real-time risk score when no intervention has been performed.

8. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 1, characterized in that, The risk quantification and assessment module will output the posterior probability of pressure ulcer occurrence from the structural causal model. By mapping the score to the 0-100 range using a linear transformation, a comprehensive causal risk score is obtained. The calculation formula is: ,in, The comprehensive causal risk score for pressure injury ranges from 0 to 100, with a higher score indicating a higher probability of pressure injury. This represents the posterior probability of a patient experiencing stress injury, as output by the structural causal model, and its value ranges from 0 to 1. The risk quantification assessment module is divided into four risk levels according to the scoring range: Low risk Medium risk is High risk is Extremely high risk Different levels correspond to different nursing response levels.

9. The intelligent quantitative risk assessment and intervention guidance system for pressure injuries according to claim 8, characterized in that, The risk quantification and assessment module calculates the proportion of the absolute value of the total causal effect of each risk factor to the sum of the absolute values ​​of the total causal effects of all factors, as the causal contribution of that factor, i.e.: ,in, Indicates the first The causal contribution of each risk factor reflects the degree to which that factor contributes to the current overall risk. Indicates the first The marginal causal effect of each risk factor; This indicates the total number of risk factors involved in the contribution calculation; Indicates the first The marginal causal effect of each risk factor.

Images