Association analysis method for sleep disorder and cardiovascular adverse event of MINOCA patient

Abstract

The invention discloses a correlation analysis method for sleep disorder and cardiovascular adverse events of an MINOCA patient, relates to the field of clinical decision of cardiovascular diseases, and aims to perform anti-factual reasoning on data of a specific patient based on a structural causal model and simulate the correlation between sleep disorder and cardiovascular adverse events of the patient under the condition of changing one or more sleep characteristic variables. The theoretical change value of the cardiovascular adverse event risk of the patient is calculated, and the expected risk reduction benefit of intervention is calculated; and generating a decision support report, and performing visual display. According to the method, by means of time sequence causal discovery and a structural causal model, the causal relationship and effect quantification of sleep features, intermediary indexes and adverse events are defined, and a core action mechanism is revealed; a personalized intervention effect is simulated through anti-factual reasoning, and a precise risk prediction and intervention basis is provided for clinic; the visual decision support report reduces the clinical application threshold, assists in optimizing the personalized treatment scheme of the MINOCA patient, and effectively reduces the risk of cardiovascular adverse events.

Description

Technical Field

[0001] This invention relates to the field of clinical decision-making technology for cardiovascular diseases, and in particular to a method for analyzing the association between sleep disorders and adverse cardiovascular events in MINOCA patients. Background Technology

[0002] Cardiovascular disease is one of the leading causes of death worldwide. Non-obstructive coronary myocardial infarction (MINOCA), an important subtype of acute myocardial infarction, accounts for approximately 5%-14% of acute myocardial infarction cases, and its long-term risk of adverse cardiovascular events is comparable to that of obstructive myocardial infarction. Sleep disorders have been proven to be an important risk factor for cardiovascular disease, exacerbating myocardial damage through factors such as affecting autonomic nervous system function and inflammatory responses. However, existing research on the association between sleep disorders and adverse cardiovascular events in MINOCA patients has significant limitations.

[0003] Traditional research often relies on single-dimensional data or cross-sectional analysis, failing to integrate multimodal time-series data such as sleep quality, dynamic physiological monitoring, clinical events, and laboratory tests, making it difficult to capture the time-series lag relationships between variables;

[0004] Meanwhile, existing analyses mostly remain at the level of correlation, lacking the quantification of causal pathways and mediating effects. They cannot clarify the core mechanisms by which sleep disorders affect adverse cardiovascular events, nor can they provide personalized assessments of intervention benefits, resulting in a lack of precise causal evidence to support clinical decisions. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for analyzing the association between sleep disorders and adverse cardiovascular events in MINOCA patients. The technical solution adopted is as follows:

[0006] The method for analyzing the association between sleep disorders and adverse cardiovascular events in MINOCA patients includes the following steps:

[0007] Step 1: Collect multimodal time-series data of MINOCA patients. The multimodal time-series data includes sleep quality data, dynamic physiological monitoring data, discrete clinical event data, and laboratory test data. The data is then aligned and standardized according to a preset time window to form a time-series panel dataset.

[0008] Step 2: Apply the temporal causal discovery algorithm to process the temporal panel dataset, automatically learn and output a preliminary causal directed acyclic graph, which reflects the temporal lag relationship between sleep characteristics, mediating physiological indicators and adverse cardiovascular events.

[0009] Step 3: Using the preliminary causal directed acyclic graph as prior knowledge, construct a structural causal model. The model takes core sleep characteristics as treatment variables, autonomic nervous function and inflammation level indicators as mediating variables, and cardiovascular adverse event risk as outcome variables. It also quantifies the direct and indirect effects from treatment variables through various mediating variables to outcome variables.

[0010] Step 4: Based on the structural causal model, counterfactual reasoning is performed on the data of a specific patient to simulate the theoretical change in the patient's risk of adverse cardiovascular events under the condition of changing one or more sleep characteristic variables, and to calculate the expected risk reduction benefit of the intervention.

[0011] Step 5: Generate a decision support report, which includes a personalized causal path diagram, intervention simulation results, and confidence assessment results, and presents them visually.

[0012] Optionally, in step 1, the sleep quality data includes structured scores based on the Pittsburgh Sleep Quality Index and continuous sleep monitoring signals collected from wearable devices, and the dynamic physiological monitoring data includes continuously collected heart rate variability data and blood pressure circadian rhythm data; data alignment adopts a sliding time window method based on clinical event timestamps, and feature extraction is performed on continuous signals to obtain statistical feature values ​​within each time window.

[0013] Optionally, in step 2, the temporal causal discovery algorithm is the PCMCI algorithm. First, conditional independence tests are performed between variables within the time window to construct an initial causal graph skeleton. Then, the edges are oriented by analyzing the direction of lag correlation across the time window, and finally, a causal directed acyclic graph with temporal lag labels is output.

[0014] Optionally, step 3, constructing the structural causal model, includes the following sub-steps:

[0015] Step 31: Based on the preliminary causal directed acyclic graph, define a set of structural equations to form a structural causal model; the structural equations represent the outcome variable Y, each mediating variable M as a deterministic function of the treatment variable T, other antecedent variables and an independent noise term; where the treatment variable T is a binary or continuous variable of the core sleep feature;

[0016] Step 32: Using the time series panel dataset, estimate the functional form and parameters in the structural equation;

[0017] Step 33: Based on the estimated structural causal model, causal effect quantification is performed; using pre-intervention mean matching, under the condition of controlling for confounding variable C, the expected change of outcome variable Y when treatment variable T changes from reference value to target value is calculated as the total effect;

[0018] Using a mediation analysis framework and based on an estimated structural causal model, by fixing the value of the mediator variable M or simulating its distribution, we decompose the indirect effects of the treatment variable T on the outcome variable Y through the specific mediator variable M, as well as the direct effects that do not go through this mediator variable.

[0019] Optionally, in step 32, when a dual robust estimator is used, a propensity score model of the treatment variable T with respect to the confounding variable C is first established, as well as an outcome regression model of the outcome variable Y with respect to the treatment variable T and the confounding variable C. Then, based on the results of the propensity score model and the outcome regression model, the average treatment effect is calculated using an enhanced inverse probability weighting formula, which is used as an estimate of the total effect.

[0020] When using a gradient method based on neural networks, each structural equation is parameterized using neural networks; by maximizing the likelihood function of the entire SCM under the observed data and utilizing backpropagation and gradient descent algorithms, the parameters of all neural networks are optimized simultaneously to learn the nonlinear functional relationships between variables.

[0021] The neural network structure includes a gated recurrent unit layer for explicitly modeling and capturing the temporal dependence and lag effects between the treatment variable and the mediating variable.

[0022] Optionally, in step 33, the causal effect decomposition is achieved through the following core formula:

[0023] ;

[0024] in Let T be the standardized total effect of the quality variable T on the sanitary ware variable Y, and p be the total number of mediating variables. For the treatment variable T, the mediator variable is the kth mediator variable. Standardized path coefficients as a mediator variable Standardized path coefficients for outcome variable Y.

[0025] Optionally, in step 4, counterfactual reasoning specifically involves: inputting the actual observational data of the target patient into the structural causal model, then keeping the confounding variables unchanged, modifying only the value of a sleep characteristic variable to the preset target intervention value, obtaining the outcome variable value under the counterfactual conditions by running the model forward calculation, and comparing it with the outcome variable value under the original observation conditions to obtain the risk change.

[0026] Optionally, step 4 may also include a confidence fusion assessment step:

[0027] The expected benefits derived from counterfactual reasoning are combined with the statistical confidence intervals obtained through Bootstrap resampling and the verification results of the rationality of the causal path based on the clinical knowledge graph to generate a three-level confidence level: high, medium, and low.

[0028] Optionally, in step 5, the decision support report also includes a personalized intervention recommendation list. The personalized intervention recommendation list is automatically generated by enumerating common sleep intervention measures and calling the counterfactual simulation engine in step 4 to calculate and rank the expected benefits of each measure for the patient.

[0029] A correlation analysis system for sleep disorders and adverse cardiovascular events in MINOCA patients is used to implement the correlation analysis method. The correlation analysis system includes a data integration and time series alignment module, which is used to collect multimodal time series data of MINOCA patients and align and standardize them according to a preset time window to form a time series panel dataset.

[0030] The causal discovery engine module applies a time-series causal discovery algorithm to process the time-series panel dataset, automatically learns and outputs a preliminary causal directed acyclic graph; and uses the preliminary causal directed acyclic graph as prior knowledge to construct a structured causal model.

[0031] The counterfactual simulation and benefit calculation module is based on a structured causal model and performs counterfactual reasoning on data from specific patients.

[0032] The decision support and visualization report generation module is used to generate decision support reports and visualize them.

[0033] In summary, the present invention has at least one of the following beneficial technical effects:

[0034] This invention provides a method for association analysis between sleep disorders and adverse cardiovascular events in MINOCA patients. By integrating and standardizing multimodal time-series data, the comprehensiveness and accuracy of the association analysis are improved. Through time-series causal discovery and structural causal models, the causal relationships and effect quantification between sleep characteristics, mediating indicators, and adverse events are clarified, revealing the core mechanisms of action. By simulating the effects of personalized interventions through counterfactual reasoning, it provides accurate risk prediction and intervention basis for clinical practice. The visualized decision support report lowers the threshold for clinical application, helps optimize personalized treatment plans for MINOCA patients, and effectively reduces the risk of adverse cardiovascular events. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the association analysis method between sleep disorders and adverse cardiovascular events in MINOCA patients according to the present invention.

[0036] Figure 2This is a schematic diagram of temporal causal discovery and causal graph construction according to a specific embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the structural causal model and counterfactual reasoning of a specific embodiment of the present invention;

[0038] Figure 4 This is a visual illustration of a personalized decision support report according to a specific embodiment of the present invention; Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] This invention discloses a method for analyzing the association between sleep disorders and adverse cardiovascular events in MINOCA patients.

[0041] Reference Figures 1-4 Example 1, a method for analyzing the association between sleep disorders and adverse cardiovascular events in MINOCA patients, includes the following steps:

[0042] Step 1: Collect multimodal time-series data of MINOCA patients. The multimodal time-series data includes sleep quality data, dynamic physiological monitoring data, discrete clinical event data, and laboratory test data. The data is then aligned and standardized according to a preset time window to form a time-series panel dataset.

[0043] Step 2: Apply the temporal causal discovery algorithm to process the temporal panel dataset, automatically learn and output a preliminary causal directed acyclic graph, which reflects the temporal lag relationship between sleep characteristics, mediating physiological indicators and adverse cardiovascular events.

[0044] Step 3: Using the preliminary causal directed acyclic graph as prior knowledge, construct a structural causal model. The model takes core sleep characteristics as treatment variables, autonomic nervous function and inflammation level indicators as mediating variables, and cardiovascular adverse event risk as outcome variables. It also quantifies the direct and indirect effects from treatment variables through various mediating variables to outcome variables.

[0045] Step 4: Based on the structural causal model, counterfactual reasoning is performed on the data of a specific patient to simulate the theoretical change in the patient's risk of adverse cardiovascular events under the condition of changing one or more sleep characteristic variables, and to calculate the expected risk reduction benefit of the intervention.

[0046] Step 5: Generate a decision support report, which includes a personalized causal path diagram, intervention simulation results, and confidence assessment results, and presents them visually.

[0047] In Example 2, step 1, the sleep quality data includes structured scores based on the Pittsburgh Sleep Quality Index and continuous sleep monitoring signals collected from wearable devices. The dynamic physiological monitoring data includes continuously collected heart rate variability data and blood pressure circadian rhythm data. Data alignment adopts a sliding time window method based on clinical event timestamps, and feature extraction is performed on continuous signals to obtain statistical feature values ​​within each time window.

[0048] By employing the aforementioned technical solutions, sleep quality, dynamic physiological monitoring, discrete clinical events, and laboratory test data respectively correspond to the inducing factors, pathophysiological processes, and objective indicators of disease occurrence. The integration of multimodal data can comprehensively cover the entire chain of exposure-mediation-outcome, avoiding the information limitations of a single data dimension. Sliding time window alignment based on clinical event timestamps can solve the problem of temporal asynchrony in data from different sources, making the temporal correlation between variables have clear clinical significance. Statistical feature extraction and data standardization of continuous signals can eliminate dimensional differences and outlier interference, providing a high-quality, comparable dataset for subsequent causal analysis.

[0049] Traditional association analysis can only reflect the correlation between variables, but cannot distinguish the causal direction and time-series lagged effects. In contrast, time-series causal discovery algorithms eliminate spurious associations through prior conditional independence tests, construct a potential causal framework among variables, and then use lagged correlation analysis across time windows to determine the causal direction, ultimately outputting a causal graph with time-series labels. This process can effectively uncover the potential causal effects of sleep characteristics on adverse cardiovascular events, while identifying the core roles of mediating variables such as autonomic nervous system function and inflammation levels, providing a reliable causal structure prior for subsequent model construction.

[0050] Based on the previously obtained causal diagram, the roles of core variables were clearly defined, consistent with the pathophysiological mechanism by which sleep disorders affect cardiovascular health. Sleep disorders can directly affect cardiovascular function and can also exacerbate cardiovascular damage through indirect pathways such as disrupting autonomic nerve rhythms and increasing inflammation levels. By defining structural equation modeling to quantify the functional relationships between variables, the direct and indirect effects of the therapeutic variable (sleep characteristics) on the outcome variable (adverse event risk) can be separated, and the contribution weights of each mediating pathway can be clarified, providing a target basis for precise intervention.

[0051] Counterfactual reasoning, by fixing the values ​​of confounding variables and changing only the target sleep characteristic variable, simulates changes in outcomes after intervention. This effectively avoids causal confounding caused by confounding factors in traditional observational studies and accurately assesses the potential benefits of personalized interventions. Its core logic is based on real individual clinical data to construct a virtual scenario of "how the risk of adverse events would change if sleep status were altered," reflecting the impact of individual differences on intervention effects and achieving a transformation from group-level analysis to individual-level decision-making.

[0052] Transforming complex causal pathways, quantified effect results, and confidence assessments into visual content can lower the barrier for clinicians to understand technical analysis results. Personalized causal pathway diagrams intuitively present patient-specific risk transmission mechanisms, intervention simulation results provide clear benefit expectations, and confidence assessments ensure the reliability of the results. The decision support report formed by combining these three elements can directly provide a scientific basis for clinical sleep intervention programs, realizing the translation of research results into clinical practice.

[0053] In Example 3, step 2, the temporal causal discovery algorithm is the PCMCI algorithm. First, conditional independence is tested among variables within the time window to construct the initial causal graph skeleton. Then, the edges are oriented by analyzing the lag correlation direction across the time window, and finally, a causal directed acyclic graph with temporal lag labels is output.

[0054] By employing the aforementioned technical solution, multimodal time-series data encompasses multiple dimensions of variables, including sleep quality, physiological monitoring, and clinical events. Spurious associations may exist between these variables (e.g., they may be jointly influenced by other confounding factors). The PCMCI algorithm first performs a conditional independence test on any two variables within each time window. By controlling for interference from other latent variables, it determines whether the association between variables is a direct dependency. After eliminating spurious associations, it retains the truly directly dependent variable pairs, forming an undirected initial causal graph skeleton, laying the foundation for subsequently clarifying the causal direction.

[0055] Lagged Correlation Analysis Across Time Windows: A key characteristic of time series data is that variables have a time lag effect; for example, the impact of sleep disorders on inflammation levels only becomes apparent after a certain period. The PCMCI algorithm analyzes the lagged correlations of variables across time windows to determine the direction of causal effects: If the change in the value of variable A in time window t can significantly explain the change in the value of variable B in the subsequent time window t+k (where k is the lag step size), and the inverse correlation is not significant, then the undirected edge between A and B in the initial skeleton is directed to A→B, and the lag step size k is marked.

[0056] By integrating the oriented causal edges with the initial skeleton, a causal directed acyclic graph is ultimately formed. This graph not only clearly presents the direct causal relationship between sleep characteristics, mediating physiological indicators, and adverse cardiovascular events, but also reflects the time span of the effects through time lag markers, providing accurate causal structure priors for subsequent structural causal models and avoiding the problems of confusion regarding causal direction and time-series effects in traditional association analysis.

[0057] Example 4, step 3, constructing the structural causal model includes the following sub-steps:

[0058] Step 31: Based on the preliminary causal directed acyclic graph, define a set of structural equations to form a structural causal model; the structural equations represent the outcome variable Y, each mediating variable M as a deterministic function of the treatment variable T, other antecedent variables and an independent noise term; where the treatment variable T is a binary or continuous variable of the core sleep feature;

[0059] Step 32: Using the time series panel dataset, estimate the functional form and parameters in the structural equation;

[0060] Step 33: Based on the estimated structural causal model, causal effect quantification is performed; using pre-intervention mean matching, under the condition of controlling for confounding variable C, the expected change of outcome variable Y when treatment variable T changes from reference value to target value is calculated as the total effect;

[0061] Using a mediation analysis framework and based on an estimated structural causal model, by fixing the value of the mediator variable M or simulating its distribution, we decompose the indirect effects of the treatment variable T on the outcome variable Y through the specific mediator variable M, as well as the direct effects that do not go through this mediator variable.

[0062] In Example 5, in step 32, when using a dual robust estimator, a propensity score model of the treatment variable T with respect to the confounding variable C and an outcome regression model of the outcome variable Y with respect to the treatment variable T and the confounding variable C are first established respectively. Then, based on the results of the propensity score model and the outcome regression model, the average treatment effect is calculated using an enhanced inverse probability weighting formula, which is used as an estimate of the total effect.

[0063] When using a gradient method based on neural networks, each structural equation is parameterized using neural networks; by maximizing the likelihood function of the entire SCM under the observed data and utilizing backpropagation and gradient descent algorithms, the parameters of all neural networks are optimized simultaneously to learn the nonlinear functional relationships between variables.

[0064] The neural network structure includes a gated recurrent unit layer for explicitly modeling and capturing the temporal dependence and lag effects between the treatment variable and the mediating variable.

[0065] In Example 6, step 33, the causal effect decomposition is achieved through the following core formula:

[0066] ;

[0067] in Let T be the standardized total effect of the quality variable T on the sanitary ware variable Y, and p be the total number of mediating variables. For the treatment variable T, the mediator variable is the kth mediator variable. Standardized path coefficients as a mediator variable Standardized path coefficients for outcome variable Y.

[0068] By employing the aforementioned technical approach, a preliminary causal directed acyclic graph (DAG) has clearly defined the causal relationship structure of sleep characteristics: the treatment variable T, the mediating variable M, the risk of adverse cardiovascular events (outcome variable Y), and the confounding variable C. The structural equation model expresses Y and each M as a function of T, antecedent variables (such as age, underlying diseases, and other confounding variables C), and an independent noise term. This approach aligns with the physiological logic of sleep disorders affecting cardiovascular health while also eliminating interference from unobserved variables through the independent noise term. The treatment variable T is set as a binary or continuous variable, allowing for flexible adaptation to the classification assessment or continuous quantification of sleep quality, meeting the analytical needs of different research scenarios.

[0069] The dual-robust estimator achieves double-protection effect estimation by simultaneously constructing a propensity score model and an outcome regression model. The propensity score model quantifies the effect of the confounding variable C on the treatment variable T, while the outcome regression model characterizes the combined effect of T and C on the outcome variable Y. The results from both models are then fused using an enhanced inverse probability weighting formula to calculate the average treatment effect. The advantage of this method is that as long as one model is accurately specified, a robust estimate of the total effect can be obtained, effectively avoiding errors caused by biases in the specification of a single model.

[0070] The neural network gradient method leverages the nonlinear fitting capability of neural networks to parameterize each structural equation, adapting to complex nonlinear relationships between variables. By maximizing the likelihood function of the SCM under observed data, it ensures the model's fit to the actual data. Using backpropagation and gradient descent algorithms, it simultaneously optimizes all neural network parameters, efficiently learning the relationships between variables. The introduction of gated recurrent unit layers can specifically capture the temporal dependence and lag effects between therapeutic and mediating variables, aligning with the characteristics of multimodal time-series data (e.g., the impact of sleep disorders on inflammation levels needs time to manifest).

[0071] Pre-intervention mean matching ensures comparability of the two compared groups (with reference and target values ​​for the treatment variable T) in terms of baseline characteristics by controlling the value of the confounding variable C, thus eliminating the interference of confounding factors. The expected change in the outcome variable Y calculated at this point represents the pure causal total effect of T on Y, avoiding the causal confounding problems in traditional correlation analysis.

[0072] Based on the mediation analysis framework, the effects of specific mediation paths are isolated by fixing the value of the mediating variable M or simulating its distribution. The direct effect reflects the strength of T's influence on Y without going through M directly, while the indirect effect reflects the strength of T's influence on Y via M; both constitute the total effect. In the core formula, the total effect TE is calculated as the sum of the products of the path coefficients of each mediating variable, where... Quantify T for the k-th mediator variable The intensity of the impact, Quantify The product of the effects on Y represents the contribution of a single mediating path, while the sum fully reflects the combined effect of all mediating paths, clarifying the core transmission mechanism by which sleep characteristics influence adverse events.

[0073] In Example 7, step 4, counterfactual reasoning specifically involves: inputting the actual observation data of the target patient into the structural causal model, keeping the confounding variables unchanged, modifying only the value of a sleep characteristic variable to the preset target intervention value, obtaining the outcome variable value under counterfactual conditions by running the model forward calculation, and comparing it with the outcome variable value under the original observation conditions to obtain the risk change.

[0074] Example 8, step 4, also includes a confidence fusion evaluation step:

[0075] The expected benefits derived from counterfactual reasoning are combined with the statistical confidence intervals obtained through Bootstrap resampling and the verification results of the rationality of the causal path based on the clinical knowledge graph to generate a three-level confidence level: high, medium, and low.

[0076] By adopting the above technical solution, counterfactual reasoning first inputs the actual observation data of the target patient (including sleep characteristics, confounding variables, mediating variables and baseline outcome risk) into the constructed structured causal model to ensure that the initial state of the model is consistent with the patient's actual clinical situation.

[0077] Keep the values ​​of confounding variables (such as age, underlying diseases, laboratory test indicators, etc.) unchanged to avoid them interfering with the causal relationship between sleep characteristics and outcome variables. Only modify the target sleep characteristic variables (such as PSQI score, sleep time, etc.) to the preset intervention values.

[0078] By using forward modeling, the theoretical values ​​of mediating variables (autonomic nervous system function, inflammation level, etc.) and outcome variables (risk of adverse cardiovascular events) under this intervention condition are simulated, i.e., the counterfactual outcome.

[0079] By calculating the difference between the counterfactual outcome and the outcome variable value under the original observation conditions, the change in the risk of adverse cardiovascular events in patients after intervention for a single sleep feature is obtained, which intuitively reflects the potential benefits of personalized intervention.

[0080] The confidence fusion assessment is based on the Bootstrap resampling method, which repeatedly samples the time-series panel dataset and reconstructs the model, calculating the risk change of each counterfactual inference result to obtain the statistical confidence interval. The narrower the confidence interval, the stronger the statistical stability of the results and the smaller the impact of sampling error on the results.

[0081] Based on a clinical knowledge graph, we verify whether the causal pathways upon which counterfactual reasoning relies are consistent with the pathophysiological mechanisms of cardiovascular diseases and the conclusions of existing clinical research. If the causal pathways are consistent with clinical consensus, they are deemed to have clinical rationality; otherwise, their confidence weights are reduced.

[0082] By combining the results of the two assessments above, and combining the width of the statistical confidence interval with the level of clinical pathway rationality, a three-tier confidence level (high, medium, and low) is generated. High confidence indicates that the results have both statistical stability and clinical rationality; medium confidence indicates that one of them has a slight deficiency; and low confidence indicates poor statistical stability or questionable clinical pathway, providing a reliable reference for clinical decision-making.

[0083] In Example 9, step 5, the decision support report also includes a personalized intervention recommendation list. The personalized intervention recommendation list is automatically generated by enumerating common sleep intervention measures and calling the counterfactual simulation engine in step 4 to calculate and rank the expected benefits of each measure for the patient.

[0084] Using the above-mentioned technical approach, the selection criteria for common sleep interventions are as follows: all listed interventions are derived from clinically validated sleep improvement programs, covering categories such as sleep schedule adjustment, environmental optimization, behavioral intervention, and medication. The selection criteria take into account the pathophysiological characteristics of MINOCA patients, prioritizing measures with minimal impact on autonomic nervous system function and inflammation levels, and excluding interventions that may increase cardiovascular burden.

[0085] Each type of intervention is transformed into a corresponding change in sleep characteristic variables within a structured causal model. For example, a fixed sleep schedule corresponds to the optimization of sleep rhythm stability indicators, pre-sleep relaxation training corresponds to the shortening of sleep latency, and controlling sleep time corresponds to the adjustment of actual sleep time. By clearly defining the variable mapping relationships, the interventions can be identified and quantified by the counterfactual simulation engine.

[0086] Counterfactual simulation engine usage and benefit calculation: Based on the target patient's baseline data (sleep characteristics, physiological indicators, clinical information), the changes in variables corresponding to each intervention are input into the counterfactual simulation engine. The engine, based on a pre-built structured causal model, simulates the theoretical reduction in the risk of adverse cardiovascular events for each intervention; this reduction represents the expected benefit of the intervention. During the calculation, confounding variables are kept constant, focusing only on the independent impact of changes in sleep characteristics on the outcome to ensure the accuracy of the benefit assessment.

[0087] Personalized ranking and list generation: Interventions are ranked in descending order of their expected benefits, prioritizing those with the highest benefits. The ranking is then fine-tuned based on feasibility and safety, ultimately generating a personalized intervention recommendation list. The list clearly presents the expected risk reduction, key operational points, and confidence level for each measure, providing direct reference for clinicians to develop targeted sleep intervention plans and achieving a closed loop from risk assessment to intervention implementation.

[0088] Example 10: A correlation analysis system for sleep disorders and adverse cardiovascular events in MINOCA patients, used to implement the correlation analysis method. The correlation analysis system includes: a data integration and time series alignment module, used to collect multimodal time series data of MINOCA patients and align and standardize them according to a preset time window to form a time series panel dataset;

[0089] The causal discovery engine module applies a time-series causal discovery algorithm to process the time-series panel dataset, automatically learns and outputs a preliminary causal directed acyclic graph; and uses the preliminary causal directed acyclic graph as prior knowledge to construct a structured causal model.

[0090] The counterfactual simulation and benefit calculation module is based on a structured causal model and performs counterfactual reasoning on data from specific patients.

[0091] The decision support and visualization report generation module is used to generate decision support reports and visualize them.

[0092] The following describes the implementation principle of the present invention using specific embodiments:

[0093] Using MINOCA patients admitted to the cardiology department of a tertiary hospital as the research subjects, a complete association analysis of sleep disorders and adverse cardiovascular events in MINOCA patients was conducted. The specific process is as follows:

[0094] Multimodal time series data acquisition and standardization:

[0095] Three hundred MINOCA patients meeting the diagnostic criteria were selected, and multimodal time-series data were continuously collected for two years after enrollment:

[0096] Sleep quality data: A questionnaire survey is conducted quarterly using the Pittsburgh Sleep Quality Index to obtain structured scores in 7 dimensions, including subjective sleep quality and sleep latency; at the same time, patients are equipped with wearable sleep monitoring devices to continuously collect signals such as nighttime sleep duration, sleep cycle, and number of awakenings.

[0097] Dynamic physiological monitoring data: Using a portable physiological monitor, the patient's heart rate variability index and blood pressure diurnal rhythm data are continuously collected, and the heart rate, systolic blood pressure and diastolic blood pressure values ​​are recorded every 5 minutes within 24 hours.

[0098] Discrete clinical event data: The occurrence time and diagnosis results of adverse cardiovascular events such as recurrence of acute myocardial infarction, cerebral infarction, and heart failure after patient enrollment are recorded through the hospital's electronic medical record system.

[0099] Laboratory test data: Blood routine, inflammatory factors, liver and kidney function and other indicators are collected from patients every six months, including white blood cell count, neutrophil percentage, C-reactive protein, homocysteine ​​and other results.

[0100] Data processing employed a sliding time window method based on clinical event timestamps, setting a 3-month time window and aligning data across dimensions according to the time window. Statistical features such as mean, standard deviation, and peak value were extracted from continuous signals collected by wearable devices and physiological monitors within each time window. All data were standardized to eliminate dimensional differences and outlier interference, ultimately forming a time-series panel dataset.

[0101] Temporal causal discovery and causal graph construction:

[0102] The PCMCI algorithm is used to process the time series panel dataset:

[0103] First, within each time window, conditional independence tests are performed on all variables, including sleep characteristics, autonomic nervous system function indicators, inflammation level indicators, and adverse cardiovascular events. The interference of potential confounding variables such as age, gender, and underlying diseases is controlled to eliminate spurious associations and retain the truly directly dependent variable pairs, thus constructing an undirected initial causal graph skeleton.

[0104] Subsequently, the direction of the lag correlation across time windows is analyzed. For example, it is determined whether the change in the value of insufficient sleep duration in time window t can significantly explain the increase in C-reactive protein in the subsequent time window t+3. If the inverse correlation is not significant, the corresponding undirected edge is directed to insufficient sleep duration → increased C-reactive protein, and the lag is marked by 3 steps.

[0105] The final output is a causal directed acyclic graph with temporal lag markers, which clearly shows the direct causal relationship and temporal action span between core sleep features and autonomic nervous function, inflammatory level indicators, and adverse cardiovascular events.

[0106] Structural causal model construction and effect quantification:

[0107] Structural equation modeling (SEM) definition: Based on a preliminary causal directed acyclic graph, a structural equation model is defined. The comprehensive sleep quality score is used as the treatment variable T (a continuous variable, ranging from 0 to 21 points). Autonomic nervous system function indicators (heart rate variability coefficient) and inflammatory level indicators (C-reactive protein, tumor necrosis factor α) are used as mediating variables M1, M2, and M3, respectively. The risk of adverse cardiovascular events is used as the outcome variable Y. Age, history of hypertension, and history of diabetes are used as confounding variables C. The structural equation model expresses Y, M1, M2, and M3 as functions of T, C, and independent noise terms, respectively.

[0108] Parameter estimation: A dual robust estimator is used for parameter estimation. First, a propensity score model of the treatment variable T with respect to the confounding variable C is established, and at the same time, an outcome regression model of the outcome variable Y with respect to T and C is established. Then, the results of the two models are fused by an enhanced inverse probability weighting formula to obtain the total effect estimate. For variable pairs with nonlinear associations, a neural network with a gated recurrent unit layer is used for parameterization to maximize the likelihood function of the model under the observed data and simultaneously optimize the parameters to capture time dependence and lagged effects.

[0109] Causal effect quantification: By controlling for confounding variables C with mean matching before intervention, the expected change in outcome variable Y when treatment variable T drops from 12 (sleep disorder) to 8 (normal sleep) is calculated as the total effect; using a mediation analysis framework, with fixed values ​​of M1, M2, and M3, the indirect effects of T on Y through each mediating variable, as well as the direct effects without mediating variables, are decomposed, and the standardized total effect is calculated using the core formula.

[0110] Figure 2 This demonstrates the process of constructing a causal directed acyclic graph using the PCMCI algorithm. The left side shows the construction process of the initial causal graph skeleton, eliminating spurious associations through conditional independence tests; the right side shows cross-time window lag correlation analysis to determine the causal direction; and the bottom shows the final causal directed acyclic graph with time lag labels, clearly presenting the time lag relationship between sleep characteristics, mediating physiological indicators, and adverse cardiovascular events.

[0111] Counterfactual reasoning and confidence assessment:

[0112] Counterfactual reasoning: A 56-year-old male patient with MINOCA was selected. His baseline sleep quality score was 14, heart rate variability coefficient was 0.15, C-reactive protein was 8 mg / L, and he had no history of hypertension or diabetes. The patient's actual observational data were input into a structural causal model, keeping the values ​​of confounding variables unchanged. Only the sleep quality score was modified to the target intervention value of 9. The risk of adverse cardiovascular events under the counterfactual conditions was calculated through forward calculation of the model, and the change in risk was obtained by comparing it with the original risk value.

[0113] Figure 3This demonstrates the process of constructing a structural causal model and counterfactual reasoning. The left side shows the definition of the structural equations and parameter estimation methods for the structural causal model; the middle shows the counterfactual reasoning process, simulating the impact of changes in sleep characteristics on the risk of adverse cardiovascular events; the right side shows the core content of the decision support report, including a list of personalized intervention recommendations and benefit assessments.

[0114] Confidence fusion assessment: The Bootstrap resampling method was used to perform 1000 repeated samplings on the time-series panel dataset and reconstruct the model. The risk change for each counterfactual inference was calculated to obtain a 95% statistical confidence interval. Based on the cardiovascular disease clinical knowledge graph, the consistency between the causal path of "improved sleep quality → increased heart rate variability → reduced risk of adverse cardiovascular events" and the clinical pathophysiological mechanism was verified. Considering the width of the statistical confidence interval and the rationality of the clinical path, the confidence level of the patient's counterfactual inference result was determined to be high.

[0115] Decision support report generation and visualization:

[0116] A personalized decision support report was generated for this patient, containing three core components:

[0117] Personalized causal path diagram: Visualizing the causal associations and time-series lag relationships between sleep quality scores, heart rate variability coefficient, C-reactive protein and the risk of adverse cardiovascular events;

[0118] Intervention simulation results: The reduction in cardiovascular adverse event risk and the corresponding confidence level were clearly indicated after the sleep quality score improved from 14 to 9.

[0119] Personalized intervention recommendation list: Enumerate clinically validated sleep intervention measures, use a counterfactual simulation engine to calculate and rank the expected benefits of each measure, and prioritize the following three measures: "Fixed sleep schedule (go to bed at 22:00 and wake up at 6:30)", "Pre-sleep relaxation training (15 minutes of meditation)" and "Optimization of sleep environment (keep the bedroom quiet and dimly lit)". Each measure is marked with its expected risk reduction, key points of operation and confidence level.

[0120] Implementation of the correlation analysis system:

[0121] This analysis was performed using a dedicated system, which comprises four main functional modules:

[0122] Data integration and time series alignment module: Connects to hospital electronic medical record system, wearable device data platform and laboratory testing system, automatically collects multimodal data, completes alignment and standardization according to 3-month time windows, and outputs time series panel dataset;

[0123] Causal discovery engine module: Built-in PCMCI algorithm, automatically processes time series panel datasets, outputs causal directed acyclic graph with time series lag labels, builds structural causal models based on the graph and completes parameter estimation and effect quantification;

[0124] Counterfactual simulation and benefit calculation module: Supports manual input or automatic reading of individual patient data, provides an interface for modifying sleep characteristic variables, calculates risk changes under counterfactual conditions with one click, and automatically completes confidence fusion assessment;

[0125] Decision support and visualization report generation module: Automatically generates decision support reports that include personalized causal path diagrams, intervention simulation results, confidence assessments, and intervention recommendation lists, and supports PDF export and sharing with the hospital's internal system.

[0126] Figure 4 This visualization interface for generating a decision support report for a single MINOCA patient employs a regional grid layout to integrate core information. The patient's basic information area displays information such as gender, age, baseline sleep quality score, follow-up duration, and underlying diseases. The confidence assessment area marks the confidence level as high, explaining that the assessment is based on a narrow statistical confidence interval and that the causal path conforms to the clinical mechanism. The personalized causal path diagram presents the patient's two core causal paths with a simplified node and line segment structure, annotating the effect contribution value of each path. The personalized intervention recommendation list area arranges three sleep intervention measures in descending order of expected benefit, with each measure indicating the expected risk reduction and key operational points. The overall interface layout is clearly hierarchical, and the information is presented intuitively, facilitating clinicians' rapid access to personalized decision-making support.

[0127] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method of analyzing the association of sleep disorders in MINOCA patients with cardiovascular adverse events, characterized in that, The method comprises the following steps: Step 1, collecting multi-modal time series data of MINOCA patients, the multi-modal time series data comprising sleep quality data, dynamic physiological monitoring data, discrete clinical event data and laboratory examination data, and aligning and standardizing the data according to a preset time window to form a time series panel data set; Step 2, applying a time series causal discovery algorithm to process the time series panel data set, automatically learning and outputting a preliminary causal directed acyclic graph, the preliminary causal directed acyclic graph reflecting time series lag relationships between sleep characteristics, intermediary physiological indicators and adverse cardiovascular events; Step 3, constructing a structural causal model based on the preliminary causal directed acyclic graph, the model taking core sleep characteristics as a treatment variable, taking autonomic nervous function and inflammation level indicators as intermediary variables, taking cardiovascular adverse event risk as an outcome variable, and quantifying direct and indirect effects from the treatment variable to the outcome variable through the intermediary variables; Step 4, based on the structural causal model, performing counterfactual reasoning on the data of a specific patient, simulating the theoretical change value of the cardiovascular adverse event risk of the patient under the condition of changing one or more sleep characteristic variables, and calculating the expected risk reduction benefit of the intervention; Step 5, generating a decision support report, the decision support report comprising an individualized causal path diagram, intervention simulation results and confidence evaluation results, and being visually displayed.

2. The MINOCA patient sleep disorder and cardiovascular adverse event association analysis method according to claim 1, characterized in that, In step 1, the sleep quality data comprises structured scores based on the Pittsburgh Sleep Quality Index and continuous sleep monitoring signals collected from wearable devices, and the dynamic physiological monitoring data comprises continuously collected heart rate variability data and blood pressure circadian rhythm data; The data alignment adopts a sliding time window method based on a clinical event timestamp, and the continuous signals are feature extracted to obtain statistical feature values in each time window.

3. The MINOCA patient sleep disorder and cardiovascular adverse event association analysis method according to claim 2, characterized in that, In step 2, the time series causal discovery algorithm is a PCMCI algorithm, which first performs conditional independence test between variables in a time window to construct an initial causal graph skeleton, then analyzes the lag correlation direction across time windows to orient the edges, and finally outputs a causal directed acyclic graph with time series lag markers.

4. The MINOCA patient sleep disorder and cardiovascular adverse event association analysis method according to claim 3, characterized in that, In step 3, constructing a structural causal model comprises the following sub-steps: Step 31, based on the preliminary causal directed acyclic graph, defining a set of structural equations to constitute a structural causal model; the structural equation represents the outcome variable Y and each intermediary variable M as a deterministic function of the treatment variable T, other preceding variables and an independent noise term; wherein the treatment variable T is a binary or continuous variable of the core sleep characteristic; Step 32, using the time series panel data set, estimating the functional form and parameters in the structural equation; Step 33, based on the estimated structural causal model, quantifying causal effects; using mean value matching before intervention, under the condition of controlling confounding variables C, calculating the expected change of the outcome variable Y when the treatment variable T changes from a reference value to a target value as the total effect; Using the mediation analysis framework, based on the estimated structural causal model, by fixing the value of the mediator variable M or simulating its distribution, the indirect effect of the treatment variable T on the outcome variable Y through a specific mediator variable M is decomposed, as well as the direct effect without the mediation of the mediator variable.

5. The method of claim 4, wherein the MINOCA patient is a female. In step 32, when using the double robust estimator, first, the propensity score model of the treatment variable T with respect to the confounding variable C is established, and the outcome regression model of the outcome variable Y with respect to the treatment variable T and the confounding variable C is established; then, based on the results of the propensity score model and the outcome regression model, the average treatment effect is calculated by the augmented inverse probability weighting formula, which is taken as the estimated value of the total effect; When using the neural network-based gradient method, a neural network is used to parameterize each structural equation; by maximizing the likelihood function of the entire SCM under the observed data, and using the back propagation and gradient descent algorithm, the parameters of all neural networks are optimized to learn the nonlinear functional relationship between variables; The structure of the neural network includes a gated recurrent unit layer for explicitly modeling and capturing the time-dependent and lag effects between the treatment variable and the mediator variable.

6. The method of claim 5, wherein the MINOCA patient is a female. In step 33, the causal effect decomposition is realized by the following core formula: ; wherein is the standardized total effect of the high quality variable T on the fixture variable Y, p is the total number of mediator variables, is the standardized path coefficient of the treatment variable T on the kth mediator variable , is the standardized path coefficient of the mediator variable on the outcome variable Y.

7. The method of claim 6, wherein the MINOCA patient is a female. In step 4, counterfactual reasoning is specifically: in the structural causal model, the actual observed data of the target patient is input, then the confounding variable is kept unchanged, only the value of one sleep characteristic variable is modified to the preset target intervention value, the outcome variable value under the counterfactual condition is obtained by running the model forward calculation, and compared with the outcome variable value under the original observation condition, the risk change amount is obtained.

8. The MINOCA patient sleep disorder and cardiovascular adverse event association analysis method according to claim 7, step 4 further comprises a confidence fusion evaluation step: The expected benefit obtained by counterfactual reasoning is combined with the statistical confidence interval obtained by Bootstrap resampling, and the verification result of the rationality of the causal path based on the clinical knowledge graph to perform fusion evaluation, and generate high, medium and low three grades of confidence levels.

9. The MINOCA patient sleep disorder and cardiovascular adverse event association analysis method according to claim 8, in step 5, the decision support report further comprises a personalized intervention recommendation list, which is automatically generated by enumerating common sleep intervention measures and calling the counterfactual simulation engine of step 4 to calculate the expected benefit of each measure for the patient.

10. A system for performing a correlation analysis of sleep disorders in MINOCA patients with cardiovascular adverse events, for implementing the correlation analysis method as claimed in claims 1-9, characterized in that, The association analysis system comprises a data integration and time series alignment module for collecting multi-modal time series data of MINOCA patients and aligning and standardizing the data according to a preset time window to form a time series panel data set; The causal discovery engine module applies a time series causal discovery algorithm to process the time series panel data set, automatically learns and outputs a preliminary causal directed acyclic graph; based on the preliminary causal directed acyclic graph as prior knowledge, a structural causal model is constructed; The counterfactual simulation and benefit calculation module performs counterfactual reasoning based on the structural causal model for the data of a specific patient; The decision support and visualization report generation module is used to generate a decision support report and perform visual display.

Images