SAH post-delayed cerebral ischemia risk feature robust screening method based on mixed causal structure learning and related equipment

Abstract

The invention provides a hybrid causal structure learning-based delayed cerebral ischemia risk feature robust screening method after SAH and related equipment, and belongs to the technical field of neurological intensive care data analysis. The method comprises the following steps: performing time sequence alignment and multi-modal preprocessing on SAH clinical data to obtain preprocessed data; establishing a mixed causal diagram based on the medical hard constraint matrix; outputting a minimum confusion variable adjustment set; performing consistency verification on the average causal effect value output by each estimator, and if the verification is passed, outputting the fused average causal effect value and the confidence interval thereof; performing multi-dimensional reverse test on the fused average causal effect value, and calculating a clinical causal robustness score; and grading and outputting the candidate risk factors corresponding to the processing variables as confirmed causal factors, potential causal factors or false correlation factors. According to the method, the conflict problem of medical prior and data driving in causal graph construction can be solved, and the interpretability and credibility of the result are improved.

Description

Technical Field

[0001] This application relates to the field of neurocritical care medicine data analysis technology, and in particular to a robust screening method and related equipment for risk features of delayed cerebral ischemia after SAH based on mixed causal structure learning. Background Technology

[0002] In related technologies, the pathophysiological mechanism of DCI is highly complex, involving multiple processes such as large vessel spasm, microcirculatory disturbance, cortical diffuse depolarization, inflammatory response, and blood-brain barrier disruption. Various clinical scores, imaging indicators, biochemical markers, and treatment measures are intertwined in multiple ways. This multifactorial and multi-pathway complexity renders the predictive model based on "single risk factor + regression analysis" unreliable in its characteristic basis and has poor generalization ability.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this application is to propose a robust screening method and related equipment for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning. This method can resolve the conflict between medical priors and data-driven approaches in causal graph construction, and achieve accurate identification from "spurious correlations" to "true causality".

[0005] To achieve the above objectives, one aspect of this application proposes a robust screening method for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning, the method comprising the following steps: Obtain SAH clinical data; The SAH clinical data were subjected to time-series alignment and multimodal preprocessing to obtain preprocessed data; At least one candidate risk factor to be analyzed is set as a treatment variable; Based on the preprocessed data and the processing variables, a hybrid causal graph based on a medical hard constraint matrix is ​​established; Based on the hybrid causal graph, causal effect identification is performed to determine whether the causal relationship between the treatment variable and the outcome variable is identifiable, and the minimum confounding variable adjustment set is output. Based on the minimum confounding variable adjustment set, an integrated strategy is used to numerically estimate the causal effect between the treatment variable and the outcome variable, and at least two types of heterogeneous estimators are run in parallel to calculate the average causal effect value output by each estimator. The average causal effect value output by each estimator is checked for consistency. If the check passes, the fused average causal effect value and its confidence interval are output. A multidimensional refutation test was performed on the fused average causal effect value, and a clinical causal robustness score was calculated based on the results of the multidimensional refutation test. Based on the confidence interval of the fused average causal effect value and the clinical causal robustness score, the candidate risk factors corresponding to the treatment variables are classified and output as confirmatory causal factors, potential causal factors, or spurious correlation factors.

[0006] In some embodiments, the step of performing time-series alignment and multimodal preprocessing on the SAH clinical data to obtain preprocessed data includes: Using the onset time of SAH as the timeline, a time axis was established, and all variables in the SAH clinical data were divided into a set of baseline variables, a set of intermediate process variables, and an outcome variable; the outcome variable included the occurrence status of delayed cerebral ischemia. Missing value processing, outlier cleaning, and standardized coding are performed on the data in the baseline variable set, the intermediate process variable set, and the outcome variable to obtain preprocessed data.

[0007] In some embodiments, setting at least one candidate risk factor to be analyzed as a treatment variable includes: At least one variable is selected from the baseline variable set or the intermediate process variable set as a treatment variable; the selected treatment variable is the candidate risk factor to be analyzed.

[0008] In some embodiments, establishing a hybrid causal graph based on a medical hard constraint matrix according to the preprocessed data and the processing variables includes: Based on the preprocessed data and prior medical knowledge, a medical hard constraint matrix is ​​defined. In the medical hard constraint matrix, by setting specific values ​​as prohibition symbols, prohibition rules for connecting variables are defined. The prohibition rules include at least: prohibiting the outcome variable from pointing to any variable in the baseline variable set. Based on the aforementioned medical hard constraint matrix, a constraint-based causal discovery algorithm is used to perform a conditional independence test on the preprocessed data, generating an initial directed acyclic graph. Edges in the initial directed acyclic graph that conflict with weak medical prior knowledge are marked as edges to be adjudicated. For each edge to be decided, calculate the change in the Bayesian information criterion value of the model as a whole when the edge is retained and removed in the initial directed acyclic graph. When the absolute value of the change exceeds the preset resolution threshold, the edge to be decided is retained; otherwise, the edge to be decided is removed, and the result of conflict resolution is obtained. Based on the conflict resolution results, the initial directed acyclic graph is adjusted to obtain a hybrid causal graph based on a medical hard constraint matrix.

[0009] In some embodiments, the step of identifying causal effects based on the mixed causal graph, determining whether the causal relationship between the treatment variable and the outcome variable is identifiable, and outputting a minimum confounding variable adjustment set includes: On the topology of the hybrid causal graph, based on the backdoor criterion or the frontdoor criterion, all connection paths are traversed and analyzed to determine whether there is a set of observed variables; the set of observed variables is used to block all non-causal paths from the processing variable to the outcome variable; If the set of observed variables does not exist, it is determined to be unidentifiable, and the subsequent analysis of the current causal relationship is terminated; If the set of observed variables exists, the current causal relationship is determined to be identifiable. Then, nodes identified as colliding variables and mediating variables by preset rules are excluded from the set of observed variables to obtain the minimum confusion variable adjustment set.

[0010] In some embodiments, the step of numerically estimating the causal effect between the treatment variable and the outcome variable using an integrated strategy based on the minimum confounding variable adjustment set, running at least two types of heterogeneous estimators in parallel, and calculating the average causal effect value output by each estimator includes: Based on the adjusted set of minimum confounding variables, an ensemble strategy is used to numerically estimate causal effects and train a heterogeneity estimator. The heterogeneity estimator includes, but is not limited to, a parametric estimator based on an outcome model, a nonparametric estimator based on a treatment allocation model, and a dual robust estimator. The parametric estimator based on the outcome model is used to fit the conditional probability distribution of the outcome variable given the treatment variable and the adjusted set of minimum confounding variables. The nonparametric estimator based on the treatment allocation model is used to estimate the propensity score of the treatment variable given the adjusted set of minimum confounding variables, and to perform weighted or matched calculations based on this score. The dual robust estimator combines information from both the parametric estimator based on the outcome model and the nonparametric estimator based on the treatment allocation model for estimation. Run at least two types of heterogeneity estimators in parallel, and calculate and output the average causal effect value estimated by each of the heterogeneity estimators.

[0011] In some embodiments, the consistency check of the average causal effect value output by each estimator, and the output of the fused average causal effect value and its confidence interval if the check passes, includes: Calculate the coefficient of variation among the average causal effect values ​​output by different estimators; The coefficient of variation is compared with a preset consistency threshold, and the effect direction of each average causal effect value is checked to see if they are consistent. If the coefficient of variation is less than the consistency threshold and the effect direction of all average causal effect values ​​is consistent, the verification is deemed to have passed. When the verification is successful, a weighted average is calculated using the reciprocal of the variance of each estimate as the weight, to obtain the fused average causal effect value and its confidence interval.

[0012] In some embodiments, performing a multidimensional rebuttal test on the fused average causal effect value and calculating a clinical causal robustness score based on the results of the multidimensional rebuttal test includes: The processing variable in the original data is replaced with a randomly generated placebo variable; the placebo variable is a variable that has no causal relationship with the outcome variable. Under the minimum confounding variable adjustment set, the heterogeneity estimator is rerun to calculate the effect value of the placebo variable on the outcome variable, which is used as the placebo effect value. Multiple data subsets are generated from the original data through random sampling with replacement. On each of the data subsets, the heterogeneity estimator is run independently to perform the causal effect identification and estimation process, calculate the fused average causal effect value, and obtain a subset effect estimation value sequence; Set a strong confounding factor that is not recorded in the data, and set the strength of its influence on both the processing variable and the outcome variable. Gradually increase the influence intensity of the strong confounding factor, calculate and determine the minimum influence intensity required to make the statistical significance of the fused average causal effect value disappear, and use it as the sensitivity threshold. Based on the placebo effect value, the subset effect estimate sequence, and the sensitivity threshold, the placebo test score, stability score, and sensitivity score are calculated using a preset scoring rule. Based on preset weighting coefficients, the placebo test score, the stability score, and the sensitivity score are weighted and summed to calculate the final clinical causal robustness score; the clinical causal robustness score is a scalar value between 0 and 1.

[0013] In some embodiments, the step of classifying and outputting the candidate risk factors corresponding to the treatment variable as confirmatory causal factors, potential causal factors, or spurious correlation factors based on the confidence interval of the fused average causal effect value and the clinical causal robustness score includes: A first scoring threshold and a second scoring threshold are preset; the first scoring threshold is higher than the second scoring threshold. Based on the confidence interval of the fused average causal effect value and the clinical causal robustness score, the following judgment rule is applied: If the confidence interval of the fused average causal effect value does not contain zero at a preset significance level, and the clinical causal robustness score is greater than or equal to the first score threshold, then the candidate risk factors in the processing variables will be identified and output as confirmatory causal factors. If the confidence interval of the fused average causal effect value does not contain zero at a preset significance level, and the clinical causal robustness score is less than the first scoring threshold but greater than or equal to the second scoring threshold, then the candidate risk factors in the processing variables will be identified and output as potential causal factors. If the confidence interval of the fused mean causal effect value contains zero, or the clinical causal robustness score is less than the second score threshold, then the candidate risk factors in the treatment variables will be identified and output as spurious correlation factors.

[0014] To achieve the above objectives, another aspect of this application proposes a robust screening device for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning, the device comprising: The first module involves acquiring SAH clinical data. The second module performs time-series alignment and multimodal preprocessing on the SAH clinical data to obtain preprocessed data; The third module sets at least one candidate risk factor to be analyzed as a treatment variable; The fourth module establishes a hybrid causal graph based on a medical hard constraint matrix, according to the preprocessed data and the processing variables. The fifth module identifies causal effects based on the hybrid causal graph, determines whether the causal relationship between the treatment variable and the outcome variable is identifiable, and outputs the minimum confounding variable adjustment set. The sixth module, based on the minimum confounding variable adjustment set, uses an integrated strategy to numerically estimate the causal effect between the treatment variable and the outcome variable, runs at least two types of heterogeneous estimators in parallel, and calculates the average causal effect value output by each estimator; The seventh module performs a consistency check on the average causal effect values ​​output by each estimator. If the check passes, it outputs the fused average causal effect value and its confidence interval. The eighth module performs a multi-dimensional rebuttal test on the fused average causal effect value, and calculates a clinical causal robustness score based on the results of the multi-dimensional rebuttal test. The ninth module, based on the confidence interval of the fused average causal effect value and the clinical causal robustness score, classifies and outputs the candidate risk factors corresponding to the treatment variables as confirmatory causal factors, potential causal factors, or spurious correlation factors.

[0015] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0016] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0017] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0018] The embodiments of this application include at least the following beneficial effects: This application provides a robust screening method, device, electronic device, storage medium, and program product for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning. This application includes: performing time-series alignment and multimodal preprocessing on SAH clinical data to obtain preprocessed data; establishing a hybrid causal graph based on a medical hard constraint matrix; outputting a minimum confounding variable adjustment set; performing consistency verification on the average causal effect values ​​output by each estimator, and if the verification passes, outputting the fused average causal effect value and its confidence interval; performing multi-dimensional rebuttal tests on the fused average causal effect value and calculating a clinical causal robustness score; and classifying and outputting the candidate risk factors corresponding to the treatment variables as confirmatory causal factors, potential causal factors, or spurious correlation factors. This application can resolve the conflict between medical priors and data-driven approaches in causal graph construction, improving the interpretability and credibility of the results. Attached Figure Description

[0019] Figure 1 This is a flowchart of a robust screening method for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning, provided in an embodiment of this application. Figure 2 This is a schematic diagram of the robust screening method for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning provided in the embodiments of this application. Figure 3 This is a schematic diagram illustrating the construction principle of a hybrid causal graph based on "medical hard constraints" provided in the embodiments of this application; Figure 4This is a logic diagram of multi-model ensemble estimation and consistency verification provided in the embodiments of this application; Figure 5 This is a diagram of the Clinical Causality Robustness Rating (CCRS) calculation model provided in the embodiments of this application; Figure 6 This is a schematic diagram of the baseline characteristics of SAH patients provided in the embodiments of this application; Figure 7 This is a schematic diagram of constructing a hybrid causal graph based on "medical hard constraints" in verifying the causal effect of Hunt-Hess classification (clinical score) on DCI, provided in the embodiments of this application. Figure 8 This is a schematic diagram of constructing a hybrid causal graph based on "medical hard constraints" in verifying the causal effect of early vasospasm (imaging indicators) on DCI, provided in the embodiments of this application. Figure 9 This is a schematic diagram of constructing a hybrid causal graph based on "medical hard constraints" in verifying the causal effect of the inflammatory factor IL-6 (a biochemical indicator) on DCI, as provided in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0023] 1) SAH, Subarachnoid Hemorrhage, refers to the rupture of a cerebral blood vessel, causing blood to flow into the subarachnoid space on the surface of the brain. This is the core research disease of this invention, and data analysis and risk screening are all centered around SAH patients.

[0024] 2) Delayed ischemic attack (DCI), one of the most common and serious complications after subarachnoid hemorrhage (SAH), usually occurs 3-14 days after the hemorrhage and is a leading cause of disability and death in patients. In this invention, DCI is the outcome variable to be studied, i.e., the target event that is being predicted and prevented.

[0025] 3) The Hunt-Hess grading is a clinical scale that grades the severity of SAH based on patients' clinical symptoms (such as headache, level of consciousness, and neurological deficits). In this invention, it is a key clinical assessment variable and is used as a "treatment variable" (candidate risk factor) to be analyzed in this case to study its causal relationship with DCI.

[0026] 4) Fisher classification, a radiological scale that grades the amount and distribution of subarachnoid hemorrhage on cranial CT scans. It is primarily used to predict the risk of cerebral vasospasm. In this invention, it is an important baseline confounding variable and upstream cause because the amount of hemorrhage (Fisher classification) affects both clinical severity (Hunt-Hess classification) and directly influences the risk of DCI.

[0027] 5) PC algorithm, a classic constraint-based causal discovery algorithm. Its core idea is to gradually eliminate irrelevant edges between variables through a series of conditional independence tests and determine the direction of the edges.

[0028] 6) GES algorithm, Greedy Equivalence Class Search algorithm, is a mainstream score-based causal discovery algorithm. It starts with an empty graph and optimizes a scoring function (such as BIC score) by greedily adding or deleting edges to search for the most likely data generation model.

[0029] 7) Do calculus, a set of formal mathematical rules proposed by Judea Pearl, is used to transform probabilistic expressions that include intervention (do operator) into probabilistic expressions that only contain observation data.

[0030] 8) TMLE, Target Maximum Likelihood Estimation, is an advanced dual robust estimation method. It first uses a machine learning model to make an initial estimate of the outcome, and then corrects the bias of the initial estimate by "targeted" updating the processing mechanism (propensity score), thereby obtaining a more robust estimate of the causal effect.

[0031] 9) AIO, Enhanced Inverse Probability Weighted, is a dual robust estimation method that combines inverse probability weighted and outcome regression models, using an enhancement term to correct potential misspecifications in propensity score models.

[0032] 10) Gamma value, a sensitivity parameter, is a key parameter used in sensitivity analysis of causal inference to quantify the strength of unobserved confounding. A high Gamma value means that a very strong unknown confounding factor is needed to overturn the current causal conclusion, thus indicating that the original conclusion is robust.

[0033] 11) High_IL6, high interleukin-6 level (binary variable), a treatment variable. In order to make causal comparisons, the researchers divided the continuously measured cerebrospinal fluid IL-6 levels into two groups, "high" (≥50, coded as 1) and "low" (<50, coded as 0), based on the median (approximately 50 pg / mL).

[0034] 12) CSF IL-6, or cerebrospinal fluid interleukin-6, is an inflammatory cytokine measured in cerebrospinal fluid that is elevated in the inflammatory response following subarachnoid hemorrhage (SAH). In related studies, it is frequently found to be associated with diabetic cirrhosis (DCI).

[0035] This invention relates to a risk factor screening method based on causal inference theory, designed for the characteristics of data on complications of subarachnoid hemorrhage (SAH), which can automatically correct for confounding bias and quantify the robustness of results.

[0036] This invention mainly relates to the following technical fields: Neurocritical Care: Subarachnoid hemorrhage (SAH) is a critical illness type in neurocritical care. Even after treatments such as aneurysm clipping or embolization, patients still face various risks of secondary brain injury. Among these, neurogenic brain injury (DCI) is one of the most common and important complications, typically occurring 3–14 days after SAH. It can manifest as new-onset or progressive neurological deterioration, leading to permanent neurological dysfunction or even death in severe cases. Numerous studies and clinical practice have shown that DCI is a key factor determining the long-term prognosis and quality of life of SAH patients; therefore, identifying and intervening in high-risk groups for DCI has significant clinical value.

[0037] The pathophysiological mechanisms of diabetic cerebral infarction (DCI) are highly complex, involving multiple processes such as macrovascular spasm, microcirculatory disturbances, diffusive cortical depolarization, inflammatory responses, and blood-brain barrier disruption. Various clinical scores, imaging indicators, biochemical markers, and treatment measures are interconnected in multiple ways. This multifactorial and multipathological complexity makes it difficult for traditional research frameworks based on "single risk factor + regression analysis" to provide a clear and reliable causal chain.

[0038] Risk factor identification and prediction models: In the medical field, identifying disease risk factors and building predictive models is an important research direction. Traditionally, statistical methods (such as regression analysis, logistic regression, Cox regression, etc.) are mainly used to analyze clinical data, identify risk factors associated with disease occurrence, and construct predictive models.

[0039] Causal inference: Traditional statistical methods primarily focus on the correlation between variables, but correlation is not equivalent to causation. Causal inference aims to infer causal relationships between variables from observational data. In recent years, causal inference methods have received increasing attention across various fields and have demonstrated great potential in medical research. DoWhy is a Python-based causal inference library that provides a structured framework for causal inference, including steps such as causal graph construction, causal effect identification, causal effect estimation, and refutation, helping researchers conduct more rigorous causal analysis.

[0040] This invention takes into account that research on the identification of risk factors for DCI after SAH in related technologies mainly focuses on the following paths: Risk factor analysis based on traditional statistical methods: Most studies employ retrospective or prospective cohort designs, collecting data on patients' clinical characteristics (e.g., age, past medical history), bleeding severity (e.g., Hunt-Hess classification, Fisher classification), imaging features (CT / CTA / DSA), laboratory indicators (e.g., inflammatory factors), and treatment information (e.g., nimodipine use, angioplasty, etc.). Statistical models such as Logistic regression and Cox regression are then used to analyze the "correlation" with the occurrence of DCI, and risk scores or predictive models are constructed accordingly. While such studies are valuable in quantifying risk and establishing empirical predictive tools, they essentially remain at the level of "correlation," making it difficult to rigorously distinguish between true "causative factors" and "comorbidities."

[0041] Conventional causal inference applications (including direct use of the DoWhy library): Although causal graphs (DAGs) are introduced, the following unresolved technical challenges remain in practical medical applications: Subjectivity and instability in causal graph construction: Currently, most methods rely entirely on doctors' experience to draw graphs or on data algorithms (such as PC algorithms) for mining. The former is too subjective and easily misses paths; the latter, on data with limited sample size (usually a few hundred cases) and high noise, is prone to constructing erroneous structures that violate common sense about medical timelines (such as "outcome" pointing to "cause").

[0042] Unity of measurement bias in estimation results: Conclusions are typically drawn using only one estimator (e.g., propensity score matching PSM). In cases of extreme imbalance in medical data (e.g., a DCI incidence of only 30%), the variance of a single model is large, making the conclusions unreliable.

[0043] Lack of clinically available quantitative robustness indicators: Existing rebuttal tests typically only output qualitative "pass / fail" results, lacking a comprehensive quantitative score to tell doctors: how "hard" is this causal conclusion? How strong is the unobserved confusion required to overturn it? In view of this, this application provides a robust screening method and related equipment for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning. This approach achieves a leap from "subjective experience / pure data mining" to "medical constraint-guided hybrid structure learning", effectively reducing the error rate of causal mapping, overcoming the bias of single statistical models on medical imbalance data, improving the accuracy of causal effect estimation, and pioneering the "Clinical Causal Robustness Score (CCRS)", which solves the problem of traditional statistical results lacking clinical intuitive interpretation, eliminates "spurious correlation" features from the source, and significantly enhances the generalization ability of subsequent DCI prediction models and intervention strategies.

[0044] The robust screening method for delayed cerebral ischemia after SAH based on hybrid causal structure learning provided in this application relates to the field of neurocritical care medicine data analysis. This method can be applied to terminals, servers, or software running on either a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, or desktop computer, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the robust screening method for delayed cerebral ischemia after SAH based on hybrid causal structure learning, but is not limited to the above forms.

[0045] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0046] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user (patient) information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to a confirmation page. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0047] Figure 1 This is an optional flowchart of a robust screening method for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning, provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S109.

[0048] Step S101: Obtain SAH clinical data; Step S102: Perform time-series alignment and multimodal preprocessing on SAH clinical data to obtain preprocessed data; Step S103: Set at least one candidate risk factor to be analyzed as a treatment variable; Step S104: Based on the preprocessed data and processing variables, establish a hybrid causal graph based on the medical hard constraint matrix; Step S105: Based on the mixed causal graph, identify causal effects, determine whether the causal relationship between the treatment variable and the outcome variable is identifiable, and output the minimum confounding variable adjustment set; Step S106: Based on the minimum confounding variable adjustment set, an ensemble strategy is used to numerically estimate the causal effect between the treatment variable and the outcome variable, and at least two types of heterogeneity estimators are run in parallel to calculate the average causal effect value output by each estimator. Step S107: Perform a consistency check on the average causal effect value output by each estimator. If the check passes, output the fused average causal effect value and its confidence interval. Step S108: Perform a multidimensional rebuttal test on the fused average causal effect value, and calculate the clinical causal robustness score based on the results of the multidimensional rebuttal test. Step S109: Based on the confidence interval of the fused average causal effect value and the clinical causal robustness score, the candidate risk factors corresponding to the treatment variables are classified and output as confirmatory causal factors, potential causal factors, or spurious correlation factors.

[0049] Steps S101 to S109, as illustrated in the embodiments of this application, involve acquiring and preprocessing SAH clinical data, setting analysis objectives, constructing a hybrid causal graph combining medical knowledge constraints and data-driven approaches, automatically identifying and estimating causal effects, performing multi-model integration and consistency verification, and finally conducting multi-dimensional stress tests on the conclusions and quantifying their robustness. Ultimately, candidate factors are output in a tiered manner, achieving an automated and standardized generation system for causal risk factor evidence chains from multi-source heterogeneous clinical data to high-reliability evidence that can directly guide clinical decision-making. This method effectively overcomes fundamental shortcomings in traditional statistical analysis, such as "correlation does not equal causation," subjective model construction, and unknown robustness of conclusions, providing a reliable method for etiological exploration and precise intervention in complex clinical scenarios.

[0050] In step S101 of some embodiments, SAH clinical data can be obtained from a hospital electronic medical record system (EMR) or a neurocritical care database (such as MIMIC), or it can be obtained through other means such as connecting to a multi-center clinical research database (such as the SAH International Registry). It is not limited to these methods. The core is to obtain structured patient data containing time-series information.

[0051] In some embodiments, step S102 may include, but is not limited to, steps S201 to S202: Step S201: Using the onset time of SAH as the timeline, establish a timeline and divide all variables in the SAH clinical data into a set of baseline variables, a set of intermediate process variables, and an outcome variable; the outcome variable includes the occurrence status of delayed cerebral ischemia. Step S202 involves handling missing values, cleaning up outliers, and standardizing the data in the baseline variable set, intermediate process variable set, and outcome variable set to obtain preprocessed data.

[0052] In steps S201 to S202 of some embodiments, a timeline anchored by the onset of illness is established, and all clinical variables are divided into three sets: baseline, process, and outcome. This enables a logical and structured modeling of complex clinical event flows. This ensures that subsequent causal analysis conforms to the temporal logic of "cause first, effect later," and provides a direct basis for applying medical hard constraints such as "temporal prohibition rules," effectively preventing fundamental errors such as causal inversion. By uniformly imputing missing values, cleaning outliers, and standardizing coding the variables in the three sets, the quality normalization and feature engineering standardization of multi-source, multimodal clinical data are achieved. This provides stable and reliable data input for subsequent causal discovery algorithms based on statistical independence (such as the PC algorithm), reduces the interference of data noise and heterogeneity on causal structure learning, and improves the robustness and reproducibility of the entire methodology.

[0053] In some embodiments, step S103 may include, but is not limited to, step S301: Step S301: Select at least one variable from the baseline variable set or the intermediate process variable set as a treatment variable; the selected treatment variable is the candidate risk factor to be analyzed.

[0054] In step S301 of some embodiments, at least one variable is actively selected as a treatment variable from either the baseline variable set (such as "age," "history of hypertension," or other characteristics existing before the onset of illness) or the intermediate process variable set (such as "vasospasm occurring within 72 hours of admission," "nimodipine used during treatment," or other events recorded from the onset of illness to the outcome). For example, "vasospasm" is selected as the target of this round of analysis. This clarifies and focuses the analytical objective. The "treatment variable" specifically refers to the "cause" (i.e., candidate risk factor) to be tested in this round of causal analysis, while the "baseline variable" and "intermediate process variable" are the pool of objective variables from which it originates. This step transforms the broad concept of "all data" into a specific, actionable causal research question, enabling the subsequent complex causal inference process to revolve around a clear "treatment-outcome" relationship, thus improving the flexibility and specificity of the method.

[0055] In some embodiments, step S104 may include, but is not limited to, steps S401 to S406: Step S401: Based on preprocessed data and prior medical knowledge, define a medical hard constraint matrix. In the medical hard constraint matrix, define prohibition rules between variables by setting specific values ​​as prohibition symbols. The prohibition rules include at least: prohibiting the outcome variable from pointing to any variable in the baseline variable set. Step S402: Based on the medical hard constraint matrix, use the constraint-based causal discovery algorithm to perform conditional independence tests on the preprocessed data and generate an initial directed acyclic graph. Step S403: Mark the edges in the initial directed acyclic graph that conflict with weak prior medical knowledge as edges to be adjudicated. Step S404: For each edge to be decided, calculate the change in the overall Bayesian information criterion value of the model when the edge is retained and removed in the initial directed acyclic graph. Step S405: When the absolute value of the change exceeds the preset resolution threshold, retain the edge to be decided; otherwise, remove the edge to be decided to obtain the result of conflict resolution. Step S406: Based on the conflict resolution results, the initial directed acyclic graph is adjusted to obtain a hybrid causal graph based on the medical hard constraint matrix.

[0056] In steps S401 to S406 of some embodiments, S401 can define a medical hard constraint matrix based on prior medical knowledge and set prohibition rules such as "prohibiting outcome variables (e.g., 'DCI') from pointing to baseline variable sets (e.g., 'age', 'gender')" within it. This transforms inviolable medical common sense into rigid constraints that can be executed by the algorithm. This ensures that any subsequent data-driven discoveries must be conducted within this logically consistent "search space," eliminating absurd causal arrows from the source. S402 can perform conditional independence tests on the data using constraint-based causal discovery algorithms (e.g., PC algorithms) under the constraints of the hard constraint matrix, enabling data-driven causal exploration within a medical logical framework. The initial directed acyclic graph generated in this step contains both the potential associations revealed by the data and has eliminated obvious errors that violate the hard constraints, representing a preliminary manifestation of the "hybrid" strategy. Step S403 identifies and problems the gray area between data discovery and expert experience by marking edges in the initial directed acyclic graph that conflict with weak medical prior knowledge (such as "a certain drug may protect nerves, but the level of evidence is not high") as edges to be decided. Steps S404 and S405 calculate the change in the Bayesian information criterion value of each edge to be decided and compare it with a preset resolution threshold to achieve objective and quantitative decision-making based on the strength of data evidence. For example, if the data strongly supports an edge that conflicts with a weak prior (ΔBIC is large), it is retained, acknowledging that the data may reveal new knowledge; otherwise, it is removed according to the prior. This achieves a scientific trade-off and automated decision-making between "data-driven" and "knowledge-guided" approaches when conflicts arise. Step S406 adjusts the initial graph by integrating all decision results, ultimately obtaining a hybrid causal graph based on a medical hard constraint matrix. This generates an optimal causal structure hypothesis that strictly follows the core logic of medicine, fully absorbs hidden information in the data, and has undergone conflict negotiation. This graph serves as a reliable foundation for subsequent causal inferences.

[0057] In some embodiments, step S105 may include, but is not limited to, steps S501 to S503: Step S501: On the topological structure of the hybrid causal graph, based on the backdoor criterion or the frontdoor criterion, traverse and analyze all connection paths to determine whether there is a set of observed variables; the set of observed variables is used to block all non-causal paths from the processing variable to the outcome variable. Step S502: If there is no set of observed variables, it is determined to be unidentifiable, and the subsequent analysis of the current causal relationship is terminated. Step S503: If a set of observed variables exists, the current causal relationship is determined to be identifiable. Then, nodes identified as collision variables and mediator variables by preset rules are excluded from the set of observed variables to obtain the minimum confusion variable adjustment set.

[0058] In steps S501 to S503 of some embodiments, step S501, based on the topological structure of the hybrid causal graph, systematically traverses and analyzes all connection paths, and searches for a set of observed variables (e.g., {age, underlying disease, bleeding volume}) that can block all non-causal paths. This is done using either the backdoor criterion (used to block confounding paths caused by common causes, such as blocking the path "treatment plan → age → rehabilitation effect" by controlling "age") or the frontdoor criterion (used to handle intermediate mechanisms where the processing variable cannot be directly observed). This automates and graph-theorizes the fundamental question of whether "causal relationships can be unbiasedly estimated from observed data," transforming the abstract causal identifiability into a concrete path-blocking problem. Step S502, by introducing a "non-identifiable" determination and process termination mechanism, achieves early identification and resource conservation for causal problems that cannot be solved from existing data. This avoids futile or even misleading numerical estimations in the absence of sufficient information (e.g., the presence of unobserved key confounding factors and the lack of instrumental variables), demonstrating the rigor of the method. Step S503, after determining that the variables are identifiable, automatically outputs a theoretically unbiased "minimum confounding variable adjustment set" by excluding nodes identified as colliding variables (such as the "hospital" node, which is simultaneously affected by "treatment" and "unobserved factors," as this would introduce bias) and mediating variable nodes (such as "changes in physiological indicators" located on the causal path from "treatment" to "outcome," as this would mask the total effect to be estimated). This solves the empirical problem of "what to adjust" and the over / under-adjustment dilemma in traditional regression analysis, providing the correct mathematical adjustment object for obtaining reliable causal estimates.

[0059] In some embodiments, step S106 may include, but is not limited to, steps S601 to S602: Step S601: Based on the minimum confounding variable adjustment set, an ensemble strategy is used to numerically estimate the causal effect and train a heterogeneity estimator. The heterogeneity estimator types include a parametric estimator based on the outcome model, a non-parametric estimator based on the treatment allocation model, and a dual robust estimator. The parametric estimator based on the outcome model is used to fit the conditional probability distribution of the outcome variable given the treatment variable and the minimum confounding variable adjustment set. The non-parametric estimator based on the treatment allocation model is used to estimate the propensity score of the treatment variable given the minimum confounding variable adjustment set, and to perform weighting or matching based on this score. The dual robust estimator is used to simultaneously combine information from the parametric estimator based on the outcome model and the non-parametric estimator based on the treatment allocation model for estimation. Step S602: Run at least two types of heterogeneity estimators in parallel, and calculate and output the average causal effect value estimated by each heterogeneity estimator.

[0060] In steps S601 to S602 of some embodiments, step S601 employs an ensemble strategy and simultaneously trains heterogeneity estimators—including: 1) a parametric estimator based on the outcome model (such as Logistic regression, directly modeling the probability of "outcome DCI" given "treatment variable vasospasm" and "adjustment set {age, Fisher grade}"); 2) a nonparametric estimator based on the treatment assignment model (such as propensity score matching / weighting, simulating random grouping by modeling the probability of "vasospasm" occurring); and 3) a dual robust estimator (such as TMLE, utilizing information from both models simultaneously)—achieving "dual defense" against single model specification errors. In this architecture, if either the outcome model or the treatment assignment model is correctly specified, the final estimate can approach unbiasedness, improving the robustness of estimation in scenarios like SAH where data distribution is complex and models are prone to misspecification. Step S602 improves computational efficiency and enables independent verification of model conclusions by running at least two of the above-mentioned heterogeneity estimators in parallel. Parallelization accelerates the analysis process, and each model independently calculates the average causal effect value based on different mathematical principles (parametric assumptions vs. non-parametric equilibrium vs. dual robustness), providing multiple comparable and convergent sources of evidence for the next step of "consistency verification". This is a key design to ensure the reliability of the conclusions.

[0061] In some embodiments, step S107 may include, but is not limited to, steps S701 to S703: Step S701: Calculate the coefficient of variation between the average causal effect values ​​output by different estimators; Step S702: Compare the coefficient of variation with the preset consistency threshold, and check whether the effect direction of each average causal effect value is consistent. If the coefficient of variation is less than the consistency threshold and the effect direction of all average causal effect values ​​is consistent, the verification is deemed to have passed. Step S703: When the verification is successful, a weighted average is calculated using the reciprocal of the variance of each estimated value as the weight, and the fused average causal effect value and its confidence interval are obtained.

[0062] In steps S701 to S703 of some embodiments, step S701 calculates the coefficient of variation (COP) between the average causal effect values ​​output by different estimators (e.g., ATE of 0.25 from Logistic Regression, ATE of 0.28 from Propensity Score Weighted Estimation, and ATE of TMLE of 0.26), thereby quantifying the dispersion of estimation results from multiple independent models. A smaller COP indicates a high degree of consistency in numerical conclusions despite different model principles, providing objective evidence of robust conclusions. Step S702 compares the calculated COP with a preset consistency threshold (e.g., 0.2) and simultaneously checks whether the effect directions (positive / negative) of each effect value are consistent, thus achieving an automated, double-standard "scientific consensus" mechanism. Verification is considered successful only when the COP is below the threshold (indicating close numerical values) and the directions are consistent (both indicating increased or decreased risk). This effectively prevents outliers deviating from mainstream conclusions due to model misconfiguration from being adopted. Step S703, after successful verification, achieves optimal statistical fusion of multiple highly consistent pieces of evidence by weighting the estimates with the reciprocals of their variances. Estimates with smaller variances (higher accuracy) have greater weight in the final fusion result. The resulting fused average causal effect value and its confidence interval demonstrate higher statistical efficiency (smaller standard errors) and theoretical reliability (based on multiple verifications) compared to estimates from any single model.

[0063] In some embodiments, step S108 may include, but is not limited to, steps S801 to S808: Step S801: Replace the processing variable in the original data with a randomly generated placebo variable; the placebo variable is a variable that has no causal relationship with the outcome variable. Step S802: Under the minimum confounding variable adjustment set, rerun the heterogeneity estimator to calculate the effect value of the placebo variable on the outcome variable, and use it as the placebo effect value; Step S803: Generate multiple data subsets from the original data by random sampling with replacement; Step S804: On each data subset, run the heterogeneity estimator independently, execute the causal effect identification and estimation process, calculate the fused average causal effect value, and obtain the subset effect estimate sequence. Step S805: Set a strong confounding factor that is not recorded in the data, and set the strength of its influence on both the treatment variable and the outcome variable. Step S806: Gradually increase the influence intensity of the strong confounding factor, calculate and determine the minimum influence intensity required to make the statistical significance of the fused average causal effect value disappear, and use it as the sensitivity threshold. Step S807: Based on the placebo effect value, subset effect estimate sequence and sensitivity threshold, the placebo test score, stability score and sensitivity score are calculated according to the preset scoring rules. Step S808: Based on the preset weighting coefficients, the placebo test score, stability score, and sensitivity score are weighted and summed to calculate the final clinical causal robustness score; the clinical causal robustness score is a scalar value between 0 and 1.

[0064] In steps S801 to S808 of some embodiments, steps S801-S802 replace the original treatment variable (e.g., "vasospasm") with a randomly generated placebo variable (a random signal theoretically unrelated to the DCI outcome) and recalculate its effect value under the same analytical procedure to obtain the placebo effect value. This achieves a systematic "stress test" on model overfitting and unadjusted confounding. If the placebo effect value is significantly non-zero, it indicates a defect in the current analytical procedure (model or adjustment set), which may misjudge noise as a signal. Steps S803-S804 can generate multiple data subsets through bootstrap random sampling with replacement and independently run a complete analysis on each subset to obtain a sequence of subset effect estimates, achieving a quantitative assessment of the sample dependence and random volatility of the conclusion. The distribution of this sequence (e.g., mean, variance, confidence interval overlap) intuitively reflects the stability of the causal estimate and is key to assessing whether the conclusion can be reproduced under different sampling conditions. Steps S805-S806 simulate a strong confounding factor not recorded in the data (e.g., a hypothetical "genetic susceptibility") and quantify the minimum influence of this confounding factor required to make the original conclusion statistically significant (i.e., P-value > 0.05 or confidence interval containing 0), i.e., the sensitivity threshold (commonly expressed as Gamma or E value), to achieve a quantitative assessment of resistance to unobserved confounding threats. A higher threshold indicates stronger evidence of absence is needed to overturn the conclusion, suggesting a more robust original conclusion. Steps S807-S808 convert the three test results (placebo effect value, statistical characteristics of the subset estimated sequence, and sensitivity threshold) into placebo test scores (the closer to zero, the higher the score), stability scores (the smaller the fluctuation and the higher the consistency, the higher the score), and sensitivity scores (the higher the threshold, the higher the score) according to preset scoring rules. These scores are then weighted and summed using preset weighting coefficients (e.g., 0.4, 0.3, 0.3) to calculate a clinical causal robustness score between 0 and 1. This process aggregates multidimensional and abstract statistical robustness evidence into a single, intuitive, comparable, and clinically interpretable quantitative indicator. This score answers the question, "How reliable is this risk factor conclusion?", filling the gap in traditional p-values' inability to measure the reliability of conclusions.

[0065] In some embodiments, step S109 may include, but is not limited to, steps S901 to S905: Step S901: Preset a first scoring threshold and a second scoring threshold; the first scoring threshold is higher than the second scoring threshold. Step S902, based on the confidence interval of the fused mean causal effect value and the clinical causal robustness score, apply the following judgment rules: Step S903: If the confidence interval of the fused average causal effect value does not contain zero at the preset significance level, and the clinical causal robustness score is greater than or equal to the first score threshold, then the candidate risk factors in the processing variables will be identified and output as confirmatory causal factors. Step S904: If the confidence interval of the fused average causal effect value does not contain zero at the preset significance level, and the clinical causal robustness score is less than the first scoring threshold but greater than or equal to the second scoring threshold, then the candidate risk factors in the processing variables will be identified and output as potential causal factors. In step S905, if the confidence interval of the fused mean causal effect value contains zero, or the clinical causal robustness score is less than the second scoring threshold, then the candidate risk factors in the processing variables will be identified and output as spurious correlation factors.

[0066] In steps S901 to S905 of some embodiments, step S901 establishes clear and adjustable objective standards for subsequent grading decisions by presetting a first scoring threshold (e.g., 0.8) and a second scoring threshold (e.g., 0.6), with the first threshold (high threshold) being higher than the second threshold (low threshold). These two thresholds divide the continuous robustness score (CCRS) into intervals with different clinical meanings, providing clear boundaries for automated decision-making. Steps S902-S905 construct and apply a set of automated judgment rules based on dual objective evidence: these rules simultaneously consider the confidence interval of the fused average causal effect value (statistical evidence, judging "whether the effect exists") and the clinical causal robustness score (robust evidence, judging "how reliable the conclusion is"), realizing the transformation from complex analysis results to discrete decisions that can directly guide clinical actions. Specifically: Step S903 identifies variables as confirmatory causal factors (e.g., analysis confirms "Hunt-Hess grade ≥ 3" as a confirmatory factor for DCI) by requiring "confidence interval not to contain zero" (statistically significant) and "CCRS ≥ high threshold," thus identifying factors with both strong statistical evidence and robustness tests. This provides evidence support for the development of mandatory intervention guidelines in clinical practice. Step S904 identifies variables as potential causal factors (e.g., "early vasospasm" effect is significant but CCRS is 0.75) by requiring "confidence interval not to contain zero" but "CCRS in the moderate range," thus identifying "signals of concern" with clear effects but whose robustness requires further verification. Step S905 can classify variables as spurious factors (e.g., the "high CSF IL-6 level" effect is weak and the CCRS is only 0.58) by determining whether the "confidence interval contains zero" (statistically insignificant) or "CCRS is below a low threshold." This systematically excludes "concomitant biomarkers" or "spurious associations" that are easily misjudged by traditional methods from important conclusions. This avoids wasting clinical effort on ineffective interventions and improves decision-making efficiency.

[0067] This application takes into account the existence of problems and shortcomings, mainly due to the following reasons: The complexity of the pathophysiological mechanism of SAH-DCI: The occurrence of DCI involves the interaction of multiple factors, such as vasospasm, inflammatory response, microcirculatory disturbance, and neuronal damage. Traditional linear statistical models are difficult to capture this complexity.

[0068] The inherent limitations of observational research data: Most current studies on SAH-DCI are observational studies, which cannot strictly control for various confounding factors like randomized controlled trials (RCTs), making it difficult to directly infer causal relationships. Ethical and practical constraints also make it difficult to conduct RCT studies targeting many potential risk factors.

[0069] The mismatch between causal discovery algorithms and clinical data characteristics: Existing mainstream causal structure learning algorithms (such as PC and GES algorithms) typically rely on the assumption of "large sample asymptotic consistency," meaning that massive amounts of data are required to accurately infer the independence between variables. However, SAH clinical data usually has a limited sample size (hundreds of cases) and is accompanied by high-dimensional noise. This mismatch in data characteristics makes it easy for purely data-driven algorithms to produce erroneous structures that violate medical timelines (such as "outcome pointing to cause") or miss critical paths.

[0070] Limitations of Single Estimation Models: Any method for estimating causal effects is based on specific mathematical assumptions. For example, regression models assume a linear relationship between variables, and propensity score matching (PSM) relies on the overlap of the "common support." In cases of extreme imbalance in SAH complication data (low incidence of DCI), a single model struggles to satisfy all assumptions simultaneously, making it impossible to systematically eliminate bias.

[0071] The disconnect between statistical significance and clinical decision confidence: Existing causal inference tools (such as the DoWhy native library) are primarily geared towards researchers, and their rebuttal tests output abstract p-values ​​or effect fluctuation ranges. These statistical indicators fail to translate into quantitative standards that intuitively reflect the "robustness of conclusions," making it difficult for clinicians to determine whether the causal conclusion remains reliable when faced with potential unobserved confounding factors (such as unrecorded gene mutations or microenvironmental changes).

[0072] The difficulties encountered in solving these problems included: Identification and control of confounding factors: SAH-DCI studies present numerous potential confounding factors, such as patient age, underlying diseases, SAH severity, and treatment regimens. Comprehensively identifying and effectively controlling these confounding factors is a challenge faced by both traditional statistical and machine learning methods.

[0073] The challenge of integrating medical prior knowledge with data-driven logic lies in transforming textualized and rule-based medical expert knowledge (such as anatomical pathways and pathological timelines) into mathematically computable "hard constraint matrices," and dynamically correcting erroneous connections generated by data mining algorithms (such as conditional independence tests) without compromising their objectivity. This is a complex optimization problem involving graph theory and Bayesian networks. If the constraints are too strong, it degenerates into manual graph drawing; if the constraints are too weak, logical fallacies cannot be eliminated.

[0074] Consistency verification for heterogeneous model ensembles: When constructing ensemble estimation strategies, different models (such as parametric regression models and non-parametric machine learning models) process data in drastically different ways. Designing a universal verification mechanism that accurately identifies "estimation divergence" caused by conflicting model assumptions while ensuring computational efficiency, and extracting a reliable average effect from multiple contradictory estimates, presents significant challenges in algorithm design.

[0075] Quantitative Modeling of Unobserved Confusion Effects: "Unobserved confusion" is inherently invisible. To establish a robustness scoring system (CCRS) that conforms to both statistical principles and clinical understanding through simulation in the absence of real-world data requires constructing complex sensitivity analysis equations. In particular, setting reasonable thresholds that allow the scoring to filter out spurious correlations without overlooking potentially weak causal signals due to overly stringent standards necessitates extensive experimental validation and parameter tuning.

[0076] The technical problem this invention aims to solve is as follows: Although general causal reasoning frameworks (such as workflows based on structural causal models, SCM) provide a basic path for the analysis of observational medical data, there is still a significant "last mile" technical gap in the specific clinical field of subarachnoid hemorrhage (SAH) complicated by delayed cerebral ischemia (DCI): the general framework lacks structural constraint mechanisms for the complex pathological time series of SAH, lacks integrated estimation strategies adapted to highly imbalanced data, and lacks clinical evaluation criteria that can quantify the reliability of causal conclusions.

[0077] To address the shortcomings of the aforementioned related technologies, this invention mainly solves the following three core technical problems: ① Resolve the conflict between medical a priori knowledge and data-driven approaches in causal graph construction, achieving accurate identification from "spurious correlations" to "true causality": In the complex clinical scenario of SAH–DCI, purely data-driven data mining algorithms are prone to producing erroneous structures that violate common medical sense (such as outcomes pointing to causes), while purely expert experience is insufficient to cover all pathways. This invention aims to address how to systematically eliminate biases introduced by confounders and colliders in large-scale observational data through the organic integration of time-series hard constraints and statistical independence tests, thereby identifying risk factors with real effects, rather than merely "accompanying variables" statistically related to DCI.

[0078] ② Address the issue of lacking quantitative robustness indicators for causal conclusions, and improve the interpretability and credibility of the results: Related technologies often only provide binary conclusions of "significant / non-significant," lacking a quantitative assessment of the robustness of the results. This invention aims to address this issue by constructing a comprehensive evaluation system that includes placebo testing, subset resampling, and unobserved confounding sensitivity analysis. This system transforms the "mathematical assumptions implicit in the algorithm" into visual, robustness evidence with explicit numerical scores (such as CCRS scores), clearly informing clinicians to the extent to which the conclusions regarding this risk factor can withstand the challenges of unknown interference.

[0079] ③ This addresses the problem of unstable feature selection leading to poor generalization ability of the prediction model, providing a reliable feature base for subsequent interventions: Traditional predictive models often suffer from poor cross-center generalization (overfitting to bias) due to the inclusion of unstable spurious correlation variables. This invention aims to construct a feature selection method that ensures that variables included in DCI predictive models or clinical intervention strategies undergo rigorous double robust estimation and causal confidence testing, thereby reducing model performance fluctuations and clinical intervention ineffectiveness caused by "spurious correlations" from the outset.

[0080] Through the above technical solutions, this invention aims to overcome the limitations of existing general causal analysis tools in the application of post-SAH DCI, such as subjective structure construction, large estimation variance, and unknown robustness, and provides reliable computational methodology support for understanding the occurrence mechanism of DCI and developing highly reliable prevention strategies.

[0081] As an optional implementation, this application provides a robust screening method for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning. This method uses clinical data from SAH patients as its foundation, employs a causal graph as its core representation, and sequentially completes data preprocessing, causal graph construction, causal effect identification, causal effect estimation, causal effect refutation, and causal risk factor output within a causal reasoning framework. Its technical solution mainly includes the following steps: Step S1: Temporal alignment and multimodal preprocessing of SAH clinical data: Data sources: hospital electronic medical record system, neurocritical care database, or multicenter clinical research database.

[0082] Data content: Collect various clinical variables that may be associated with the occurrence of DCI, including but not limited to: Patient basic information: age, gender, past medical history (such as hypertension, diabetes, smoking history, etc.).

[0083] SAH-related characteristics: SAH etiology (aneurysm, trauma, etc.), Hunt-Hess classification, Fisher classification, Glasgow Coma Scale (GCS) score on admission, bleeding site, bleeding volume, and early vasospasm (assessed by TCD or DSA).

[0084] Treatment-related information: surgical procedures (craniotomy, interventional embolization), treatment of vasospasm (nimodipine, balloon angioplasty, etc.), fluid management, blood pressure management, ventilator support, etc.

[0085] DCI occurrence details: whether DCI occurred (determined based on clinical criteria and imaging evidence), the time of DCI occurrence, the severity of DCI, and the duration of DCI.

[0086] Laboratory test indicators: serum sodium, blood glucose, blood lipids, and inflammatory markers (such as C-reactive protein and white blood cell count) upon admission.

[0087] Imaging data: CT, CTA, MRI, DSA and other imaging data.

[0088] Time-series anchor setting: Establish a time axis with "SAH onset time" as T0 and "discharge or DCI occurrence" as Tend. Divide all variables into three categories: baseline variables ( V base (Existing before T0), intermediate process variables () V mid (occurring between T0 and Tend, such as vasospasm, infection) and outcome variables ( Y (i.e., DCI).

[0089] Data preprocessing: Missing value handling: Use appropriate imputation methods (such as multiple imputation) or delete unqualified samples; Outlier identification and handling: Clean up unreasonable observations based on clinical common sense and statistical rules; Variable transformation: Appropriately encode categorical variables (such as one-hot encoding), and standardize or normalize continuous variables; Data partitioning: Divide the dataset into training and test sets according to a ratio (e.g., 7:3 or 8:2).

[0090] Step S2: Construct a hybrid causal graph based on "medical hard constraints": Define the constraint matrix: Establish a temporal constraint matrix reflecting prior medical knowledge. This matrix defines the rules for prohibiting connections, including: Temporal prohibition: Outcome variables cannot be the parent node of baseline variables (i.e., prohibited). Y →V base ); Anatomical prohibition: Downstream complications must not lead to upstream pathological changes (e.g., "DCI" must not lead to "aneurysm rupture").

[0091] Constrained structure learning: Under the constraints of the constraint matrix, the preprocessed data is tested for conditional independence using a constraint-based causal discovery algorithm (such as the PC-Stable algorithm) to initially generate a data-driven causal directed acyclic graph (DAG).

[0092] Conflict resolution and graph generation: For edges that conflict with weak prior medical knowledge (non-hard constraints) discovered by the algorithm, calculate the change in Bayesian Information Criterion (BIC) when retaining and removing the edge, select the structure with the least information loss, and finally generate the optimal SAH-DCI causal graph that takes into account both medical logic and data characteristics.

[0093] Variable selection: Treatment Variable: This refers to the potential risk factors to be investigated, such as early vasospasm, Hunt-Hess classification, Fisher classification, etc. One or more potential risk factors can be investigated at a time.

[0094] Outcome Variable: namely, delayed cerebral ischemia (DCI).

[0095] Confounding variables: These are variables that may affect both the treatment variable and the outcome variable. They need to be carefully identified and included based on domain knowledge. Examples include age, severity of SAH, and underlying diseases.

[0096] Mediator Variables: These are intermediate variables located in the causal path between Treatment Variable and Outcome Variable. For example, vasospasm may act as a mediator of the influence of Hunt-Hess grading on DCI.

[0097] Collider Variables: These are variables that are affected by two or more other variables. They need to be handled carefully to avoid introducing bias.

[0098] Step S3: Causal Effect Identification Based on Graph Structure In this step, based on the hybrid causal graph (DAG) constructed in step S2, the system automatically determines whether the causal relationship between the variable to be analyzed and the outcome variable is statistically "identifiable" using graph theory rules, and outputs the corresponding mathematical expression. Specifically, this includes: Causal effect type definition: The type of causal indicator to be estimated is set according to the specific needs of the clinical study. The default setting is usually Average Treatment Effect (ATE), which is the average effect of changes in the treatment variable (such as Hunt-Hess classification) on the probability of occurrence of the outcome variable (DCI) in the overall population. When specific subgroup analyses are involved, Conditional Average Causal Effect (CATE) can be set, which is the average causal effect given a specific covariate (such as age > 60 years).

[0099] Automatic Identifiability Verification: Based on the constructed causal graph topology, the Do-calculus rule system is applied for path analysis. The system automatically traverses all connection paths between Treatment and Outcome, and based on the Backdoor Criterion or Frontdoor Criterion, determines whether there exists a set of observed variables sufficient to block non-causal paths (confusion paths). Judgment Logic: If a set of observed variables satisfying the above criteria exists, the causal effect is determined to be "identifiable," meaning it can be observed from a distribution. P ( Y , T , X The intervention distribution was derived from this. P ( Y | do ( T Otherwise, it is determined to be unidentifiable, and the analysis of that variable is terminated.

[0100] Identification Strategy and Adjustment Set Generation: Once the target effect is determined to be identifiable, the system will automatically match the optimal identification strategy based on the graph structure characteristics. Priority is given to finding a backdoor adjustment strategy, outputting a minimum adjustment set (Z). This set includes all necessary confounding factors (such as age and underlying diseases) while automatically excluding collider and mediator variables, thus providing the correct mathematical expression for subsequent numerical estimation.

[0101] If there are unobserved confounding variables but instrumental variable nodes exist, the system will automatically switch to the instrumental variable strategy.

[0102] Step S4: Causal effect estimation based on the "integrated strategy": In this step, to overcome the potential bias of a single statistical model on small SAH samples and imbalanced data, this invention does not rely on a single algorithm, but instead employs a "heterogeneous multi-model ensemble strategy" to numerically estimate the causal relationships identified in step S3. The specific process is as follows: S4.1 Construction and selection of heterogeneity estimator array: Based on the identification strategy determined in step 3 (such as backdoor adjustment), the system selects at least two types of estimators with different mathematical principles to construct an ensemble array to capture different aspects of the data distribution: Estimation based on outcome-based methods: Construction P ( Y | T , Z The model uses linear or logistic regression to directly estimate the marginal effect of T on Y by adjusting the covariate Z.

[0103] Estimation based on treatment assignment models (Treatment-assignment methods): Construction P ( T | Z The model employs propensity score matching (PSM) or inverse probability weighting (IPW) to simulate a randomized controlled trial environment by balancing the covariate distributions of the treatment and control groups.

[0104] Double Robust methods employ Target Maximum Likelihood Estimation (TMLE) or AIO (Augmented Inverse Probability Weighting). This method combines both the outcome model and the treatment allocation model; if either model is correctly configured, the final estimate is unbiased.

[0105] S4.2 Model Training and Parallel Computation: Using the SAH training set data, the selected estimator models are trained in parallel. The system calculates the average treatment effect (ATE) and its corresponding 95% confidence interval (CI) for each estimator.

[0106] S4.3, Consistency Verification of Effects and Fusion of Results: To ensure the reliability of the conclusions, the system introduces a consistency verification mechanism: Calculate the coefficient of variation (CV) of the ATE values ​​obtained from different estimators.

[0107] Judgment rule: If the CV is less than a preset threshold (e.g., 0.2) and the directions (positive / negative) of the estimated values ​​from each model are consistent, then the estimation result is considered convergent. The system will output the variance-weighted ensemble ATE value as the final estimation result. If the CV is greater than the threshold or there is a conflict in direction, it indicates a risk of model dependency, and the factor is marked as "effect uncertain," requiring manual intervention or an increase in the sample size.

[0108] S4.4 Final Effect Assessment: The statistical significance of the fused ATE value is assessed. Only when the confidence interval of the fused ATE does not contain 0 (i.e., P < 0.05) is the treatment variable considered to have a statistically significant causal effect on the outcome variable, and the process proceeds to the next step of robustness refutation.

[0109] Step S5: Refutation and Scoring of Causal Effects To avoid potential biases from relying solely on a single model's estimation results, this invention does not directly output conclusions after estimating causal effects. Instead, it executes a multi-dimensional rebuttal and stress testing process. For each treatment variable to be validated, the system sequentially performs the following algorithmic steps and calculates the Clinical Causal Robustness Score (CCRS): S5.1 Placebo Interference Test (Spurious Correlation Screening): Operational logic: While keeping the adjustment set of confusion variables unchanged, the system replaces the treatment variable in the original data with a randomly generated "placebo variable" that has no real causal relationship with the outcome.

[0110] Calculate the effect size: Rerun the estimator from step S4 to calculate the effect size of the placebo variable on the DCI. ATE placebo .

[0111] Judgment criteria: In theory ATE placeboThe value should approach 0. The system calculates the degree to which it deviates from 0 and includes it as a negative penalty in the total score. If the placebo effect is significant (P<0.05), it indicates that the model has severe overfitting or unadjusted structural bias.

[0112] S5.2 Subset Resampling Test (Stability Check): Operational logic: Use the Bootstrap method or random sampling strategy to extract multiple subsets of the original data (e.g., extract 80% of the samples and repeat N times; or divide them according to different hospital centers).

[0113] Calculation of indicators: Repeat the estimation process of step S4 independently on each subset to obtain the effect size distribution sequence. ATE 1, ATE 2, ..., ATE n}

[0114] Judgment criteria: Calculate the confidence interval overlap rate and effect direction consistency rate of the sequence. If the estimation results of each subset fluctuate within a reasonable range and the direction remains consistent, then the causal relationship is determined to be insensitive to sample fluctuations and possesses high stability.

[0115] S5.3, Sensitivity Analysis - Latent Variable Screening: Operational logic: Simulate a hypothetical, unobserved confounder that is not recorded in clinical data, and assume it affects both Treatment and DCI. The system progressively increases the effect strength of this hypothetical variable.

[0116] Calculation index: Calculate the minimum interference threshold (Gamma value or E-value) required to make the original causal effect disappear (i.e., the P value becomes >0.05) or the effect value return to zero.

[0117] Judgment criteria: The higher the threshold, the more likely it is that extremely strong (clinically unrealistic) unknown factors are needed to overturn the current conclusion, thus proving that the original result is "robust" against unobserved confounding.

[0118] S5.4 Calculation and determination of Clinical Causality Robustness Score (CCRS): Scoring Formula: The system normalizes the test results from the above three dimensions and calculates a comprehensive score based on preset weights. CCRS =ω1·(1 | Norm ( ATE placebo)∣)+ω2· Consistency subset +ω3·Γ sensitivity ; Interpretation of results: The score is a value between 0 and 1, which quantifies the reliability of causal conclusions.

[0119] If the refutation test results in a CCRS score that is too low (e.g., <0.6), the system will automatically trigger an alarm, prompting a need to re-examine the causal graph hypothesis in step S2 or increase the sample size, and will mark the variable as "low confidence".

[0120] Step S6: Causal risk factor classification output: Based on the estimation and refutation tests of causal effects, this invention sets a unified causal determination criterion for each treatment variable, and divides candidate variables into three levels of output according to the CCRS score and the significance of ATE: Confirmatory causal factors: ATE is significant and CCRS is higher than the high threshold (e.g., 0.8), which is recommended as the core target for clinical intervention; Potential causal factors: ATE was significant but CCRS was in the intermediate range, suggesting the need for further multicenter validation; False correlation factors: If ATE is not significant or CCRS is below the low threshold, it is considered an accompanying phenomenon and it is recommended to remove it from the prediction model.

[0121] The output set of causal risk factors can be further used for: Construct DCI prediction models based solely on causal features (such as Logistic regression, random forest, deep learning models, etc.) to improve the stability and generalization ability of the models; guide the design of clinical intervention strategies for this risk factor, emphasizing the focus on "etiology" rather than just "companion markers".

[0122] The process conditions provided in the embodiments of this application are as follows: The implementation of this invention relies on a specific medical big data computing environment and algorithm parameter configuration. To ensure the accuracy of the hybrid causal graph construction and the convergence of the integrated estimation strategy, the specific process conditions are as follows: 1. Computing hardware environment: Since this invention relates to combinatorial search of structure learning algorithms (such as PC-Stable) and high-frequency resampling (Bootstrap) calculation in step S5, it places clear requirements on computing power: Processor: It is recommended to use a multi-core processor (8 cores or more) to support parallel computation of multiple estimators (regression, IPW, TMLE) in step S4, as well as parallel processing of data subset refutation in step S5.

[0123] Memory: It is recommended to configure 32GB or more of RAM to support the calculation of SAH multimodal time-series feature matrix and the storage of high-dimensional adjacency matrix in the causal graph search process.

[0124] Storage: High-speed I / O capability (SSD) is required to meet the needs of frequent reading of large-scale clinical data and writing of intermediate process data (such as the P-value distribution of tens of millions of simulations).

[0125] 2. Software architecture and dependency libraries: This invention is built on the Python ecosystem, but it does not only rely on general libraries; it also integrates domain-specific algorithm modules. Core language: Python 3.8 and above.

[0126] Data processing layer: Pandas, NumPy (used for time alignment and matrix operations).

[0127] Causal structure learning layer: Causal-Learn or pgmpy library (used to implement the PC-Stable algorithm and conditional independence test in step S2).

[0128] Causal inference framework layer: DoWhy (as a top-level encapsulation), combined with EconML or zEpid (for implementing the dual robust estimation TMLE in step S4).

[0129] Custom modules: The unique "Constraint Matrix Generator", "Integration Consistency Verifier" and "CCRS Scoring Calculator" of this invention need to be implemented through custom Python classes.

[0130] 3. Critical Hyperparameters: The effectiveness of this invention highly depends on the precise configuration of the following three sets of key parameters, which need to be preset according to the specific characteristics of the SAH data (sample size, sparsity): Group A: Cause-effect graph construction parameters (for step S2): Independence test significance level (α): It is recommended to set it to 0.01 or 0.05. This value determines the strictness with which the PC algorithm determines whether two variables are "independent". In small sample data of SAH, it is recommended to relax it to 0.05 to avoid missing weak signals.

[0131] BIC conflict resolution threshold (δ): It is recommended to set it to 2.0-10.0. When data-driven edges conflict with weak medical priors, prior knowledge is only allowed to be overridden if the optimization amount of the BIC score exceeds this threshold.

[0132] Group B: Integrated estimation and verification parameters (for step S4): Consistency check threshold ( CV threshold ) It is recommended to set it to 0.20. That is, the conclusion is considered convergent only when the coefficient of variation of ATE calculated by the three different estimators is less than 20%.

[0133] Trimming Threshold: When performing IPW (Inverse Probability Weighted) estimation, it is recommended to set a cutoff value of [0.05, 0.95] to prevent extreme propensity score weights from causing an explosion in the estimated variance.

[0134] Group C: Robustness scoring parameters (for step S5): Number of resampling times ( N boot It is recommended to set it to at least 50-100 times to ensure... S subset Statistical stability of the scores.

[0135] Unobserved confusion intensity limit (Γ) max ): It is recommended to set the scan range to [1.0, 5.0] and the step size to 0.1.

[0136] CCRS score weights (ω1, ω2, ω3): Recommended configuration is ω1=0.4 (placebo weight), ω2=0.3 (subset stability weight), ω3=0.3 (sensitivity weight).

[0137] Through the above-mentioned strict control of process conditions, the present invention can overcome the uncertainty of general software when processing complex clinical data of SAH, and ensure that the final output of DCI risk factors has reproducibility and clinical-grade high confidence.

[0138] refer to Figure 2 This application's embodiments demonstrate the complete logical flow of "a robust screening method for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning." This process consists of six core steps, reflecting the transformation from raw data to high-confidence causal conclusions: S1. Temporal alignment and multimodal preprocessing of SAH patient clinical data: Input SAH clinical data, establish a temporal axis with the onset time as the anchor point, and complete missing value imputation and feature alignment.

[0139] S2. Constructing a hybrid causal graph based on "medical hard constraints": This is one of the core steps of this invention. The system does not directly use data mining, but combines "medical hard constraint matrix" with "data-driven structure learning (PC algorithm)" to generate a hybrid causal graph (DAG) that conforms to pathological common sense and reflects data characteristics.

[0140] S3. Causal effect identification based on graph structure: Based on the generated DAG, the backdoor criterion is used to automatically find the minimum set of confounding variables.

[0141] S4. Causal effect estimation based on "integration strategy": Three heterogeneous models (regression, IPW, and dual robustness) are run in parallel, and the average causal effect (ATE) is determined by consistency check (CV threshold).

[0142] S5. Causal effect refutation and robustness quantification score: The estimation results are "stress-tested" and a comprehensive CCRS score including placebo effect, subset stability and sensitivity threshold is calculated.

[0143] S6. Causal Risk Factor Classification Output: Based on the CCRS score, risk factors are classified into three categories: "confirmatory", "potential", or "spurious".

[0144] refer to Figure 3 This application provides a schematic diagram of the construction principle of a hybrid causal graph based on "medical hard constraints" (corresponding to step S2), which shows in detail the technical solution for resolving the conflict between "expert experience subjectivity" and "data mining bias" in step S2.

[0145] The input on the left displays a preset medical hard constraint matrix. This matrix stores the prohibition rules in a two-dimensional table, for example... Matrix [ Outcome ][ Baseline ]=0 indicates that connections from the outcome variable to the baseline variable are prohibited.

[0146] Intermediate processing: This demonstrates the learning process for constrained structures. Data-driven algorithms (such as PC-Stable) attempt and orient edges only within the search space allowed by the constraint matrix when performing conditional independence tests.

[0147] Conflict resolution module: When the data statistics results (such as BIC score) indicate the existence of an undefined strong correlation edge, but the edge conflicts with the weak medical prior, the system calculates ΔBIC and retains the edge only if it is greater than the threshold δ.

[0148] Output on the right: The final generated optimal SAH-DCI causal graph, which eliminates erroneous paths that violate temporal logic and retains hidden paths found in the data.

[0149] refer to Figure 4 The multi-model ensemble estimation and consistency verification logic diagram (corresponding to step S4) provided in the embodiments of this application demonstrates how the present invention overcomes the bias of a single statistical model through its ensemble strategy.

[0150] The process begins with the input of the "minimum adjustment set Z" and is divided into three parallel computation branches: Branch A (parameterization): Directly fit using a Logistic regression model. P ( Y | T , Z ).

[0151] Branch B (Weighted Method): Uses inverse probability weighting (IPW) to balance the distribution of covariates. P ( T | Z ).

[0152] Branch C (semi-parametric): Uses dual robust estimation (TMLE) while utilizing information from both models mentioned above.

[0153] Convergence node (consistency check): The system calculates the coefficient of variation (CV) of the ATE values ​​output by the three branches.

[0154] Logical judgment: If CV <0.2 and consistent in direction: Output "High Confidence Estimate" (weighted average).

[0155] like CV ≥0.2 or directional conflict: Triggers an "uncertainty alert," halting subsequent processes or prompting manual intervention. This mechanism effectively prevents erroneous conclusions caused by model misspecification.

[0156] refer to Figure 5 This application provides a Clinical Causality Robustness Rating (CCRS) calculation model diagram (corresponding to step S5), demonstrating the calculation logic of the original CCRS scoring system of this invention. The diagram vividly explains how the score quantifies the "hardness of causal conclusions." The diagram includes dashboards with three input dimensions: Placebo Test Dimension: Displays the ATE distribution of the placebo variable; the closer to 0, the higher the score. Subset Stability Dimension: Displays the ATE confidence intervals after 10 resampling iterations; the higher the interval overlap rate, the higher the score. Sensitivity Analysis Dimension: Displays the Gamma curve; the slower the curve declines (i.e., the stronger the unknown confusion required to overturn the conclusion), the higher the score.

[0157] Central Computing Core: Demonstrating the weighted summation formula CCRS =ω1 S placebo +ω2 S subset +ω3 S sensitivity .

[0158] Output: A scalar value between 0 and 1, used for the final hazard factor classification.

[0159] Key points of this application: The main contribution of this invention lies in overcoming three major shortcomings of general causal inference software when processing complex clinical data of SAH: "subjective structure construction," "single-model bias," and "lack of quantitative evaluation standards." The key innovations and the technical solution to be protected are as follows: Key Point 1: Hybrid Cause-Effect Graph Construction Technique Based on "Medical Hard Constraint Matrix" (corresponding to step S2): The essence of innovation: to resolve the contradiction between "expert experience in graphing is prone to omissions" and "pure data mining (such as PC algorithms) is prone to violating time sequences".

[0160] Core logic: This invention does not directly let the algorithm run on the data, but predefines an N×N hard constraint matrix.

[0161] Protected content: The rules for defining matrices include the temporal prohibition rule (prohibiting the outcome variable from pointing to the baseline variable) and the anatomical necessity rule.

[0162] Constrained search algorithm: Perform conditional independence tests within the "legal subspace" defined by the constraint matrix.

[0163] Conflict resolution mechanism: When a data-driven edge discovery conflicts with a weak prior, a decision logic based on the change in the Information Criterion (BIC) is used.

[0164] Key Point 2: Causal Effect Debiasing Estimation Strategy Based on "Heterogeneous Multi-Model Integration" (corresponding to Step S4): The essence of the innovation is to solve the problem that imbalanced SAH data leads to unreliable results from a single model (such as using only PSM).

[0165] Core logic: Construct an integrated estimator that includes an outcome regression model, an treatment allocation model (IPW), and a dual robust model (TMLE).

[0166] Protected content: Parallel computing architecture: running the above three heterogeneous models simultaneously.

[0167] Consistency check logic: The coefficient of variation (CV) is defined as the output gating. The result is considered valid only when CV < a preset threshold (e.g., 0.2) and the effect direction is consistent. This is an algorithmic mechanism to prevent model misspecification.

[0168] Key Point 3: Clinical Causality Robustness Rating (CCRS) Quantitative Evaluation System (corresponding to Step S5): The essence of the innovation is to transform the abstract statistical indicators output by tools such as DoWhy into 0-1 scores that are readable by clinicians.

[0169] The core logic is to establish a weighted calculation formula to comprehensively measure the "hardness" of the conclusion.

[0170] Protected content: Scoring formula structure: Protection CCRS =ω1 S placebo +ω2 S subset +ω3 S sensitivity This computational model.

[0171] Specification of rating dimensions: S placebo : Penalty term based on the placebo effect value.

[0172] S subset Stability term based on resampling confidence interval overlap rate.

[0173] S sensitivity : Anti-interference term based on the unobserved confusion intensity threshold (Gamma / E-value).

[0174] Key Point 4: Multimodal Temporal Data Alignment and Feature Engineering for SAH-DCI (corresponding to Step S1): The essence of innovation: solving the problem that raw clinical data cannot be directly used for causal graph analysis.

[0175] Protected content: "Event anchoring" mechanism: based on the time of onset T Using 0 as the baseline, variables are strictly divided into a baseline set. V base process set V mid and the final episode Y This serves as the basis for generating the constraint matrix.

[0176] Compared with related technologies (i.e., traditional multiple regression analysis, black-box machine learning prediction models, and unoptimized general causal inference software applications), this invention has the following three significant technical advantages: 1. From "subjective mapping / blind digging" to "medically guided precision discovery," significantly reducing the error rate in causal structure analysis: This invention pioneers a hybrid construction strategy combining "temporal hard constraint matrix + data-driven structure learning." By pre-setting medical ironclad rules such as "the outcome cannot point to the baseline," the search space of the algorithm is forcibly constrained, correcting logical fallacies in data mining while retaining the ability to discover potential new associations. It can automatically eliminate over 90% of spurious paths that violate medical common sense due to data noise, ensuring that the generated causal graph is anatomically and pathophysiologically sound.

[0177] 2. From "single-model bias" to "dual robust ensemble," the estimation challenge of extremely imbalanced medical data has been overcome: This invention introduces a "heterogeneous multi-model ensemble strategy," running parallel result regression (Parametric), inverse probability weighting (Non-parametric), and semi-parametric estimation, and uniquely incorporates a coefficient of variation (CV) consistency verification mechanism. It achieves "double robustness," meaning that as long as either the result model or the allocation model is correctly defined, the final result is an unbiased estimate. If there are conflicts between model conclusions, the system automatically issues an alert instead of outputting erroneous conclusions, greatly improving the scientific validity and reliability of the conclusions.

[0178] 3. From "abstract P-value" to "Clinical Causality Robustness Score (CCRS)," a gap has been filled in the evaluation criteria for clinical decision-making: This invention establishes a multidimensional quantitative evaluation system—the CCRS score. This score integrates the results of placebo tests (to rule out spurious correlations), subset resampling (to rule out sample fluctuations), and sensitivity analysis (to rule out latent variables), transforming complex statistical robustness into an intuitive score between 0 and 1. It provides clinicians with clear grading criteria (e.g., CCRS > 0.8 indicates a confirmatory factor and can be included in guidelines; CCRS < 0.6 indicates a spurious correlation and should be excluded). This transforms causal inference from a research tool into a decision support system that can directly guide clinical practice.

[0179] 4. Improved the cross-center generalization ability of subsequent predictive models and intervention strategies: This invention serves as a pre-filter for feature engineering, eliminating "accompanying markers" that are only statistically correlated but have no causal effect at the source.

[0180] The predictive model constructed based on the "causal features" screened by this invention captures the essential mechanism of disease occurrence rather than the surface phenomenon. Therefore, it has a stronger cross-center generalization ability, and interventions formulated for these causal features (such as targeting specific inflammatory pathways) are more likely to produce actual therapeutic benefits.

[0181] The embodiments of the present invention are described in detail and illustrated with specific figures and application examples. These examples demonstrate how to apply the core steps (S1-S6) of the present invention to overcome the shortcomings of existing technologies that can only reveal correlations or perform "black box" predictions. The data in the following specific examples are all derived from data of 200 SAH patients. Figure 6 This is a table showing the patient's clinical baseline data.

[0182] Case 1: Validating the causal effect of Hunt-Hess classification (clinical score) on DCI: Background: The Hunt-Hess classification is shown to be associated with DCI. But is this association causally related, or are both factors contributing to the association with other factors such as age and underlying medical conditions? This case aims to address this question.

[0183] Objective: To identify whether the Hunt-Hess classification is a causal risk factor for DCI.

[0184] Detailed steps (S1-S6): Step S1: Temporal alignment and multimodal preprocessing of clinical data from SAH patients: 1. Data loading and cleaning: Data source: Data from 200 SAH patients.

[0185] Missing value handling: No missing values.

[0186] Variable encoding: sex: F ->0, M ->1; surgical_method: clipping ->0, coiling ->1.

[0187] 2. Definition of key variables (based on anchor point mechanism): Treatment(processing variable): Binarizes hunt_hess.

[0188] Define High_HH (T=1): Hunt-Hess classification 3-5 (severe).

[0189] Define Low_HH (T=0): Hunt-Hess grade 1-2 (mild cases).

[0190] Data statistics: Among the 200 patients, there were 113 high-risk patients (T=1) and 87 low-risk patients (T=0).

[0191] Outcome (outcome variable): DCI (1 = occurred, 0 = did not occur).

[0192] Original incidence: 33% (66 / 200) overall.

[0193] Covariates (baseline / confounding variables): V base = [age, sex, hypertension, diabetes, fisher_grade].

[0194] Note: early_vasospasm and csf_il6 are intermediate process variables. V mid This step marks the graph, but it needs to be processed according to the graph structure when identified in step S3.

[0195] Step S2: Reference Figure 7 Construct a hybrid causal graph based on "medical hard constraints": 1. Define the hard constraint matrix ( M c ): Prohibition Rule 1: DCI hunt_hess (The ending cannot affect the admission status).

[0196] Prohibited rule 2: early_vasospasm hunt_hess (Complications occurred after admission).

[0197] Required rule: fisher_grade (bleeding volume) → early_vasospasm (common pathology).

[0198] 2. Restricted Structure Learning (PC-Stable Algorithm): exist M c Run the independence test under constraints (α=0.05).

[0199] Conflict detected: Pure data mining suggests a positive correlation between nimodipine_use and DCI (because the more severe the condition, the more medication is used).

[0200] Conflict resolution: Medical a priori indications suggest that nimodipine is a protective factor. Calculation ΔBIC The forced correction directions are hunt_hess→nimodipine_use and nimodipine_use→DCI (negative).

[0201] 3. Generate the final DAG: Confirm path: Age→Hunt_Hess→DCI.

[0202] Confirmed obfuscation path: Fisher_Grade→Hunt_Hess (relevance) and Fisher_Grade→DCI.

[0203] Step S3: Identification of causal effects based on graph structure: 1. Minimum Adjustment Set Identification: Based on DAG, the backdoor criterion is applied.

[0204] Target path: Hunt_Hess→DCI.

[0205] Identified minimum adjustment set Z : {age, sex, hypertension, fisher_grade}.

[0206] Note: early_vasospasm was identified as a mediator and was removed; wbc was identified as a collider (affected by both infection and stress) and was removed.

[0207] Step S4: Causal effect estimation based on the "ensemble strategy": For the SAH data distribution (severe patients have a high risk of DCI, but a worse baseline), three models were used to calculate ATE (Risk Difference) in parallel: Model A: Logistic Regression (Outcome Regression): formula: Logit(P(DCI=1)) = β 0+ β 1 HH_High + β 2 Age + β 3 Sex + β 4 Hypertension + β 5 Fishe r; Results: Coefficients β 1 HH_High =1.28, converted into marginal utility ATE reg =0.265 (95% CI: 0.15, 0.38).

[0208] Explanation: After controlling for confounding, the risk of DCI increased by 26.5% in the high-risk group.

[0209] Model B: Inverse Probability Weighted (IPW): Calculate propensity score P ( T =1∣ ZThe truncation weights are [0.05, 0.95].

[0210] result: ATE ipw =0.258.

[0211] Model C: Dual Robust Estimation (TMLE): Combine the Q-model (Outcome) and the g-model (Treatment).

[0212] result: ATE tmle =0.262.

[0213] Consistency check: Calculate the coefficient of variation: CV = σ (0.265, 0.258, 0.262) / μ≈0.01<0.2 (threshold). Validation passed. Output integration. ATE =0.262 (P<0.001).

[0214] Step S5: Causality Refutation and Robustness Quantification Rating (CCRS): To verify whether this 26.2% increase in risk is reliable, a stress test was conducted: Placebo test S placebo ): Randomly shuffle hunt_hess and repeatedly calculate ATE.

[0215] result: ATE placebo =0.003 (close to 0, P=0.85).

[0216] Score: Score 1=1 |0.003|≈1.0.

[0217] Data subset test ( S subset ): 80% of the samples were randomly selected and repeated 20 times.

[0218] Result: The ATE fluctuation range is [0.21, 0.32], and the direction is always positive.

[0219] Score: Score 2 = 0.95 (highly consistent).

[0220] Sensitivity analysis: Simulating unobserved confusion. Calculating the Gamma value. Result: Γ=1.95. This means that an unknown latent variable (not associated with Fisher's classification) that would nearly double the risk of DCI would be needed to offset the current significance. This is extremely rare in clinical practice.

[0221] Score: after normalization Score 3 = 0.85.

[0222] CCRS total score calculation: Set weight w=[0.4,0.3,0.3], CCRS=0.4×1.0+0.3×0.95+0.3×0.85=0.94.

[0223] Step S6: Causal risk factor classification output: Summary of indicators: Mean causal effect (ATE): 0.262 (26.2% risk attribution increase).

[0224] Clinical robustness score (CCRS): 0.94 (Class I).

[0225] Final judgment: The system determined that the Hunt-Hess classification (levels 3-5) was a "confirmatory causal risk factor" for DCI.

[0226] Clinical significance: For patients with Hunt-Hess > 3, even if the Fisher grade is not high, it is recommended to immediately initiate DCI prophylactic intervention (such as using nimodipine at an adequate dose), because the risk stems from the severity of the condition itself and has a clear causal driving force, rather than just being an accompanying phenomenon of bleeding.

[0227] Case Study 2: Verifying the causal effect of early vasospasm (imaging indicators) on DCI: Background: Vasospasm is considered one of the main mechanisms of DCI, but the two are highly correlated and it is difficult to distinguish whether it is causal or accompanied.

[0228] Objective: To identify whether early vasospasm (EV) is a causal risk factor for DCI.

[0229] Detailed steps (S1-S6): Step S1: Temporal alignment and multimodal preprocessing of clinical data from SAH patients: 1. Variable mapping and timing anchor point setting: Anchor point ( T 0): Admitted to hospital due to SAH onset.

[0230] Processing variables (T ): early_vasospasm(EV).

[0231] Definition: Vasospasm confirmed by TCD or CTA within 72 hours of admission (1=yes, 0=no).

[0232] Data distribution: Among the 200 patients, 72 had EV=1 and 128 had EV=0.

[0233] Outcome variables ( Y ): DCI (1 = occurred, 0 = did not occur).

[0234] Baseline confounding variable set ( V base ): age, fisher_grade (bleeding amount), smoking (smoking history), hypertension (hypertension).

[0235] Intermediate / concomitant variable: nimodipine_use. Note: This is a key variable because spasms may lead doctors to increase medication use.

[0236] 2. Preprocessing steps: Perform ordered categorical variable encoding on fisher_grade and hunt_hess.

[0237] Perform Standardization (Z-score) on the continuous variable age.

[0238] Step S2: Construct a hybrid causal graph based on "medical hard constraints": 1. Import hard constraint matrix ( M c ): Timing Forbidden: DCI early_vasospasm (The ending cannot lead to a precursor).

[0239] Anatomical requirement: fisher_grade → early_vasospasm. Medical explanation: Subarachnoid hemorrhage is a direct chemical stimulus that causes vasospasm.

[0240] 2. Constrained Structure Learning (PC-Stable): Reference Figure 8 The algorithm runs under constraints.

[0241] Conflict detection and resolution: The algorithm found a high correlation between early_vasospasm and nimodipine_use in the data.

[0242] Determine whether it's nimodipine_use → early_vasospasm (prevention) or early_vasospasm → nimodipine_use (treatment)? Resolution: Based on medical treatment guidelines (add medication if spasms occur) and combined with changes in the BIC score, the direction was determined to be early_vasospasm→nimodipine_use.

[0243] Therefore, nimodipine_use is a downstream node (Descendant) of EV.

[0244] 3. Final DAG structure: Confusion path: Fisher_Grade → EV, and Fisher_Grade → DCI. (Note: Bleeding volume causes both spasm and DCI through other mechanisms).

[0245] Causal path: EV→DCI.

[0246] Exclusion path: EV → Nimodipine → DCI (This is the mediating path; it must be included to calculate the total effect, and blocked to calculate the direct effect. This case calculates the total causal effect (ATE)).

[0247] Step S3: Identification of causal effects based on graph structure: 1. Automatic optimization of the minimum adjustment set: Utilize the backdoor criterion.

[0248] Target: P ( DCI | do ( EV )).

[0249] Identification result: The backdoor path that must be blocked is the path via Fisher_Grade and Smoking.

[0250] Minimal Adjustment Set Z : {fisher_grade, hunt_hess, age, smoking, hypertension}.

[0251] Key point: The system automatically removed nimodipine_use because it was a processed variable, and including it in the adjustment set would introduce "over-adjustment bias".

[0252] Step S4: Causal effect estimation based on the "ensemble strategy": To address the imbalance in EV distribution in the data (approximately 36% of patients have EV), an ensemble strategy was employed to calculate the risk difference: Model A: Logistic Regression (LR): Fitting P ( DCI | EV , Z ).

[0253] result: OR =4.2.

[0254] Transformation ATE (marginal effect): 0.315 (i.e., EV leads to a 31.5% increase in DCI risk).

[0255] Model B: Propensity Score Matching (PSM): Predict the probability of EV using Z as a covariate (Propensity Score). Use 1:1 matching with a caliper value of 0.2.

[0256] Post-matching sample balance check: Standardized Mean Difference (SMD) < 0.1.

[0257] result: ATE psm =0.298.

[0258] Model C: Dual Robust Estimation (TMLE): Simultaneously, both the outcome model and the treatment model are modeled.

[0259] result: ATE tmle =0.308.

[0260] Consistency check: Data vector Vest =[0.315,0.298,0.308].

[0261] Calculate the coefficient of variation CV ≈0.02<0.2.

[0262] Verification passed. Output integration. ATE = 0.307 (95% CI: 0.21 - 0.40, P<0.001).

[0263] Step S5: Causality Refutation and Robustness Quantification Rating (CCRS): Although the p-value is highly significant, we need to know whether this conclusion is "strong enough".

[0264] Placebo test Splacebo ): Generate a random variable to replace early_vasospasm.

[0265] result: ATE new = 0.01 (P=0.76).

[0266] Score: Score 1 = 1.0.

[0267] Data subset test ( S subset Bootstrap resampled 50 times.

[0268] Results: The mean ATE was 0.30, and the 95% percentile was [0.22, 0.39]. The interval did not cross 0.

[0269] Score: Score 2 = 0.98.

[0270] Sensitivity analysis ( S sensitivity ): Simulated scenario: Suppose there is an undocumented "genetic susceptibility" variable that causes both spasms and DCI.

[0271] Calculate the Gamma value: The result shows Γ=2.8.

[0272] Explanation: This means that the unobserved gene must increase the risk of DCI by 2.8 times and be strongly associated with spasms to explain the current causal effect. Given that Fisher classification is known to be controlled, the likelihood of such strong confounding is extremely low.

[0273] Score: Score 3 = 0.92.

[0274] CCRS total score: CCRS = 0.4(1.0) + 0.3(0.98) + 0.3(0.92) = 0.97.

[0275] Step S6: Causal risk factor classification output: Final judgment: Causal attribute: Confirmatory causal risk factor (Class I).

[0276] Quantitative conclusion: Early vasospasm independently increases the risk of DCI by 30.7%, and this conclusion is highly robust (CCRS=0.97).

[0277] Comparison with existing technologies (correlation analysis): Traditional view: Some literature suggests that EV is merely a marker of bleeding volume, and that EV is not significant after controlling for Fisher grade.

[0278] This invention reveals that, through rigorous backdoor adjustment and dual robust estimation, patients who experience endocardial infarction (EV) still have a significantly increased risk of stroke (DCI) even under the same Fisher classification. This indicates that EV is an indispensable key causal node in the pathway from hemorrhage to stroke.

[0279] Clinical recommendation: EVs should not be considered merely as a marker of disease severity, but rather as a direct intervention target. It is recommended that patients with EVs detected by TCD monitoring, regardless of Fisher classification, should have their preventative measures (such as hypertension treatment) escalated.

[0280] Case 3: Verifying the causal effect of the inflammatory cytokine IL-6 (a biochemical marker) on DCI: Background: Levels of inflammatory factors such as IL-6 in cerebrospinal fluid have been found to be associated with brain injury-related cognitive impairment (DCI). However, is IL-6 a "driver" of DCI or merely a "result (marker)" of brain injury? Objective: To identify whether "high levels of IL-6" is a causal risk factor for DCI.

[0281] Detailed steps (S1-S6): Step S1: Temporal alignment and multimodal preprocessing of clinical data from SAH patients: 1. Variable mapping and threshold setting: Processing variables ( T ): csf_IL-6. Continuous to binary: To facilitate causal comparison, based on the data distribution (median approximately 45-50), High_IL6 (T=1) is set to ≥50 pg / mL, and Low_IL6 (T=0) is set to <50 pg / mL.

[0282] Data distribution: Among the 200 patients, there were 108 in the High_IL6 group and 92 in the Low_IL6 group.

[0283] Outcome variables ( Y ): DCI (1 = occurred, 0 = did not occur).

[0284] Key confounding variables ( Z ): fisher_grade (bleeding volume), hunt_hess (clinical grade), early_vasospasm (early spasm).

[0285] 2. Preprocessing steps: Missing value imputation complete.

[0286] Preliminary correlation analysis: Simple statistics show that the incidence of DCI was 52% in the High_IL6 group and 12% in the Low_IL6 group. The Pearson correlation coefficient was as high as 0.45. Traditional statistics would undoubtedly classify it as an "extremely significant risk factor".

[0287] Step S2: Construct a hybrid causal graph based on "medical hard constraints": 1. Import hard constraint matrix ( M c ): Dissection must follow the rule: fisher_grade→csf_IL-6.

[0288] Medical logic: The breakdown of red blood cells in the subarachnoid space (high Fisher grade) stimulates immune cells to release IL-6. Therefore, the amount of bleeding is an upstream cause of IL-6 release.

[0289] Timing Forbidden: DCI csf_IL-6.

[0290] 2. Constrained Structure Learning (PC-Stable): The algorithm detected strong pairwise correlations (triangular relationships) among fisher_grade, csf_IL6, and DCI.

[0291] Due to the existence of hard constraints, the edge csf_IL-6→fisher_grade is prohibited.

[0292] The algorithm preserves both fisher_grade→DCI (direct path) and csf_IL-6→DCI (potential path).

[0293] 3. Final DAG structure: refer to Figure 9 The key obfuscated structure is: Fisher_Grade → IL-6 → DCI (intermediate path?) vs. Fisher_Grade → DCI (direct path).

[0294] Key point to identify: It is necessary to isolate the impact of Fisher_Grade on IL-6 and see if IL-6 still has an independent causal effect.

[0295] Step S3: Identification of causal effects based on graph structure: 1. Automatic optimization of the minimum adjustment set: Apply the backdoor criterion. (To calculate...) P ( DCI | do (IL 6) The backdoor path must be blocked: IL-6←Fisher_Grade→DCI.

[0296] Similarly, Hunt_Hess is also an upstream obfuscation.

[0297] Minimal Adjustment Set Z : {fisher_grade, hunt_hess, age, surgical_method}.

[0298] Step S4: Causal effect estimation based on the "ensemble strategy": This demonstrates the significant differences between the present invention and conventional methods.

[0299] Unadjusted (Naive Estimate - corresponding to existing technology): Calculate the risk difference directly: ATE naive =0.52 0.12 = 0.40.

[0300] Conclusion: IL-6 increases the risk of DCI by 40% (extremely overestimated).

[0301] This invention integrates estimation: Model A ( Logistic (Regression): Add Fisher and Hunt-Hess controls.

[0302] Results: The regression coefficient of IL-6 decreased significantly, and the p-value became 0.048 (marginally significant). Transformation ATE reg =0.08.

[0303] Model B ( IPW ): Patients with higher Fisher classification tend to have higher IL-6 levels; after weighting, the difference between the two groups narrowed.

[0304] result: ATE ipw =0.06.

[0305] Model C(TMLE): Results: ATE tmle =0.07.

[0306] Consistency check: The three model results [0.08, 0.06, 0.07] have very small coefficients of variation, indicating that they are consistent.

[0307] integrated ATE =0.07 (7% risk increase).

[0308] Interim conclusion: After controlling the bleeding volume, the pathogenicity of IL-6 plummeted from 40% to 7%. Although it was still statistically significant (P<0.05), the effect size was very weak.

[0309] Step S5: Causality Refutation and Robustness Quantification Rating (CCRS): To determine whether this 7% effect warrants clinical intervention, a CCRS score was used.

[0310] Placebo test S placebo ): Randomly replace IL-6.

[0311] result: ATE placebo ≈0.

[0312] Score: Score 1 = 1.0 (Passed).

[0313] Data subset test ( S subset ): Resample 20 times.

[0314] Results: Due to the small effect size (0.07), the ATE dropped to 0.01 or even -0.02 in some subsets, with the 95% confidence interval crossing 0.

[0315] Score: Score 2 = 0.4 (Poor stability, indicating that the results are highly dependent on specific samples).

[0316] Sensitivity analysis ( S sensitivity ): Simulated scenario: Are there unobserved confounding events (such as systemic inflammatory response syndrome, SIRS)? Calculate the Gamma value: The result shows Γ=1.15.

[0317] Explanation: This means that by introducing even a very weak unknown confounding factor (increasing the risk by 1.15 times), a marginally significant p-value can become insignificant. This is extremely common in complex clinical settings.

[0318] Score: Score 3 = 0.2 (extremely fragile).

[0319] CCRS total score: CCRS = 0.4(1.0) + 0.3(0.4) + 0.3(0.2) = 0.58.

[0320] Step S6: Causal risk factor classification output: Summary of indicators: Mean causal effect (ATE): 0.07 (only 7%, and marginally significant).

[0321] Clinical robustness score (CCRS): 0.58 (below the threshold of 0.6).

[0322] Final judgment: The system determined that CSF IL-6 was a "pseudo-related factor / accompanying biomarker (Class III)".

[0323] Results Interpretation and Technical Advantages Demonstration: Misjudgment by existing technology: If only traditional correlation analysis or regression without sensitivity analysis is used, IL-6 may be falsely reported as an important target because P<0.05.

[0324] This invention provides insight: The CCRS score (0.58) reveals the fragility of this causal relationship. This explains why IL-6 antagonists (such as tocilizumab) have been ineffective in clinical trials for preventing DCI—because IL-6 is more of a “shadow” of bleeding severity than a “driving force” of DCI.

[0325] Clinical decision support: Physicians are advised not to make lowering IL-6 levels the primary treatment goal for preventing DCI, but to focus on managing upstream bleeding (Fisher grade) or confirmatory Hunt-Hess classification.

[0326] This application also provides a robust screening device for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning, which can implement the above method. The device includes: The first module involves acquiring SAH clinical data. The second module performs time-series alignment and multimodal preprocessing on SAH clinical data to obtain preprocessed data. The third module sets at least one candidate risk factor to be analyzed as a treatment variable; The fourth module establishes a hybrid causal graph based on a medical hard constraint matrix, using preprocessed data and processing variables. The fifth module identifies causal effects based on mixed causal graphs, determines whether the causal relationship between treatment variables and outcome variables is identifiable, and outputs the minimum confounding variable adjustment set. The sixth module, based on the minimum confounding variable adjustment set, employs an ensemble strategy to numerically estimate the causal effect between the treatment variable and the outcome variable, runs at least two types of heterogeneous estimators in parallel, and calculates the average causal effect value output by each estimator. The seventh module performs a consistency check on the average causal effect values ​​output by each estimator. If the check passes, it outputs the fused average causal effect value and its confidence interval. The eighth module performs a multidimensional rebuttal test on the fused mean causal effect value and calculates the clinical causal robustness score based on the results of the multidimensional rebuttal test. The ninth module, based on the confidence interval of the fused average causal effect value and the clinical causal robustness score, classifies the candidate risk factors corresponding to the treatment variables into confirmatory causal factors, potential causal factors, or spurious correlation factors.

[0327] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0328] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0329] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0330] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0331] The robust screening method, apparatus, electronic device, storage medium, and program product for delayed cerebral ischemia after SAH based on hybrid causal structure learning provided in this application have good technical effects: ① It has achieved a leap from "subjective experience / pure data mining" to "medical constraint-guided hybrid structure learning", effectively reducing the error rate of causal mapping.

[0332] This invention introduces a hard constraint matrix containing "temporal prohibition rules" and "anatomical necessity rules" to limit the search space of data-driven algorithms. This "hybrid construction strategy" retains the ability of data mining to discover potential new associations while forcibly correcting logical fallacies that violate medical common sense (such as "outcome points to cause"), significantly improving the accuracy and medical compliance of causal graph structures.

[0333] ② It overcomes the bias of single statistical models on imbalanced medical data and improves the accuracy of causal effect estimation.

[0334] This invention employs a "multi-model ensemble estimation strategy," simultaneously integrating parametric (regression), non-parametric (IPW), and semi-parametric (dual robust TMLE) models. Through a consistency verification mechanism (coefficient of variation screening), only results from convergent multi-model conclusions are retained. This gives the estimation results "dual robustness," meaning that as long as either the outcome model or the treatment allocation model is correctly defined, an unbiased estimate can be obtained, greatly improving the reliability of the conclusions.

[0335] ③ The first "Clinical Causality Robustness Score (CCRS)" was created, which solved the problem that traditional statistical results lacked intuitive clinical interpretation.

[0336] This invention establishes a comprehensive quantitative evaluation index—the CCRS score—that includes placebo testing, consistency of data subset resampling, and sensitivity to unobserved confounding. This score transforms abstract robustness into a concrete numerical value between 0 and 1, intuitively informing physicians: "How strong of the unknown interference is needed to overturn the current conclusion?" This provides a grading basis for clinical decision-making (e.g., confirmatory factors vs. latent factors), filling the gap in evaluation standards between causal inference theory and clinical application.

[0337] ④ Eliminating the "spurious correlation" feature at the source significantly enhances the generalization ability of subsequent DCI prediction models and intervention strategies.

[0338] This invention is not only an analytical tool, but also a "pre-filter for feature engineering." By rigorously eliminating confounding factors and colliding variables, the features selected by this invention possess causal stability across populations and centers. Subsequent predictive models or interventions built based on these causal features target the "mechanism" of disease occurrence rather than its "appearance," thus possessing stronger generalization ability and practical clinical intervention value.

[0339] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0340] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A robust screening method for risk features of delayed cerebral ischemia after SAH based on hybrid causal structure learning, characterized in that, The method includes the following steps: Obtain SAH clinical data; The SAH clinical data were subjected to time-series alignment and multimodal preprocessing to obtain preprocessed data; At least one candidate risk factor to be analyzed is set as a treatment variable; Based on the preprocessed data and the processing variables, a hybrid causal graph based on a medical hard constraint matrix is ​​established; Based on the hybrid causal graph, causal effect identification is performed to determine whether the causal relationship between the treatment variable and the outcome variable is identifiable, and the minimum confounding variable adjustment set is output. Based on the minimum confounding variable adjustment set, an integrated strategy is used to numerically estimate the causal effect between the treatment variable and the outcome variable, and at least two types of heterogeneous estimators are run in parallel to calculate the average causal effect value output by each estimator. The average causal effect value output by each estimator is checked for consistency. If the check passes, the fused average causal effect value and its confidence interval are output. A multidimensional refutation test was performed on the fused average causal effect value, and a clinical causal robustness score was calculated based on the results of the multidimensional refutation test. Based on the confidence interval of the fused average causal effect value and the clinical causal robustness score, the candidate risk factors corresponding to the treatment variables are classified and output as confirmatory causal factors, potential causal factors, or spurious correlation factors.

2. The method according to claim 1, characterized in that, The SAH clinical data undergoes time-series alignment and multimodal preprocessing to obtain preprocessed data, including: Using the onset time of SAH as the timeline, a time axis was established, and all variables in the SAH clinical data were divided into a set of baseline variables, a set of intermediate process variables, and an outcome variable; the outcome variable included the occurrence status of delayed cerebral ischemia. Missing value processing, outlier cleaning, and standardized coding are performed on the data in the baseline variable set, the intermediate process variable set, and the outcome variable to obtain preprocessed data.

3. The method according to claim 2, characterized in that, Setting at least one candidate risk factor to be analyzed as a treatment variable includes: At least one variable is selected from the baseline variable set or the intermediate process variable set as a treatment variable; the selected treatment variable is the candidate risk factor to be analyzed.

4. The method according to claim 1, characterized in that, The step of establishing a hybrid causal graph based on a medical hard constraint matrix according to the preprocessed data and the processed variables includes: Based on the preprocessed data and prior medical knowledge, a medical hard constraint matrix is ​​defined. In the medical hard constraint matrix, by setting specific values ​​as prohibition symbols, prohibition rules for connecting variables are defined. The prohibition rules include at least: prohibiting the outcome variable from pointing to any variable in the baseline variable set. Based on the aforementioned medical hard constraint matrix, a constraint-based causal discovery algorithm is used to perform a conditional independence test on the preprocessed data, generating an initial directed acyclic graph. Edges in the initial directed acyclic graph that conflict with weak medical prior knowledge are marked as edges to be adjudicated. For each edge to be decided, calculate the change in the Bayesian information criterion value of the model as a whole when the edge is retained and removed in the initial directed acyclic graph. When the absolute value of the change exceeds the preset resolution threshold, the edge to be decided is retained; otherwise, the edge to be decided is removed, and the result of conflict resolution is obtained. Based on the conflict resolution results, the initial directed acyclic graph is adjusted to obtain a hybrid causal graph based on a medical hard constraint matrix.

5. The method according to claim 1, characterized in that, The step of identifying causal effects based on the mixed causal graph, determining whether the causal relationship between the treatment variable and the outcome variable is identifiable, and outputting a minimum confounding variable adjustment set includes: On the topology of the hybrid causal graph, based on the backdoor criterion or the frontdoor criterion, all connection paths are traversed and analyzed to determine whether there is a set of observed variables; the set of observed variables is used to block all non-causal paths from the processing variable to the outcome variable; If the set of observed variables does not exist, it is determined to be unidentifiable, and the subsequent analysis of the current causal relationship is terminated; If the set of observed variables exists, the current causal relationship is determined to be identifiable. Then, nodes identified as colliding variables and mediating variables by preset rules are excluded from the set of observed variables to obtain the minimum confusion variable adjustment set.

6. The method according to claim 1, characterized in that, Based on the minimum confounding variable adjustment set, an ensemble strategy is used to numerically estimate the causal effect between the treatment variable and the outcome variable, running at least two types of heterogeneous estimators in parallel, and calculating the average causal effect value output by each estimator, including: Based on the adjusted set of minimum confounding variables, an ensemble strategy is used to numerically estimate causal effects and train a heterogeneity estimator. The heterogeneity estimator includes, but is not limited to, a parametric estimator based on an outcome model, a nonparametric estimator based on a treatment allocation model, and a dual robust estimator. The parametric estimator based on the outcome model is used to fit the conditional probability distribution of the outcome variable given the treatment variable and the adjusted set of minimum confounding variables. The nonparametric estimator based on the treatment allocation model is used to estimate the propensity score of the treatment variable given the adjusted set of minimum confounding variables, and to perform weighted or matched calculations based on this score. The dual robust estimator combines information from both the parametric estimator based on the outcome model and the nonparametric estimator based on the treatment allocation model for estimation. Run at least two types of heterogeneity estimators in parallel, and calculate and output the average causal effect value estimated by each of the heterogeneity estimators.

7. The method according to claim 1, characterized in that, The consistency check of the average causal effect value output by each estimator is performed. If the check passes, the fused average causal effect value and its confidence interval are output, including: Calculate the coefficient of variation among the average causal effect values ​​output by different estimators; The coefficient of variation is compared with a preset consistency threshold, and the effect direction of each average causal effect value is checked to see if they are consistent. If the coefficient of variation is less than the consistency threshold and the effect direction of all average causal effect values ​​is consistent, the verification is deemed to have passed. When the verification is successful, a weighted average is calculated using the reciprocal of the variance of each estimate as the weight, to obtain the fused average causal effect value and its confidence interval.

8. The method according to claim 1, characterized in that, The process of performing a multidimensional rebuttal test on the fused average causal effect value and calculating a clinical causal robustness score based on the results of the multidimensional rebuttal test includes: The processing variable in the original data is replaced with a randomly generated placebo variable; the placebo variable is a variable that has no causal relationship with the outcome variable. Under the minimum confounding variable adjustment set, the heterogeneity estimator is rerun to calculate the effect value of the placebo variable on the outcome variable, which is used as the placebo effect value. Multiple data subsets are generated from the original data through random sampling with replacement. On each of the data subsets, the heterogeneity estimator is run independently to perform the causal effect identification and estimation process, calculate the fused average causal effect value, and obtain a subset effect estimation value sequence; Set a strong confounding factor that is not recorded in the data, and set the strength of its influence on both the processing variable and the outcome variable. Gradually increase the influence intensity of the strong confounding factor, calculate and determine the minimum influence intensity required to make the statistical significance of the fused average causal effect value disappear, and use it as the sensitivity threshold. Based on the placebo effect value, the subset effect estimate sequence, and the sensitivity threshold, the placebo test score, stability score, and sensitivity score are calculated using a preset scoring rule. Based on preset weighting coefficients, the placebo test score, the stability score, and the sensitivity score are weighted and summed to calculate the final clinical causal robustness score; the clinical causal robustness score is a scalar value between 0 and 1.

9. The method according to claim 1, characterized in that, The step of classifying and outputting candidate risk factors corresponding to the treatment variable as confirmatory causal factors, potential causal factors, or spurious correlation factors based on the confidence interval of the fused average causal effect value and the clinical causal robustness score includes: A first scoring threshold and a second scoring threshold are preset; the first scoring threshold is higher than the second scoring threshold. Based on the confidence interval of the fused average causal effect value and the clinical causal robustness score, the following judgment rule is applied: If the confidence interval of the fused average causal effect value does not contain zero at a preset significance level, and the clinical causal robustness score is greater than or equal to the first score threshold, then the candidate risk factors in the processing variables will be identified and output as confirmatory causal factors. If the confidence interval of the fused average causal effect value does not contain zero at a preset significance level, and the clinical causal robustness score is less than the first scoring threshold but greater than or equal to the second scoring threshold, then the candidate risk factors in the processing variables will be identified and output as potential causal factors. If the confidence interval of the fused mean causal effect value contains zero, or the clinical causal robustness score is less than the second score threshold, then the candidate risk factors in the treatment variables will be identified and output as spurious correlation factors.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 9.

Images