Causal inference-based evaluation and prediction method for intervention effect of chronic disease

Abstract

The present application relates to medical care informatics, and discloses a kind of based on causal inference's chronic disease intervention effect evaluation and prediction method, comprising: obtaining long-period medical record to construct physiology and intervention sequence matrix, the conditional probability value of each time observation step intervention occurrence is calculated, the clinical contraindication threshold matrix corresponding to physiological state is determined, if physiological state falls into forbidden interval, the probability reciprocal calculation path is blocked, the intervention frequency deviation residual in historical stationary window is extracted to cover time-varying confounding adjustment weight, according to weight, sequence matrix is resampled to generate quasi-state equilibrium sequence, mapping output evolution track, the present application uses residual compensation mechanism to hedge the calculation divergence risk caused by the probability extreme value induced by guideline constraint, guarantee data time sequence integrity under critical condition, eliminate survivor bias.

Description

Technical Field

[0001] This invention belongs to the field of healthcare informatics technology, and in particular relates to a method for evaluating and predicting the effects of chronic disease interventions based on causal inference. Background Technology

[0002] Currently, building healthcare information systems capable of assessing the net effect of interventions is key to improving the quality of decision support. Such systems eliminate time-varying confounding effects between physiological states and interventions by processing longitudinal electronic medical records containing physiological state characteristics and treatment sequence.

[0003] However, existing technologies still face challenges in processing such complex medical data. Traditional medical insurance policy simulations or health management technologies mostly focus on macro-level simulations such as population structure and population health, with coarse granularity, making it difficult to provide dynamic and quantitative explanations of individual medical choices. At the control algorithm level, existing methods for quantifying health behavior also have shortcomings. For example, Chinese invention patent application CN116994764A discloses a method, device, electronic device, and storage medium for constructing a health behavior model. It simulates health behavior and predicts medical insurance expenditures by constructing a utility function containing feature vectors such as medication cost, medication effect, and drug supply, and by minimizing a loss function to update the feature parameters. However, the actual clinical environment is constrained by… Medical guidelines stipulate that when a patient's physiological indicators reach the medical safety boundary, intervention actions may abruptly change due to safety constraints. This rigid constraint causes the probability of a specific intervention in the underlying data to tend to 0 or 1. Conventional assessment methods require the intervention probability to be greater than zero. The clinical guideline constraint conflicts with this positive definiteness requirement, causing the denominator to tend to a minimum when calculating the inverse probability weight, leading to a numerical explosion in the calculation matrix and causing the assessment link to collapse. Existing technologies mostly adopt a preprocessing method of removing samples that reach the boundary. This processing method destroys the temporal integrity of longitudinal medical information, causing the decision support system to lose its assessment capability under critical conditions such as disease deterioration, resulting in survivor bias. There is a lack of effective solutions in the industry that maintain the coherence of data throughout the entire disease course while ensuring computational stability.

[0004] Therefore, the technical problem to be solved by this invention is how to maintain the numerical stability and data integrity of the long-term intervention assessment process while adhering to the rigid constraints of clinical guidelines. Summary of the Invention

[0005] This invention aims to address the problem of numerical explosion and assessment bias in causal inference models caused by the rigid constraints of clinical medical guidelines under conditions of worsening disease.

[0006] In this technical solution, a method for evaluating and predicting the effectiveness of chronic disease interventions based on causal inference includes the following steps: Step S101: Obtain a physiological-intervention sequence matrix containing physiological state feature components and current intervention action codes. The physiological-intervention sequence matrix is ​​composed of time-series medical data records of the target patient. Step S102: Calculate the conditional probability of the current intervention action occurring within each time observation step in the physiological-intervention sequence matrix; Step S103: Determine the clinical contraindication threshold matrix corresponding to the physiological state characteristic components, wherein the clinical contraindication threshold matrix defines the extreme value range of physiological indicators that are mutually exclusive with the current intervention action; Step S104: Based on the mapping results between physiological state feature components and the extreme value range of physiological indicators, determine the time-varying hybrid adjustment weight corresponding to the current time observation step, wherein: when the physiological state feature components do not fall into the extreme value range of physiological indicators, the reciprocal of the conditional occurrence probability is determined as the time-varying hybrid adjustment weight; when the physiological state feature components fall into the extreme value range of physiological indicators, the calculation link from the conditional occurrence probability to the time-varying hybrid adjustment weight is blocked, the intervention frequency deviation residual within the preset sample smoothing window in the physiological-intervention sequence matrix is ​​extracted, and the intervention frequency deviation residual is used to cover the time-varying hybrid adjustment weight. Step S105: Based on the determined time-varying hybrid adjustment weights of each time observation step, the physiological-intervention sequence matrix is ​​logically resampled to generate a pseudo-equilibrium sequence. Step S106: Perform counterfactual state mapping on the pseudo-equilibrium sequence and output the evolution trajectory of the physiological state of chronic diseases.

[0007] Preferably, step S103 includes the following sub-steps: step S1031, extracting rigid contraindications for specific chronic diseases from the digital clinical guideline library; step S1032, converting the rigid contraindications into logical trigger thresholds within the electronic computing device; step S1033, constructing a clinical contraindication threshold matrix in a multi-dimensional probability space based on the logical trigger thresholds, the clinical contraindication threshold matrix being used to lock the blocking of the computing link in step S104.

[0008] Preferably, in step S101, the time-series medical data recording includes electronic medical record data, physiological parameter monitoring streams, and medication adherence records; the physiological state characteristic components include blood glucose fluctuation rate, blood pressure coefficient of variation, and biochemical parameters reflecting organ function; the current intervention action code includes drug dosage adjustment indicators, dietary intervention intensity, and exercise prescription compliance rate.

[0009] Preferably, step S102 includes: using a structural marginal model to analyze the coupling relationship between physiological state feature components and subsequent current intervention actions; determining the induced response generated by the historical evolution of physiological state feature components to the current intervention action, and using the induced response as a bias term for the conditional occurrence probability.

[0010] Preferably, in step S104, when using the intervention frequency deviation residual to cover the time-varying hybrid adjustment weight, the method further includes: step S1041, real-time transmission of the probability of the occurrence of the masking condition to the time-varying hybrid adjustment weight; step S1042, locking the upper limit of the value of the time-varying hybrid adjustment weight to limit the divergence of the calculation matrix.

[0011] Preferably, in step S105, the process of generating the pseudo-equilibrium sequence includes: applying time-varying confounding adjustment weights to the physiological-intervention sequence matrix, and constructing a counterfactual deduction space that is logically decoupled from the time-varying confounding factors by performing distribution recalibration on the observed samples.

[0012] Preferably, in step S106, outputting the evolution trajectory of the physiological state of chronic disease includes: fitting the evolution trend of the physiological state on the mimicry equilibrium sequence according to the preset candidate intervention path; calculating the average therapeutic effect of the target intervention program relative to the baseline program, and outputting the intervention effect evaluation curve containing confidence intervals.

[0013] Preferably, the method further includes: performing survivor bias correction on sample points that meet the clinical contraindication threshold matrix; and retaining high-risk samples through the weight coverage mechanism in step S104 to eliminate prediction baseline drift caused by sample removal.

[0014] Preferably, after step S106, the method further includes: comparing the evolution trajectory of the physiological state of chronic diseases with the real-time acquired physiological parameter monitoring stream; if the deviation between the two exceeds a preset deviation threshold, resetting the model parameters used to calculate the probability in step S102 to achieve online optimization of the evaluation logic.

[0015] Compared with existing technologies, the chronic disease intervention effect evaluation and prediction method based on causal inference of this invention has the following advantages: 1. In the evaluation of the effectiveness of chronic disease intervention, by deeply coupling the clinical contraindication boundary matrix with the time-series calculation link, the system can automatically trigger the conditional truncation of the logical path when it determines that the physiological state characteristic components touch the medical safety red line. This prevents the risk of numerical explosion of inverse probability weights caused by the extreme value of intervention probability induced by the rigid constraints of clinical guidelines, and ensures that the long-term evaluation matrix maintains the numerical convergence and computational stability at the physical level under complex working conditions.

[0016] 2. This in-situ compensation mechanism based on discrete residuals eliminates the need to remove high-risk or severe observation nodes to maintain system operation during the assessment process, thereby ensuring the temporal integrity of longitudinal electronic medical record data, eliminating sample selection bias and survivor bias caused by human intervention, and enabling the intervention effect prediction trajectory output by the system to cover the entire disease cycle, effectively solving the problem of population prediction benchmark drift that is common in the long-term dynamic management of chronic diseases.

[0017] 3. This invention orthogonally injects the discrete variance between the actual intervention action and its expected probability into the deduction process, successfully transforming the non-standard intervention behavior taken by clinicians for a specific individual into a proxy parameter for measuring the patient's own implicit characteristics. It achieves the goal of improving the logical purity of the quantitative indicators of the intervention net effect by simply rearranging the existing medical information flow topology without introducing additional hardware components such as external environmental sensors, and by removing the temporal confounding bias caused by unobservable environmental factors. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intervention effect evaluation of the clinical contraindication threshold constraint and mimicry sequence generation of the present invention; Figure 2 This is a system architecture diagram of the residual in-situ compensation and counterfactual mapping logic of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0021] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0022] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0023] A method for evaluating and predicting the effectiveness of chronic disease interventions based on causal inference, comprising the following steps: Step S101: Obtain a physiological-intervention sequence matrix containing physiological state feature components and current intervention action codes. The physiological-intervention sequence matrix is ​​composed of time-series medical data records of the target patient. Step S102: Calculate the conditional probability of the current intervention action occurring within each time observation step in the physiological-intervention sequence matrix; Step S103: Determine the clinical contraindication threshold matrix corresponding to the physiological state characteristic components, wherein the clinical contraindication threshold matrix defines the extreme value range of physiological indicators that are mutually exclusive with the current intervention action; Step S104: Based on the mapping results between physiological state feature components and the extreme value range of physiological indicators, determine the time-varying hybrid adjustment weight corresponding to the current time observation step, wherein: when the physiological state feature components do not fall into the extreme value range of physiological indicators, the reciprocal of the conditional occurrence probability is determined as the time-varying hybrid adjustment weight; when the physiological state feature components fall into the extreme value range of physiological indicators, the calculation link from the conditional occurrence probability to the time-varying hybrid adjustment weight is blocked, the intervention frequency deviation residual within the preset sample smoothing window in the physiological-intervention sequence matrix is ​​extracted, and the intervention frequency deviation residual is used to cover the time-varying hybrid adjustment weight. Step S105: Based on the determined time-varying hybrid adjustment weights of each time observation step, the physiological-intervention sequence matrix is ​​logically resampled to generate a pseudo-equilibrium sequence. Step S106: Perform counterfactual state mapping on the pseudo-equilibrium sequence and output the evolution trajectory of the physiological state of chronic diseases.

[0024] Preferably, step S103 includes the following sub-steps: step S1031, extracting rigid contraindications for specific chronic diseases from the digital clinical guideline library; step S1032, converting the rigid contraindications into logical trigger thresholds within the electronic computing device; step S1033, constructing a clinical contraindication threshold matrix in a multi-dimensional probability space based on the logical trigger thresholds, the clinical contraindication threshold matrix being used to lock the blocking of the computing link in step S104.

[0025] Preferably, in step S101, the time-series medical data recording includes electronic medical record data, physiological parameter monitoring streams, and medication adherence records; the physiological state characteristic components include blood glucose fluctuation rate, blood pressure coefficient of variation, and biochemical parameters reflecting organ function; the current intervention action code includes drug dosage adjustment indicators, dietary intervention intensity, and exercise prescription compliance rate.

[0026] Preferably, step S102 includes: using a structural marginal model to analyze the coupling relationship between physiological state feature components and subsequent current intervention actions; determining the induced response generated by the historical evolution of physiological state feature components to the current intervention action, and using the induced response as a bias term for the conditional occurrence probability.

[0027] Preferably, in step S104, when using the intervention frequency deviation residual to cover the time-varying hybrid adjustment weight, the method further includes: step S1041, real-time transmission of the probability of the occurrence of the masking condition to the time-varying hybrid adjustment weight; step S1042, locking the upper limit of the value of the time-varying hybrid adjustment weight to limit the divergence of the calculation matrix.

[0028] Preferably, in step S105, the process of generating the pseudo-equilibrium sequence includes: applying time-varying confounding adjustment weights to the physiological-intervention sequence matrix, and constructing a counterfactual deduction space that is logically decoupled from the time-varying confounding factors by performing distribution recalibration on the observed samples.

[0029] Preferably, in step S106, outputting the evolution trajectory of the physiological state of chronic disease includes: fitting the evolution trend of the physiological state on the mimicry equilibrium sequence according to the preset candidate intervention path; calculating the average therapeutic effect of the target intervention program relative to the baseline program, and outputting the intervention effect evaluation curve containing confidence intervals.

[0030] Preferably, the method further includes: performing survivor bias correction on sample points that meet the clinical contraindication threshold matrix; and retaining high-risk samples through the weight coverage mechanism in step S104 to eliminate prediction baseline drift caused by sample removal.

[0031] Preferably, after step S106, the method further includes: comparing the evolution trajectory of the physiological state of chronic diseases with the real-time acquired physiological parameter monitoring stream; if the deviation between the two exceeds a preset deviation threshold, resetting the model parameters used to calculate the probability in step S102 to achieve online optimization of the evaluation logic.

[0032] Example 1: For a specific long-term clinical decision support system for diabetes, when evaluating the net intervention effect of a particular oral hypoglycemic agent in patients at risk of chronic renal insufficiency, clinical guidelines require that the patient's glomerular filtration rate be below 30 mL / min / 1.73 mL / min. 2When the use of this type of drug is stopped, the conditional probability of the intervention tends to 0 in the inverse probability weighted calculation, causing the weight values ​​to explode and the entire evaluation matrix calculation to diverge. At this time, the system obtains the physiological-intervention sequence matrix containing physiological state feature components and the current intervention action code according to step S101, and simultaneously determines the clinical contraindication threshold matrix corresponding to the physiological state feature components according to step S103. The clinical contraindication threshold matrix defines the extreme value range of physiological indicators that are mutually exclusive with the current intervention action.

[0033] As the patient's disease progresses and the physiological state characteristics fall into the extreme range of physiological indicators, the system identifies that the conditional probability of intervention cannot meet the positive definiteness assumption. It then triggers conditional truncation of the logical path and blocks the original reciprocal calculation link of the conditional probability. Instead, it extracts the intervention frequency deviation residual for the patient within the historical stable sampling window according to step S104. This intervention frequency deviation residual is used to update the time-varying mixed adjustment weights corresponding to the current time observation step in situ, thereby avoiding numerical overflow caused by the denominator tending to a minimum value. This ensures the numerical convergence and stability of the long-term assessment matrix under critical illness boundary conditions. To mathematically bridge the gap between residuals and inverse probability weights... To address the dimensional differences between weights and ensure the statistical convergence of the causal inference equation, the system updates weights in situ using the intervention frequency deviation residuals, rather than directly replacing them with equivalent values. Instead, it achieves this by constructing an exponential mapping transformation logic. In practice, the system inputs the absolute value of the extracted intervention frequency deviation residuals into an exponential activation function with the natural constant as its base, converting it into a dimensionless positive scalar compensation coefficient. This coefficient is then added to the basic constant 1 to generate the upper limit of the pseudo-inverse probability weight under the truncation state. Thus, in terms of both physical and mathematical logic, the degree of specific fluctuations deviating from conventional treatment guidelines is rigorously mapped into scalar weight inputs that can be used to balance the distribution of covariates.

[0034] According to step S105, the system performs logical resampling on the original physiological-intervention sequence matrix using time-varying mixed adjustment weights. The resulting pseudo-equilibrium sequence removes the causal relationship between physiological state and intervention action without removing high-risk observation nodes. Finally, step S106 performs counterfactual state mapping on the pseudo-equilibrium sequence and outputs risk parameters containing the trajectory of physiological state evolution. These risk parameters cover the entire disease cycle from the compensated stage to the decompensated stage, eliminating survivor bias caused by the removal of samples that touch the medical safety red line in traditional methods. This enables the healthcare information system to maintain the rigor of temporal causal chain mapping in the real-world sparse and guideline-constrained medical data environment.

[0035] Example 2: In an experiment verifying the accuracy of assessing the progression of diabetes mellitus complicated with chronic renal insufficiency, the goal was to determine the assessment bias rate when the system reached clinical safety thresholds under physiological conditions. The experimental data was sourced from a publicly available electronic medical record database conforming to de-identification standards, containing longitudinal follow-up records of 3500 observation steps. To construct a stress testing environment closely resembling clinical practice, the experimental platform injected a root mean square amplitude of 2.5 mL / min / 1.73 m into the glomerular filtration rate measurement channel. 2 Random Gaussian noise is used to simulate the interference of physiological electrical signal fluctuations in medical equipment acquisition, with a sampling period of... The determination depends on the slope of the physiological index decay. Computational load of the model The trade-off function between them, and its judgment rule stipulates that when the target patient The absolute value exceeds the dynamic evolution threshold of 1.2 mL / min / 1.73 m 2 Monthly sampling cycle The system tends to use a lower limit of 14d to capture transient causal jumps, while a higher limit of 90d is selected during periods of stable physiological conditions to reduce data redundancy. In the specific underlying calculation logic, this trade-off function is configured as an inverse proportional adaptive mapping. The computing device extracts the absolute value of the continuous evolution slope of physiological indicators within the recent sampling window in real time, calculates its ratio with a preset dynamic evolution threshold, multiplies this ratio by the cosine scaling factor of the system's underlying model to construct an adjustment term, and then subtracts the product of this adjustment term and the range of values ​​from the upper limit of 90d. After rounding the calculation result down, continuous dynamic numerical output and distribution within the preset sampling period can be achieved.

[0036] The experimental procedure selected individuals with a glomerular filtration rate below 30 mL / min / 1.73 m³ / min. 2 As a rigid contraindication indicator for specific oral hypoglycemic drug interventions, a clinical contraindication threshold matrix for this drug was constructed. At the 18th month of disease progression, the physiological characteristic component measurement value of the target patient was 28.4 mL / min / 1.73 m 2If the value falls within the extreme range of the physiological index defined by the clinical contraindication threshold matrix, the control group uses an inverse probability weighting method to calculate the conditional occurrence probability of the current intervention action in each time observation step. Since the clinical guidelines truncate the drug administration action in this state, the measured value of its conditional occurrence probability is 0.0015, which causes the corresponding time-varying weight value to increase to 666.67 and triggers the divergence of the evaluation matrix calculation. The sample group of this invention identifies the extreme state of the physiological index according to step S104 and triggers the conditional truncation, extracts the intervention frequency deviation residual of the patient in the historical stable sampling window, and the residual measurement value is 0.082. The time-varying mixed adjustment weight is updated in situ using the residual, so that the adjustment weight of this observation step is stabilized at 1.082. According to step S105, the original physiological-intervention sequence matrix is ​​logically resampled using the weight, thereby generating a pseudo-equilibrium sequence.

[0037] To verify the rationality of the parameter boundaries, an out-of-range control group was set up in the experiment, and the clinical contraindication threshold was set at 60 mL / min / 1.73 m 2 Experimental data recordings show that when the physiological characteristic component is at 45.2 mL / min / 1.73 m 2 In the control group, due to premature truncation of the triggering condition, intervention response samples were lost in the mimicry equilibrium sequence. The mean square error of the risk parameter output by the counterfactual state mapping increased from 3.15 to 15.42. The appearance of this performance inflection point confirms the correlation between the clinical contraindication threshold matrix and the connotation of medical guidelines. By comparing the prediction trajectory of the sample group of this invention with that of the partially missing control group, the prediction deviation of the sample group of this invention for the time of entry into the decompensation period was 1.4 months, while the control group without the residual compensation mechanism suffered from survivor bias due to the removal of severe cases. Its prediction result deviated from the actual disease course by 6.8 months. The conclusion of this experiment confirms the effectiveness of the in-situ compensation mechanism based on discrete residuals in ensuring the temporal integrity of long-term medical data, enabling the evaluation of intervention effect to produce technical indicators with reference value under the condition of touching the red line of medical safety.

[0038] Example 3: A long-term follow-up system for diabetic patients with hypertension complications, using computing devices to obtain indicators including glycated hemoglobin. Systolic blood pressure index The time-series medical data records, including the current intervention action code containing angiotensin-converting enzyme inhibitors, are hash-mapped to de-identify the time-series medical data records. A de-identification algorithm is used to convert patient names and social identification codes into logical indexes to isolate natural person privacy information. Based on step S102, a probability mapping logic is constructed, and normalized physiological state feature components are injected into a logistic regression model to calculate the conditional probability of the current intervention action occurring within each time observation step. Its calculation process utilizes a preset feature weight vector in memory. Physiological state characteristic components The linear weighted sum, substituted into the logistic activation function, is calculated as follows: ,in, This represents the conditional probability, which ranges from 0 to 1. It is a vector composed of physiological state feature components; The feature weight vector; subscript This represents the matrix transpose operation; It is a natural constant.

[0039] Simultaneously, the system accesses the digital clinical guideline database according to step S103, uses its pre-converted triplet mapping rules to determine the clinical medication guidelines for the corresponding angiotensin-converting enzyme inhibitor, retrieves the discontinuation criteria for systolic blood pressure below 90 mmHg, and determines the extreme value range of physiological indicators as [0, 90). When the target patient's systolic blood pressure measurement falls into this extreme value range, the computing device identifies that the current intervention action is restricted and triggers conditional truncation to stop calculating the inverse probability weight. According to step S104, the discrete residual between the patient's average intervention frequency and actual intervention action within the previous stable sampling window is extracted, and the time-varying hybrid adjustment weight of the current time observation step is updated using this discrete residual. By constructing a covariate equilibrium pseudo-population based on the inverse probability weight for individual observations, the computing device constructs a diagonal weight matrix containing the time-varying hybrid adjustment weight of all time observation steps, extracts the physiological state vector and intervention action vector from the physiological-intervention sequence matrix, performs a dot product operation between the diagonal weight matrix and the physiological-intervention sequence matrix, and outputs a weighted correction matrix to establish a pseudo-equilibrium sequence. Based on the parameters of the generalized estimation equation... The convergence mechanism utilizes the known sample covariance structure to solve the longitudinal data regression coefficients, retrieves the pseudo-equilibrium sequence as the input boundary, and executes the Newton-Raphson iterative algorithm to solve the expected correlation equation of the marginal structure model for each candidate intervention path pre-read into memory. It calculates the expected predicted values ​​of the physiological state feature components corresponding to each future time observation step under the current candidate intervention path. Specifically, the mathematical form of the expected correlation equation of the marginal structure model is constructed as a generalized linear link function. Its input variable dimension includes a two-dimensional matrix tensor composed of the normalized physiological parameters of each time observation step in the pseudo-equilibrium sequence and the corresponding intervention action encoding. The output variable dimension is the conditional expectation vector of the physiological state of the future observation step. The objective function of the corresponding iterative process is set as minimizing the weighted sum of squared residuals. The Newton-Raphson iterative algorithm uses the longitudinal data regression coefficients as initial values, calculates the Hessian matrix of the objective function, and updates the model parameters in the reverse direction until the difference in the norm of the parameter vector between two adjacent iterations is less than the preset convergence precision. This establishes a deterministic multidimensional mapping path from the logical resampling sequence to the counterfactual expected physiological evolution trajectory.

[0040] The system ultimately resamples the physiological-intervention sequence matrix based on steps S105 and S106 to generate a pseudo-equilibrium sequence to reflect the lack of intervention caused by low systolic blood pressure. The final output is a risk parameter that covers the entire disease course. When the systolic blood pressure index is at the edge of the working condition, the prediction mean square error is kept within 5%. Because the system maintains the temporal integrity of the longitudinal medical data record, the output physiological state evolution trajectory is consistent with the actual clinical outcome trend, thus solving the numerical convergence problem of healthcare information systems under guideline constraints.

[0041] Example 4: In the initial deployment of the cross-regional chronic disease collaborative management system, the system accesses the digital clinical guideline database storing medication guidelines for diabetic nephropathy according to step S103. It uses entity recognition operators to retrieve triplets containing contraindicated drug names and physiological indicator boundary values ​​from medical texts, thereby establishing an initial clinical contraindication threshold matrix for the target intervention. To improve the accuracy of the clinical contraindication threshold matrix within specific medical institutions, the computing device retrieves anonymized historical electronic medical records from the past 24 months of that institution, statistically analyzes the frequency of medication discontinuation by physicians at different glomerular filtration rate levels, and uses a maximum likelihood estimation algorithm to determine the central trend shift of the actual discontinuation point distribution, thereby correcting the initial physiological indicators. Extreme value intervals are marked to generate quantitative judgment benchmarks. The interval correction operation here does not change the inherent biological pathological absolute contraindication extreme value boundary. Instead, it constructs a pre-emptive safety warning buffer by introducing a central trend shift. The physician behavior shift obtained by the maximum likelihood estimation is used as a conservative algebraic bias term and scalarly summed with the original guideline rigid biological threshold, thereby defining a more stringent preventive extreme value interval. When the patient's physiological characteristics have not yet crossed the biological absolute red line but have fallen into this correction buffer, the system will preemptively determine and trigger the above-mentioned calculation link blocking action, thereby seamlessly and reasonably converting the macro-level group medical behavior deviation into a pre-emptive risk control threshold for the quantitative evaluation of individual intervention effects.

[0042] When the system is applied to patients with irregular follow-up frequencies and the time-varying confounding adjustment weights are determined according to step S104, the computing device calculates the coefficient of variation of the physiological state characteristic components using the preceding 12 time observation steps as the search interval. To determine the historical stable sampling window, where, A sequence segment with a value less than 0.05 and a duration of at least 3 time observation steps is defined as a historical stationary sampling window. The computing device uses the arithmetic mean of the intervention action codes within the historical stationary sampling window as the expected intervention frequency. It calculates the discrete residual between the actual intervention action code and the expected intervention frequency at the current time observation step to cover the weighted branches truncated due to the non-positive definite probability of conditional occurrence. Based on the Bayesian smoothing prior principle, when individual time series data exhibits extremely sparse or continuously fluctuating characteristics, a prior distribution constraint is introduced to estimate the divergence interval of the posterior estimate. This is applied when there is no coefficient of variation within the prior search interval. For oscillating conditions of stationary sequence segments less than 0.05, a baseline adaptive update procedure is triggered. A subset of historical patients containing the target contraindication indicator is extracted from the regional medical data center. The arithmetic mean of intervention frequencies within the subset at the same disease stage is calculated and set as the population prior parameter. The preceding search interval is expanded to 24 time observation steps. An exponential decay weight is determined based on the time interval between each time observation step and the current time. The exponential decay weight is used to perform a weighted summation operation on the actual intervention action codes within the expanded preceding search interval and the population prior parameter. The output baseline fusion expectation replaces the original arithmetic mean in calculating the intervention frequency deviation residual. In this weighted summation operation logic, the specific calculation rule for the exponential decay weight is: based on the current extrapolation time... Using zero time as the reference point, a decay index is constructed based on the time interval steps between each historical observation step and the current time. The average disease duration follow-up span of the chronic disease sample set within the regional medical data center is uniformly used as the scaling constant for the calculation base. Simultaneously, based on the Bayesian smoothing prior principle, when the temporal variation characteristics of the actual intervention actions of individual patients cross the upper boundary of the confidence interval of the distribution of the entire validation dataset, the computing device forces the allocation ratio of decay weights to converge and constrain to the group prior parameters. This fusion mapping rule is used to mathematically curb the divergence of posterior estimates caused by extreme free samples, thereby ensuring that the system produces risk parameters that reflect the evolution logic of the disease under the coexistence of medical guideline constraints and sparse sampling interference.

[0043] Example 5: In the scenario of processing cross-hospital diabetes data integration in a regional medical data center, the computing device retrieves preset reference ranges for various physiological indicators from the memory and uses linear mapping to convert the glycated hemoglobin index. and systolic blood pressure index The physiological characteristics are normalized to the range of 0 to 1. Simultaneously, the computing device determines the dynamic evolution threshold according to step S103. The calibration method includes retrieving a validation dataset containing 500 historical follow-up patients, calculating the monthly decline slope distribution of glomerular filtration rate in the validation dataset under the natural course of the disease, and extracting the 95th percentile value of the distribution as the dynamic evolution threshold, thereby calculating the conditional occurrence probability. Provide scale-uniform feature inputs and sampling periods The adaptive switching establishes a quantitative judgment benchmark.

[0044] When the system is applied to sparse medical records containing missing records and determines the time-varying heterogeneous adjustment weights according to step S104, the computing device scans the physiological state feature component vectors using a sliding window, and calculates the mean for each physiological indicator component within the window. and standard deviation To determine the coefficient of variation The calculation formula is as follows: ,in, The coefficient of variation is 1. The standard deviation of the physiological state characteristic components within the sliding window; The mean of the physiological state characteristic components within the sliding window; comparison of computing devices. Compared to the preset threshold of 0.05, when three consecutive time observation steps correspond to... When all values ​​are below the preset threshold, the current sequence segment is determined as a historical stationary sampling window. The arithmetic mean of the current intervention action codes within the historical stationary sampling window is calculated as the expected intervention frequency. The expected intervention frequency and the actual intervention action codes are used to calculate the discrete residual and update the weighted branches. This generates a pseudo-equilibrium sequence to correct the weight allocation bias induced by irregular sampling frequency, so that the risk parameters maintain numerical convergence even when the data missing rate reaches 20%.

[0045] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for evaluating and predicting the effectiveness of chronic disease interventions based on causal inference, characterized in that, Includes the following steps: Step S101: Obtain a physiological-intervention sequence matrix containing physiological state feature components and current intervention action codes. The physiological-intervention sequence matrix is ​​composed of time-series medical data records of the target patient. Step S102: Calculate the conditional probability of the current intervention action occurring within each time observation step in the physiological-intervention sequence matrix; Step S103: Determine the clinical contraindication threshold matrix corresponding to the physiological state characteristic components, wherein the clinical contraindication threshold matrix defines the extreme value range of physiological indicators that are mutually exclusive with the current intervention action; Step S104: Based on the mapping results between physiological state feature components and the extreme value range of physiological indicators, determine the time-varying hybrid adjustment weight corresponding to the current time observation step, wherein: when the physiological state feature components do not fall into the extreme value range of physiological indicators, the reciprocal of the conditional occurrence probability is determined as the time-varying hybrid adjustment weight; when the physiological state feature components fall into the extreme value range of physiological indicators, the calculation link from the conditional occurrence probability to the time-varying hybrid adjustment weight is blocked, the intervention frequency deviation residual within the preset sample smoothing window in the physiological-intervention sequence matrix is ​​extracted, and the intervention frequency deviation residual is used to cover the time-varying hybrid adjustment weight. Step S105: Based on the determined time-varying hybrid adjustment weights of each time observation step, the physiological-intervention sequence matrix is ​​logically resampled to generate a pseudo-equilibrium sequence. Step S106: Perform counterfactual state mapping on the pseudo-equilibrium sequence and output the evolution trajectory of the physiological state of chronic diseases.

2. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference according to claim 1, characterized in that, Step S103 includes the following sub-steps: Step S1031, extracting rigid contraindications for specific chronic diseases from the digital clinical guideline library; Step S1032, converting the rigid contraindications into logical trigger thresholds within the electronic computing device; Step S1033, constructing a clinical contraindication threshold matrix in a multi-dimensional probability space based on the logical trigger thresholds, the clinical contraindication threshold matrix being used to lock the blocking of the computing link in step S104.

3. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference according to claim 1, characterized in that, In step S101, the time-series medical data recording includes electronic medical record data, physiological parameter monitoring streams, and medication adherence records; physiological state characteristic components include blood glucose fluctuation rate, blood pressure coefficient of variation, and biochemical parameters reflecting organ function; current intervention action coding includes drug dosage adjustment indicators, dietary intervention intensity, and exercise prescription compliance rate.

4. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference as described in claim 1, characterized in that, Step S102 includes: using a structural marginal model to analyze the coupling relationship between physiological state feature components and subsequent current intervention actions; determining the induced response generated by the historical evolution of physiological state feature components to the current intervention action, and using the induced response as a bias term of the conditional occurrence probability.

5. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference according to claim 1, characterized in that, In step S104, when using the intervention frequency deviation residual to cover the time-varying hybrid adjustment weight, the following steps are also included: step S1041, real-time transmission of the probability of the occurrence of the masking condition to the time-varying hybrid adjustment weight; step S1042, locking the upper limit of the value of the time-varying hybrid adjustment weight to limit the divergence of the calculation matrix.

6. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference according to claim 1, characterized in that, In step S105, the process of generating the pseudo-equilibrium sequence includes: applying time-varying confounding adjustment weights to the physiological-intervention sequence matrix, and constructing a counterfactual inference space that is logically decoupled from the time-varying confounding factors by performing distribution recalibration on the observed samples.

7. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference according to claim 1, characterized in that, In step S106, outputting the evolution trajectory of the physiological state of chronic disease includes: fitting the evolution trend of the physiological state on the mimicry equilibrium sequence according to the preset candidate intervention path; calculating the average treatment effect of the target intervention program relative to the baseline program, and outputting the intervention effect evaluation curve containing confidence intervals.

8. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference according to claim 1, characterized in that, The method also includes: performing survivor bias correction on sample points that meet the clinical contraindication threshold matrix; and retaining high-risk samples through the weight coverage mechanism in step S104 to eliminate prediction baseline drift caused by sample removal.

9. The method for evaluating and predicting the effect of chronic disease intervention based on causal inference according to claim 1, characterized in that, After step S106, the method further includes: comparing the evolution trajectory of the physiological state of chronic diseases with the real-time acquired physiological parameter monitoring stream; if the deviation between the two exceeds a preset deviation threshold, resetting the model parameters used to calculate the probability in step S102 to achieve online optimization of the evaluation logic.

Images