Auxiliary follow-up triggering method based on clinical diagnosis and treatment path deviation detection

Abstract

The application provides an auxiliary follow-up triggering method based on clinical diagnosis and treatment path deviation detection. The method collects historical data containing individual covariates, intervention decisions and follow-up results, extracts clinical path deviation characteristics such as node missing, time offset, intervention category deviation and follow-up interval anomaly, and constructs a decoupling representation space composed of causal confounding, intervention tool and result adjustment subspace; reduce the coupling of subspace by minimizing mutual information, combine the double-head network to predict the fact and counter-fact results, and use the maximum mean difference item to align the confounding distribution of the intervention group and the control group to obtain the individual causal effect; then determine the non-negative improvement deficiency according to the individual causal effect, construct a decision function by fusing the uncertainty measure obtained after fusing the follow-up cost and the disturbance of the three types of subspace, dynamically adjust the triggering threshold, and generate an auxiliary follow-up triggering signal when the function output is out of limit, thereby improving the rationality of follow-up resource allocation and triggering judgment as a whole.

Description

Technical Field

[0001] This application belongs to the field of detection, and in particular relates to an auxiliary follow-up triggering method based on the detection of deviations in clinical diagnosis and treatment pathways. Background Technology

[0002] With the continuous improvement of medical informatization and the popularization of the concept of whole-life-cycle health management, post-discharge follow-up work plays an increasingly important role in chronic disease management, postoperative rehabilitation monitoring, and disease recurrence prevention. Faced with a rapidly growing patient base, medical institutions' follow-up resources are relatively limited. How to identify individuals who most need follow-up and can benefit most from follow-up interventions is key to improving the efficiency of medical services. Traditional follow-up decision support typically relies on clinical experience rules or conventional machine learning prediction models. In real-world scenarios, assessing the true causal effect of clinical interventions, such as drug interventions, on individual recovery outcomes is crucial. Historical follow-up data is mostly observational, containing numerous and complex confounding factors—variables that simultaneously influence intervention allocation and prognostic outcomes. If these confounding factors are not removed from the data at the data level, their entanglement with pure intervention and pure outcome factors will result in severe distributional bias in the assessment of causal effects, leading to a decline in the scientific basis of follow-up decisions.

[0003] In real-world business environments, every follow-up intervention involves corresponding costs. An ideal follow-up trigger mechanism not only needs to assess the individual's benefit but also must consider the intervention cost to maximize resource utility. Due to the diversity of individual characteristics and the noise in real-world data, models inevitably face uncertainty when making causal inferences and counterfactual predictions. If the decision-making system blindly trusts a single predictive output without considering the potential uncertainties in the complex representation space, the triggering decision will become overly rigid. This may not only lead to ineffective follow-ups and wasted resources but also miss individuals whose predicted outcomes are on the edge of high uncertainty who actually urgently need intervention. Therefore, there is an urgent need for a comprehensive solution that can effectively decouple confounding factors from biased historical data, align distributional biases to obtain individual causal effects, and organically combine intervention costs with model predictive uncertainty to generate follow-up trigger signals, thereby optimizing the allocation of medical and other related resources while ensuring intervention benefits. Summary of the Invention

[0004] To improve the efficiency of follow-up resource allocation and the scientific nature of intervention decisions, this invention provides an auxiliary follow-up triggering method based on clinical treatment pathway deviation detection.

[0005] According to one aspect of the present invention, an auxiliary follow-up triggering method based on clinical treatment pathway deviation detection is provided, comprising the following steps: Historical follow-up data containing individual covariates, intervention decisions, and follow-up results are obtained. A decoupled representation space is constructed, and the individual covariates are mapped to decoupled representations. The decoupled representation space consists of a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace. By adding a mutual information minimization regularization term during the training process of the machine learning model, the degree of information coupling between the corresponding representations of the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace is reduced, so as to respectively characterize the confounding factors, intervention-related factors, and outcome-related factors in the data. Based on the decoupled representation, a dual-head network is constructed to predict factual and counterfactual results. A maximum mean difference term is added to the loss function during the model training phase. The distribution distance between the intervention group and the control group is minimized using the causal confounding subspace representation to align the confounding bias in the data. The difference between the predicted factual and counterfactual results is calculated to obtain the individual causal effect value. A decision function is constructed, with inputs including the individual causal effect value, cost parameters, and uncertainty measure. The uncertainty measure is obtained by evaluating the changes in prediction results caused by perturbations applied to the three subspace representations respectively. The trigger threshold is adjusted according to the uncertainty measure. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up.

[0006] Optionally, obtaining historical follow-up data containing individual covariates, intervention decisions, and follow-up results includes: Missing values ​​in the physical examination data records are imputed by mean. Age, systolic blood pressure, blood glucose values ​​and clinical treatment path deviation features are concatenated into a one-dimensional feature vector as the individual covariate. The binary identifier of whether drug intervention is performed in the electronic medical record is extracted as the intervention decision. The number of days of the patient's recovery period is extracted as the follow-up result.

[0007] Optionally, the construction of the decoupled representation space maps the individual covariates to decoupled representations. The decoupled representation space comprises a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace, including: Three independent feedforward neural networks are constructed, corresponding to the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace, respectively. The individual covariates are input into the three feedforward neural networks, which output the corresponding causal confounding feature vector, intervention tool feature vector, and outcome adjustment feature vector, respectively. The three feature vectors are concatenated and combined to generate a one-dimensional representation vector, which serves as the decoupled representation.

[0008] Optionally, the step of adding a maximum mean difference term to the loss function during model training, using a causal confounding subspace to minimize the distribution distance between the intervention group and the control group, includes: The radial basis function is used to map the feature vectors of the intervention group and the control group in the causal mixed subspace to the regenerative kernel Hilbert space. The empirical means of the intervention group and the control group in the space are calculated respectively, and the square of the L2 norm of the difference between the two empirical means is taken as the maximum mean difference term. The maximum mean difference term is integrated into the mean squared error loss function used for regression prediction for joint optimization.

[0009] Optionally, the difference between the calculated predicted factual outcome and the counterfactual outcome, yielding an individual causal effect value, includes: Extract the first prediction result under the intervention condition and the second prediction result under the non-intervention condition from the output of the dual-head network. Calculate the difference between the second prediction result and the first prediction result. Use the difference as the individual causal effect value that characterizes the degree of improvement in follow-up results after intervention.

[0010] Optionally, the construction of the decision function, with inputs including the individual causal effect value, cost parameters, and uncertainty measures, includes: The fixed labor cost and communication resource consumption cost for a single follow-up visit are obtained. The sum of the fixed labor cost and communication resource consumption cost is divided by a preset benchmark cost to convert it into a dimensionless cost coefficient greater than zero, which is used as the cost parameter. Based on a preset benchmark improvement threshold with the same dimension as the individual causal effect value and the individual causal effect value, the improvement inadequacy amount representing the risk of deviation from the treatment path is determined. The improvement inadequacy amount and the uncertainty measure are divided by a preset benchmark rehabilitation cycle greater than zero for dimensionless processing. The sum of the dimensionless improvement inadequacy amount and the dimensionless uncertainty measure is used as the numerator, and the cost parameter is used as the denominator. The follow-up rate of return is calculated by division operation, and the follow-up rate of return is used as the function output of the decision function.

[0011] Optionally, the uncertainty measure is obtained by evaluating the changes in prediction results caused by perturbations applied to the three subspace representations, including: A Gaussian noise vector with zero mean and fixed variance is generated. The Gaussian noise vector is repeatedly sampled according to a preset number of samplings. The Gaussian noise vector is injected into the causal confounding feature vector, the intervention tool feature vector, and the result adjustment feature vector, respectively, to generate a perturbed decoupled representation. The perturbed decoupled representation is input into a dual-head network to obtain the perturbed prediction result. The Euclidean distance between the perturbed prediction result and the original prediction result is calculated. The mean value obtained by summing the Euclidean distances obtained from each perturbing and dividing it by the preset number of samplings is used as the uncertainty measure.

[0012] Optionally, the trigger threshold is adjusted according to an uncertainty metric. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up, including: An initial trigger threshold is set, the dimensionless uncertainty measure is multiplied by a preset adjustment coefficient, and the product result is subtracted from the initial trigger threshold to obtain the trigger threshold. The trigger threshold is then subjected to a preset lower limit truncation. The follow-up rate of return is compared with the trigger threshold in real time. When the follow-up rate of return is greater than the trigger threshold, a digital control signal is generated to drive the communication module to initiate an automatic follow-up action.

[0013] According to another aspect of the present invention, an auxiliary follow-up triggering system based on clinical treatment pathway deviation detection is provided, comprising the following modules: A construction module is used to acquire historical follow-up data containing individual covariates, intervention decisions, and follow-up results, and construct a decoupled representation space to map the individual covariates to decoupled representations. The decoupled representation space consists of a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace. By adding a mutual information minimization regularization term during the training process of the machine learning model, the degree of information coupling between the corresponding representations of the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace is reduced, so as to respectively characterize the confounding factors, intervention-related factors, and outcome-related factors in the data. The calculation module is used to construct a dual-head network based on the decoupled representation to predict the factual and counterfactual results. In the loss function during the model training phase, a maximum mean difference term is added. The distribution distance between the intervention group and the control group is minimized using the causal confounding subspace representation to align the confounding bias in the data. The difference between the predicted factual and counterfactual results is calculated to obtain the individual causal effect value. The generation module is used to construct a decision function. The input includes the individual causal effect value, cost parameters, and uncertainty measure. The uncertainty measure is obtained by evaluating the changes in the prediction results caused by applying perturbations to the three subspace representations respectively. The trigger threshold is adjusted according to the uncertainty measure. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up.

[0014] Furthermore, the acquisition of historical follow-up data, including individual covariates, intervention decisions, and follow-up results, includes: Missing values ​​in the physical examination data records are imputed by mean. Age, systolic blood pressure, blood glucose values ​​and clinical treatment path deviation features are concatenated into a one-dimensional feature vector as the individual covariate. The binary identifier of whether drug intervention is performed in the electronic medical record is extracted as the intervention decision. The number of days of the patient's recovery period is extracted as the follow-up result.

[0015] Furthermore, the construction of the decoupled representation space maps the individual covariates to decoupled representations. The decoupled representation space comprises a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace, including: Three independent feedforward neural networks are constructed, corresponding to the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace, respectively. The individual covariates are input into the three feedforward neural networks, which output the corresponding causal confounding feature vector, intervention tool feature vector, and outcome adjustment feature vector, respectively. The three feature vectors are concatenated and combined to generate a one-dimensional representation vector, which serves as the decoupled representation.

[0016] Furthermore, the addition of a maximum mean difference term to the loss function during model training, utilizing a causal confounding subspace to minimize the distribution distance between the intervention group and the control group, includes: The radial basis function is used to map the feature vectors of the intervention group and the control group in the causal mixed subspace to the regenerative kernel Hilbert space. The empirical means of the intervention group and the control group in the space are calculated respectively, and the square of the L2 norm of the difference between the two empirical means is taken as the maximum mean difference term. The maximum mean difference term is integrated into the mean squared error loss function used for regression prediction for joint optimization.

[0017] Furthermore, the difference between the calculated predicted factual outcome and the counterfactual outcome yields an individual causal effect value, including: Extract the first prediction result under the intervention condition and the second prediction result under the non-intervention condition from the output of the dual-head network. Calculate the difference between the second prediction result and the first prediction result. Use the difference as the individual causal effect value that characterizes the degree of improvement in follow-up results after intervention.

[0018] Furthermore, the construction of the decision function, with inputs including the individual causal effect value, cost parameters, and uncertainty measures, includes: The fixed labor cost and communication resource consumption cost for a single follow-up visit are obtained. The sum of the fixed labor cost and communication resource consumption cost is divided by a preset benchmark cost to convert it into a dimensionless cost coefficient greater than zero, which is used as the cost parameter. Based on a preset benchmark improvement threshold with the same dimension as the individual causal effect value and the individual causal effect value, the improvement inadequacy amount representing the risk of deviation from the treatment path is determined. The improvement inadequacy amount and the uncertainty measure are divided by a preset benchmark rehabilitation cycle greater than zero for dimensionless processing. The sum of the dimensionless improvement inadequacy amount and the dimensionless uncertainty measure is used as the numerator, and the cost parameter is used as the denominator. The follow-up rate of return is calculated by division operation, and the follow-up rate of return is used as the function output of the decision function.

[0019] Furthermore, the uncertainty measure is obtained by evaluating the changes in prediction results caused by applying perturbations to the three subspace representations, including: A Gaussian noise vector with zero mean and fixed variance is generated. The Gaussian noise vector is repeatedly sampled according to a preset number of samplings. The Gaussian noise vector is injected into the causal confounding feature vector, the intervention tool feature vector, and the result adjustment feature vector, respectively, to generate a perturbed decoupled representation. The perturbed decoupled representation is input into a dual-head network to obtain the perturbed prediction result. The Euclidean distance between the perturbed prediction result and the original prediction result is calculated. The mean value obtained by summing the Euclidean distances obtained from each perturbing and dividing it by the preset number of samplings is used as the uncertainty measure.

[0020] Furthermore, the trigger threshold is adjusted based on an uncertainty metric. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up, including: An initial trigger threshold is set, the dimensionless uncertainty measure is multiplied by a preset adjustment coefficient, and the product result is subtracted from the initial trigger threshold to obtain the trigger threshold. The trigger threshold is then subjected to a preset lower limit truncation. The follow-up rate of return is compared with the trigger threshold in real time. When the follow-up rate of return is greater than the trigger threshold, a digital control signal is generated to drive the communication module to initiate an automatic follow-up action.

[0021] This invention constructs a decoupled representation space consisting of causal confounding factors, intervention tools, and outcome adjustments. By adding a mutual information minimization regularization term during model training, it can separate and extract correlated factors from historical follow-up data, reducing coupling interference between features and improving the independence and distinguishability of data representation. By comparing actual patient treatment records with standard clinical treatment pathway templates, it can extract clinical treatment pathway deviation features such as missing treatment nodes, time window offsets, intervention execution deviations, and abnormal follow-up intervals, providing a clear basis for auxiliary follow-up trigger judgments. By aligning the distribution differences between the intervention and control groups using the maximum mean difference term, it solves the problem of confounding bias in the data and improves the stability of individual causal effect estimation. The constructed decision-making mechanism utilizes the improvement insufficiency, follow-up costs, and uncertainty measures obtained based on feature perturbation assessment. It scientifically adjusts the trigger threshold according to this uncertainty level, improving the rationality and stability of follow-up trigger signal generation and optimizing the overall execution process and management efficiency of auxiliary follow-up. Attached Figure Description

[0022] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram illustrating the inference of individual causal effects; Figure 3 This is a diagram illustrating the implementation of follow-up decisions. Detailed Implementation

[0023] In this application, the terms "exemplarily" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplarily" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0024] In the first embodiment, the present invention proposes an auxiliary follow-up triggering method based on clinical treatment pathway deviation detection, such as... Figure 1 As shown, it includes the following steps: S1. Obtain historical follow-up data containing individual covariates, intervention decisions, and follow-up results. Construct a decoupled representation space and map the individual covariates to decoupled representations. The decoupled representation space consists of a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace. By adding a mutual information minimization regularization term during the training process of the machine learning model, the degree of information coupling between the corresponding representations of the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace is reduced, so as to respectively characterize the confounding factors, intervention-related factors, and outcome-related factors in the data.

[0025] In this embodiment, clinical treatment pathway deviation refers to the execution offset of the patient's actual treatment record relative to the standard clinical treatment pathway template for the target disease. The standard clinical treatment pathway template includes preset treatment nodes, time windows corresponding to each treatment node, recommended intervention categories, and planned follow-up nodes. The patient's actual electronic medical record, medical order record, examination record, and follow-up record are aligned with the standard clinical treatment pathway template in chronological order. Node missing identifiers, treatment time offsets, intervention category deviation identifiers, and follow-up interval deviation values ​​are extracted respectively. After normalizing the above values, clinical treatment pathway deviation features are generated. The clinical treatment pathway deviation features are input as part of the individual covariates into the subsequent decoupled representation space to characterize at least one of treatment node missingness, treatment time offset, intervention category deviation, or abnormal follow-up interval.

[0026] The system reads a tabular file containing historical follow-up data, standardizes individual covariates, converts intervention decisions into binary labels, and retains follow-up results as continuous regression labels. Three parallel and independent feedforward neural networks are constructed, each consisting of a fully connected layer, a ReLU activation function, and a Dropout layer. These networks map the standardized individual covariates to represent the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace, respectively. A mutual information minimization regularization term is added to the model training loss function. The MINE mutual information neural estimator algorithm is used to estimate the pairwise mutual information values ​​between the three subspace representation vectors. Specifically, a separately constructed statistical network is used to approximate the KL divergence between the product of the joint distribution and the marginal distribution through gradient descent optimization. The estimated mutual information values ​​are added as a regularization term to the total loss function. Backpropagation is performed using the torch.optim.Adam optimizer to update the parameters of the three feedforward neural networks and the statistical network until the mutual information regularization term meets the preset convergence condition, thus completing the construction of the decoupled representation space. The statistical network is preferably a multilayer perceptron used for MINE algorithm estimation. Its structure consists of a fully connected layer and a nonlinear activation function cascaded together. The input is a pairwise combination of subspace representation vectors, which are sampled from the product of the joint distribution and the marginal distribution, respectively. The output is a scalar value, i.e., a mutual information estimate, used to approximate the KL divergence between the two distributions.

[0027] In an optional embodiment, obtaining historical follow-up data including individual covariates, intervention decisions, and follow-up outcomes includes: Missing values ​​in the physical examination data records are imputed by mean. Age, systolic blood pressure, blood glucose values ​​and clinical treatment path deviation features are concatenated into a one-dimensional feature vector as the individual covariate. The binary identifier of whether drug intervention is performed in the electronic medical record is extracted as the intervention decision. The number of days of the patient's recovery period is extracted as the follow-up result.

[0028] During the historical follow-up data acquisition phase, the system connects to the medical institution's physical examination database, such as the HIS system or physical examination center database, extracting the individual's age, systolic blood pressure in mmHg, and fasting blood glucose level in mmol / L as core variables. Simultaneously, it connects to the electronic medical record system, medical order system, and clinical pathway management system to read the patient's actual treatment nodes and their occurrence times, and matches them with the standard clinical treatment pathway template for the target disease. If the actual record lacks a required follow-up node, a node missing marker is assigned a value of 1; otherwise, it is assigned a value of 0. If the actual medication time is later than the template's recommended time, the time difference between the actual medication time and the recommended time is calculated and normalized to a time offset. If the actual intervention category is inconsistent with the template's recommended category, an intervention category deviation marker is generated. If the actual follow-up interval exceeds the follow-up window set in the template, a follow-up interval deviation value is calculated. The node missing marker, time offset, intervention category deviation marker, and follow-up interval deviation value are combined to form the clinical treatment pathway deviation feature. If some records contain missing values, the mean of the corresponding variable in the same patient group is calculated for imputation. For example, the mean age is 45 years, the mean systolic blood pressure is 120 mmHg, and the mean blood glucose is 5.5 mmol / L. After imputation, these three values ​​are concatenated with the clinical treatment pathway deviation characteristics to generate a one-dimensional feature vector, which is used as an individual covariate. In the medical order module of the electronic medical record system, it is identified whether a specific drug intervention has been issued, and the intervention status is mapped to a binary identifier. If drug intervention has been carried out, a value of 1 is assigned, and if no intervention has been carried out, a value of 0 is assigned, which is used as the intervention decision. From the discharge or follow-up records, the number of days required for the patient to recover or reach the target is extracted from the preset starting point of the discharge time, first treatment time, enrollment observation time, or intervention start time as a continuous numerical variable, and the number of days is used as the follow-up result.

[0029] In an optional embodiment, the construction of the decoupled representation space, which maps the individual covariates to decoupled representations, comprises a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace, including: Three independent feedforward neural networks are constructed, corresponding to the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace, respectively. The individual covariates are input into the three feedforward neural networks, which output the corresponding causal confounding feature vector, intervention tool feature vector, and outcome adjustment feature vector, respectively. The three feature vectors are concatenated and combined to generate a one-dimensional representation vector, which serves as the decoupled representation.

[0030] In constructing the model network architecture, three parallel and independent multi-layer feedforward neural networks are built. Each of the three feedforward neural networks sequentially includes an input layer, a first hidden layer, a second hidden layer, and an output layer. For example, the first hidden layer has 64 nodes, the second hidden layer has 32 nodes, and the output layer has 16 nodes. The ReLU activation function is used. The input to each of the three feedforward neural networks is an individual covariate vector, and the outputs are a causal confounding feature vector, an intervention tool feature vector, and an outcome adjustment feature vector, respectively. When the individual covariate vector is input to the input layer of these three feedforward neural networks, it includes both basic physiological indicators such as age, systolic blood pressure, and blood glucose, as well as path deviation features obtained from clinical treatment path deviation detection. Each network maps the input to a low-dimensional latent space through its first and second hidden layers, outputting causal confounding feature vectors, intervention tool feature vectors, and outcome adjustment feature vectors, each with a dimension of 1×16, respectively. To form a complete representation system, these three feature vectors are concatenated along the feature dimension to generate a one-dimensional joint representation vector with a dimension of 1×48. This representation vector is the decoupled representation. In this process, the mutual information between the three feature vectors is calculated using a mutual information neural estimator, and a mutual information minimization regularization term is added to the overall loss function. The regularization coefficient is set to 0.2 for example. The network weights are updated during backpropagation, so that the extracted subspace features are statistically less correlated and tend to be independent of each other.

[0031] S2, based on the decoupled representation, construct a dual-head network to predict the factual and counterfactual results. Add a maximum mean difference term to the loss function during the model training phase. Use the causal confounding subspace representation to minimize the distribution distance between the intervention group and the control group, align the confounding bias in the data, calculate the difference between the predicted factual and counterfactual results, and obtain the individual causal effect value.

[0032] A dual-head network architecture with two independent multilayer perceptron branches was constructed. The first multilayer perceptron branch served as the prediction head for the intervention group, and the second multilayer perceptron branch served as the prediction head for the control group. The causal confounding subspace representation vector, the intervention tool subspace representation vector, and the outcome adjustment subspace representation vector were concatenated. Based on the intervention decision labels of the samples, the concatenated representation of the intervention group samples was input into the first multilayer perceptron branch to predict the factual outcome, and the concatenated representation of the control group samples was input into the second multilayer perceptron branch to predict the factual outcome. The intervention group samples were input into the control group prediction head to predict the counterfactual outcome, and the control group samples were input into the intervention group prediction head to predict the counterfactual outcome. Since the follow-up result is the number of days of the recovery period, which is a continuous regression variable, the mean squared error loss function was used to calculate the prediction error of the factual outcome.

[0033] During model training, the causal confounding subspace representation vector sets of the intervention group and the control group are extracted. The maximum mean difference (MMD) value between the two representation vector sets is calculated using the radial basis function (RBF) kernel function. This MMD value is multiplied by a preset balance coefficient and added to the mean squared error loss function to form the overall causal loss. The balance coefficient is used to scale the MMD and MSD values, or to normalize the follow-up result labels before training to ensure they are on the same optimization scale. The parameters of the three feedforward neural networks and the dual-head network are jointly optimized through error backpropagation to minimize the factual outcome prediction error and reduce the distribution distance between the intervention group and the control group in the causal confounding subspace. After model convergence, for any new input individual, the rehabilitation cycle prediction results for the new individual under intervention and non-intervention states are output through the intervention group prediction head and the control group prediction head, respectively. The rehabilitation cycle prediction result under intervention state is subtracted from the rehabilitation cycle prediction result under non-intervention state using a subtraction operation to obtain the individual causal effect value for that individual. Figure 2 As shown.

[0034] In an optional embodiment, the step of adding a maximum mean difference term to the loss function during model training, using a causal confounding subspace to minimize the distribution distance between the intervention group and the control group, includes: The radial basis function is used to map the feature vectors of the intervention group and the control group in the causal mixed subspace to the regenerative kernel Hilbert space. The empirical means of the intervention group and the control group in the space are calculated respectively, and the square of the L2 norm of the difference between the two empirical means is taken as the maximum mean difference term. The maximum mean difference term is integrated into the mean squared error loss function used for regression prediction for joint optimization.

[0035] The maximum mean difference metric mechanism was employed in the optimization of the causal confounding subspace representation. Causal confounding feature vectors corresponding to patients in the intervention group (intervention label 1) and the control group (intervention label 0) were extracted. A radial basis function (RBF) kernel was selected, with its bandwidth parameter calculated using a median heuristic algorithm, for example, set to a range of 1.0 to 5.0. These two sets of low-dimensional features were implicitly mapped to an infinite-dimensional regenerative kernel Hilbert space. Empirical means were calculated for the mapping results of the intervention group and the control group in the regenerative kernel Hilbert space, and the squared L2 norm of the difference between the two empirical mean vectors was calculated using the following formula: ,in and These represent the empirical means of the two groups, and this value is the maximum mean difference term representing the difference in distribution between the two groups. This maximum mean difference term is weighted and summed with the mean squared error loss of the network prediction results to construct the overall joint optimization objective function, where the balance coefficient is set to, for example, 5.0. This joint loss is backpropagated using the gradient descent algorithm, so that the probability distribution of the causal confounding features learned by the network tends to be similar between the intervention group and the control group, thus aligning against confounding biases.

[0036] In an optional embodiment, the calculation of the difference between the predicted factual outcome and the counterfactual outcome to obtain an individual causal effect value includes: Extract the first prediction result under the intervention condition and the second prediction result under the non-intervention condition from the output of the dual-head network. Calculate the difference between the second prediction result and the first prediction result. Use the difference as the individual causal effect value that characterizes the degree of improvement in follow-up results after intervention.

[0037] The dual-head network structure includes a common input layer and two independent fully connected regression branches at the top layer, simulating the potential responses under drug intervention and without drug intervention, respectively. Each fully connected regression branch contains a fully connected hidden layer and a regression output node. The input to the dual-head network is a decoupled representation vector, and the outputs are a first prediction result and a second prediction result. During the inference phase, the generated 1×48 decoupled representation vector is fed into the common input layer of the dual-head network, entering the two branches. The first branch sets the intervention flag to 1. After processing by the fully connected hidden layer of this branch, the regression output node of this branch outputs the predicted number of days of the patient's recovery period under drug intervention conditions, for example, a prediction of 15 days, as the first prediction result output by the dual-head network. The second branch sets the intervention flag to 0. After processing by the fully connected hidden layer of this branch, the regression output node of this branch outputs the predicted number of days of the same patient's recovery period under no drug intervention conditions, for example, a prediction of 22 days, as the second prediction result output by the dual-head network. Perform element-wise subtraction, subtracting the first prediction result from the second prediction result to obtain the difference. The calculation formula is as follows: Where Δ is the individual causal effect value, y1 is the first predicted result, and y0 is the second predicted result. For example, subtracting the first predicted result of 15 days from the second predicted result of 22 days yields 7 days. This value represents the extent to which drug intervention shortens the recovery period, and is used as the individual causal effect value for that specific individual.

[0038] S3, construct a decision function, the input of which includes the individual causal effect value, cost parameter and uncertainty measure. The uncertainty measure is obtained by evaluating the changes in the prediction results caused by applying perturbations to the three subspace representations respectively. Adjust the trigger threshold according to the uncertainty measure. When the function output exceeds the trigger threshold, generate a follow-up trigger signal for assisting follow-up.

[0039] A Gaussian noise vector with zero mean and a preset small variance is generated. In each sampling, a Gaussian noise vector with the same dimension as the causal confounding subspace representation vector, the intervention tool subspace representation vector, and the result adjustment subspace representation vector is generated. These are then superimposed on the corresponding subspace representation vectors and concatenated to form a perturbed decoupled representation. The perturbed decoupled representation is re-inputted into the dual-head network to obtain the perturbed recovery cycle prediction result. The Euclidean distance between the multiple perturbed recovery cycle prediction results and the original recovery cycle prediction result is calculated, and the mean of each Euclidean distance is used as the uncertainty measure. A decision function based on risk-reward scoring is constructed. The input variables of this function are the individual causal effect value, the uncertainty measure, and the cost parameter. The improvement shortfall is calculated based on the preset baseline improvement threshold and the individual causal effect value. The improvement shortfall and the uncertainty measure are both dimensionless. The sum of the dimensionless improvement shortfall and the dimensionless uncertainty measure is divided by the cost parameter to obtain the follow-up return rate. A trigger threshold calculation formula is constructed, with the initial threshold set as a fixed constant. The dimensionless uncertainty metric is multiplied by a preset adjustment coefficient as the threshold adjustment amount. This adjustment amount is subtracted from the initial threshold to obtain the adjusted trigger threshold. A preset lower limit is applied to the adjusted trigger threshold to make individuals with higher prediction uncertainty more likely to be included in the auxiliary follow-up focus. An if conditional statement is written to compare the output value of the decision function with the adjusted trigger threshold. When the output value of the decision function is greater than the trigger threshold, an HTTP request is sent via the system interface or an email containing the patient's identifier and follow-up suggestions is sent using Python's smtplib library. This serves as the generated auxiliary follow-up trigger signal, which is then written into the hospital information management system database. Figure 3 As shown.

[0040] In an optional embodiment, the construction of the decision function, with inputs including the individual causal effect value, cost parameters, and uncertainty measures, includes: The fixed labor cost and communication resource consumption cost for a single follow-up visit are obtained. The sum of the fixed labor cost and communication resource consumption cost is divided by a preset benchmark cost to convert it into a dimensionless cost coefficient greater than zero, which is used as the cost parameter. Based on a preset benchmark improvement threshold with the same dimension as the individual causal effect value and the individual causal effect value, the improvement inadequacy amount representing the risk of deviation from the treatment path is determined. The improvement inadequacy amount and the uncertainty measure are divided by a preset benchmark rehabilitation cycle greater than zero for dimensionless processing. The sum of the dimensionless improvement inadequacy amount and the dimensionless uncertainty measure is used as the numerator, and the cost parameter is used as the denominator. The follow-up rate of return is calculated by division operation, and the follow-up rate of return is used as the function output of the decision function.

[0041] The decision function is constructed by combining the benefits of causal inference with the cost of follow-up resource consumption to obtain financial indicators. For example, the fixed labor cost of a single follow-up, including the hourly rate of the follow-up personnel, is 15 yuan, and the communication resource consumption cost, including SMS and voice line fees, is 5 yuan. Adding these two together yields a total absolute cost of 20 yuan. This absolute cost is divided by a preset baseline cost, for example, set to 50 yuan, and normalized to obtain a dimensionless cost coefficient of 0.4, which serves as the cost parameter for that individual. During the evaluation phase, the previously calculated individual causal effect value is extracted, assumed to be 7, and a preset baseline improvement threshold is set to 10. The improvement deficiency of 3 is obtained by subtracting the individual causal effect value from the preset baseline improvement threshold. If the result is less than zero, the improvement deficiency is set to zero. A preset baseline rehabilitation period of 30 days is set. The improvement deficit of 3 divided by 30 yields a dimensionless improvement deficit of 0.1. The uncertainty measure of 1.2 divided by 30 yields a dimensionless uncertainty measure of 0.04. The dimensionless improvement deficit of 0.1 and the dimensionless uncertainty measure of 0.04 are added to obtain a benefit numerator of 0.14. This benefit numerator is then divided by the cost parameter of 0.4, resulting in a division operation of 0.35. This result is the dimensionless follow-up rate of return indicator, which is used as the output of the decision function.

[0042] In an optional embodiment, the uncertainty metric is obtained by evaluating the changes in prediction results caused by perturbations applied to the three subspace representations, including: A Gaussian noise vector with zero mean and fixed variance is generated. The Gaussian noise vector is repeatedly sampled according to a preset number of samplings. The Gaussian noise vector is injected into the causal confounding feature vector, the intervention tool feature vector, and the result adjustment feature vector, respectively, to generate a perturbed decoupled representation. The perturbed decoupled representation is input into a dual-head network to obtain the perturbed prediction result. The Euclidean distance between the perturbed prediction result and the original prediction result is calculated. The mean value obtained by summing the Euclidean distances obtained from each perturbing and dividing it by the preset number of samplings is used as the uncertainty measure.

[0043] A Gaussian noise vector with a mean of 0 and a variance of, for example, 0.05, and a dimension of 1×16 is generated. This noise vector is sampled multiple times during the inference process, and this vector is injected as an additional term into the original causal confounding feature vector, intervention tool feature vector, and result adjustment feature vector, respectively, to generate perturbed feature vectors. These three perturbed feature vectors are then reassembled to generate multiple sets of perturbed decoupled representation vectors, with the sampling number set to, for example, 50 times. These 50 sets of perturbed decoupled representation vectors are then batch-input into a dual-head network for forward propagation, resulting in 50 corresponding perturbed prediction results. The Euclidean distance between the two-dimensional prediction vector composed of the intervention condition prediction result and the non-intervention condition prediction result obtained after each perturbing step and the original two-dimensional prediction vector without added noise is calculated using the following formula: Where D is the Euclidean distance. For the disturbance prediction results, The original prediction result is obtained by summing the 50 distances and dividing by the number of samples, 50, to obtain the cumulative mean of these 50 distances. This cumulative mean represents the sensitivity of the model to input fluctuations and is used as a measure of uncertainty for decision-making.

[0044] In an optional embodiment, a trigger threshold is adjusted based on an uncertainty metric. When the function output exceeds the trigger threshold, a follow-up trigger signal for assisting follow-up is generated, including: An initial trigger threshold is set, the dimensionless uncertainty measure is multiplied by a preset adjustment coefficient, and the product result is subtracted from the initial trigger threshold to obtain the trigger threshold. The trigger threshold is then subjected to a preset lower limit truncation. The follow-up rate of return is compared with the trigger threshold in real time. When the follow-up rate of return is greater than the trigger threshold, a digital control signal is generated to drive the communication module to initiate an automatic follow-up action.

[0045] A scalar value is preset as the initial trigger threshold, for example, 0.30. The dimensionless uncertainty metric is extracted, assuming it to be 0.04, and multiplied by a predetermined adjustment coefficient, for example, 0.5, to calculate a threshold adjustment margin of 0.02. This adjustment margin of 0.02 is subtracted from the initial trigger threshold of 0.30 to generate a trigger threshold of 0.28. This trigger threshold is compared with a preset lower limit. If the trigger threshold is lower than the preset lower limit, it is adjusted to the preset lower limit. The follow-up yield of the current individual, for example, 0.35, is compared in real time with the generated trigger threshold of 0.28. Since 0.35 is greater than 0.28, a digital control signal containing the patient ID and communication command type is generated, driving the communication module to initiate an automatic follow-up action.

[0046] In a second embodiment, the present invention also provides an auxiliary follow-up triggering system based on clinical treatment pathway deviation detection, comprising the following modules: A construction module is used to acquire historical follow-up data containing individual covariates, intervention decisions, and follow-up results, and construct a decoupled representation space to map the individual covariates to decoupled representations. The decoupled representation space consists of a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace. By adding a mutual information minimization regularization term during the training process of the machine learning model, the degree of information coupling between the corresponding representations of the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace is reduced, so as to respectively characterize the confounding factors, intervention-related factors, and outcome-related factors in the data. The calculation module is used to construct a dual-head network based on the decoupled representation to predict the factual and counterfactual results. In the loss function during the model training phase, a maximum mean difference term is added. The distribution distance between the intervention group and the control group is minimized using the causal confounding subspace representation to align the confounding bias in the data. The difference between the predicted factual and counterfactual results is calculated to obtain the individual causal effect value. The generation module is used to construct a decision function. The input includes the individual causal effect value, cost parameters, and uncertainty measure. The uncertainty measure is obtained by evaluating the changes in the prediction results caused by applying perturbations to the three subspace representations respectively. The trigger threshold is adjusted according to the uncertainty measure. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An auxiliary follow-up triggering method based on clinical treatment pathway deviation detection, characterized in that, Includes the following steps: Historical follow-up data containing individual covariates, intervention decisions, and follow-up results are obtained. A decoupled representation space is constructed, and the individual covariates are mapped to decoupled representations. The decoupled representation space consists of a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace. By adding a mutual information minimization regularization term during the training process of the machine learning model, the degree of information coupling between the corresponding representations of the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace is reduced, so as to respectively characterize the confounding factors, intervention-related factors, and outcome-related factors in the data. Based on the decoupled representation, a dual-head network is constructed to predict factual and counterfactual results. A maximum mean difference term is added to the loss function during the model training phase. The distribution distance between the intervention group and the control group is minimized using the causal confounding subspace representation to align the confounding bias in the data. The difference between the predicted factual and counterfactual results is calculated to obtain the individual causal effect value. A decision function is constructed, with inputs including the individual causal effect value, cost parameters, and uncertainty measure. The uncertainty measure is obtained by evaluating the changes in prediction results caused by perturbations applied to the three subspace representations respectively. The trigger threshold is adjusted according to the uncertainty measure. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up.

2. The method according to claim 1, characterized in that, The acquisition of historical follow-up data, including individual covariates, intervention decisions, and follow-up results, includes: Missing values ​​in the physical examination data records are imputed by mean. Age, systolic blood pressure, blood glucose values ​​and clinical treatment path deviation features are concatenated into a one-dimensional feature vector as the individual covariate. The binary identifier of whether drug intervention is performed in the electronic medical record is extracted as the intervention decision. The number of days of the patient's recovery period is extracted as the follow-up result.

3. The method according to claim 1, characterized in that, The construction of the decoupled representation space maps the individual covariates to decoupled representations. The decoupled representation space consists of a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace, including: Three independent feedforward neural networks are constructed, corresponding to the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace, respectively. The individual covariates are input into the three feedforward neural networks, which output the corresponding causal confounding feature vector, intervention tool feature vector, and outcome adjustment feature vector, respectively. The three feature vectors are concatenated and combined to generate a one-dimensional representation vector, which serves as the decoupled representation.

4. The method according to claim 1, characterized in that, The addition of a maximum mean difference term to the loss function during model training, utilizing a causal confounding subspace to minimize the distribution distance between the intervention and control groups, includes: The radial basis function is used to map the feature vectors of the intervention group and the control group in the causal mixed subspace to the regenerative kernel Hilbert space. The empirical means of the intervention group and the control group in the space are calculated respectively, and the square of the L2 norm of the difference between the two empirical means is taken as the maximum mean difference term. The maximum mean difference term is integrated into the mean squared error loss function used for regression prediction for joint optimization.

5. The method according to claim 1, characterized in that, The difference between the calculated predicted factual outcome and the counterfactual outcome yields an individual causal effect value, including: Extract the first prediction result under the intervention condition and the second prediction result under the non-intervention condition from the output of the dual-head network. Calculate the difference between the second prediction result and the first prediction result. Use the difference as the individual causal effect value that characterizes the degree of improvement in follow-up results after intervention.

6. The method according to claim 1, characterized in that, The constructed decision function, with inputs including the individual causal effect value, cost parameters, and uncertainty measures, includes: The fixed labor cost and communication resource consumption cost for a single follow-up visit are obtained. The sum of the fixed labor cost and communication resource consumption cost is divided by a preset benchmark cost to convert it into a dimensionless cost coefficient greater than zero, which is used as the cost parameter. Based on a preset benchmark improvement threshold with the same dimension as the individual causal effect value and the individual causal effect value, the improvement inadequacy amount representing the risk of deviation from the treatment path is determined. The improvement inadequacy amount and the uncertainty measure are divided by a preset benchmark rehabilitation cycle greater than zero for dimensionless processing. The sum of the dimensionless improvement inadequacy amount and the dimensionless uncertainty measure is used as the numerator, and the cost parameter is used as the denominator. The follow-up rate of return is calculated by division operation, and the follow-up rate of return is used as the function output of the decision function.

7. The method according to claim 3, characterized in that, The uncertainty metric is obtained by evaluating the changes in prediction results caused by perturbations applied to the representations of the three subspaces, including: A Gaussian noise vector with zero mean and fixed variance is generated. The Gaussian noise vector is repeatedly sampled according to a preset number of samplings. The Gaussian noise vector is injected into the causal confounding feature vector, the intervention tool feature vector, and the result adjustment feature vector, respectively, to generate a perturbed decoupled representation. The perturbed decoupled representation is input into a dual-head network to obtain the perturbed prediction result. The Euclidean distance between the perturbed prediction result and the original prediction result is calculated. The mean value obtained by summing the Euclidean distances obtained from each perturbing and dividing it by the preset number of samplings is used as the uncertainty measure.

8. The method according to claim 6, characterized in that, The trigger threshold is adjusted based on an uncertainty metric. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up, including: An initial trigger threshold is set, the dimensionless uncertainty measure is multiplied by a preset adjustment coefficient, and the product result is subtracted from the initial trigger threshold to obtain the trigger threshold. The trigger threshold is then subjected to a preset lower limit truncation. The follow-up rate of return is compared with the trigger threshold in real time. When the follow-up rate of return is greater than the trigger threshold, a digital control signal is generated to drive the communication module to initiate an automatic follow-up action.

9. An auxiliary follow-up triggering system based on clinical diagnosis and treatment pathway deviation detection, characterized in that, Includes the following modules: A construction module is used to acquire historical follow-up data containing individual covariates, intervention decisions, and follow-up results, and construct a decoupled representation space to map the individual covariates to decoupled representations. The decoupled representation space consists of a causal confounding subspace, an intervention tool subspace, and an outcome adjustment subspace. By adding a mutual information minimization regularization term during the training process of the machine learning model, the degree of information coupling between the corresponding representations of the causal confounding subspace, the intervention tool subspace, and the outcome adjustment subspace is reduced, so as to respectively characterize the confounding factors, intervention-related factors, and outcome-related factors in the data. The calculation module is used to construct a dual-head network based on the decoupled representation to predict the factual and counterfactual results. In the loss function during the model training phase, a maximum mean difference term is added. The distribution distance between the intervention group and the control group is minimized using the causal confounding subspace representation to align the confounding bias in the data. The difference between the predicted factual and counterfactual results is calculated to obtain the individual causal effect value. The generation module is used to construct a decision function. The input includes the individual causal effect value, cost parameters, and uncertainty measure. The uncertainty measure is obtained by evaluating the changes in the prediction results caused by applying perturbations to the three subspace representations respectively. The trigger threshold is adjusted according to the uncertainty measure. When the function output exceeds the trigger threshold, a follow-up trigger signal is generated to assist in follow-up.

10. The system according to claim 9, characterized in that, The acquisition of historical follow-up data, including individual covariates, intervention decisions, and follow-up results, includes: Missing values ​​in the physical examination data records are imputed by mean. Age, systolic blood pressure, blood glucose values ​​and clinical treatment path deviation features are concatenated into a one-dimensional feature vector as the individual covariate. The binary identifier of whether drug intervention is performed in the electronic medical record is extracted as the intervention decision. The number of days of the patient's recovery period is extracted as the follow-up result.

Images