A system for assessing the value of a grass transaminase

Abstract

The present application relates to the technical field of medical data processing, in particular to a system for evaluating glutamic oxalacetic transaminase value, which comprises a data acquisition and standardization module, a feature engineering and proxy variable construction module, a multi-path joint regression prediction module, an interpretable risk stratification output module and a continuous learning and federal deployment module. The system takes blood lipids, body weight / body mass index and diastolic blood pressure as input features, innovatively constructs a thyroid function proxy score as an intermediate latent variable, realizes rapid regression prediction of glutamic oxalacetic transaminase value through a three-path parallel sub-network hierarchical fusion network, outputs point estimates and Bayesian confidence intervals; at the same time, based on the SHAP algorithm, an interpretable feature attribution report and a four-level risk stratification result are output, and multi-institutional compliance collaborative training is realized through horizontal federated learning and differential privacy mechanism.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically to a numerical assessment system for aspartate aminotransferase (AST). Background Technology

[0002] Aspartate aminotransferase (AST) is an important serum enzyme indicator clinically reflecting hepatocellular damage, myocardial damage, and thyroid dysfunction. AST is mainly distributed in the mitochondria and cytoplasm of hepatocytes, cardiomyocytes, and skeletal muscle cells. When these tissues and cells are damaged, a large amount of AST is released into the blood, leading to an increase in serum AST levels. Therefore, AST levels are one of the core biochemical indicators for assessing the severity of various diseases.

[0003] In traditional clinical testing procedures, obtaining AST values ​​relies on venous blood collection followed by analysis using a fully automated biochemical analyzer. This entire process typically takes 1 to 4 hours, which is insufficient to meet the needs for rapid decision-making in primary healthcare institutions and mobile health checkup scenarios. Furthermore, the cost of a complete set of biochemical tests is relatively high, which may lead some individuals to forgo these tests, resulting in missed diagnoses of individuals with potential liver function abnormalities.

[0004] In recent years, researchers have attempted to use machine learning methods to build auxiliary diagnostic models, but existing technologies have the following shortcomings: First, existing predictive models mostly use liver function-related indicators (such as alanine aminotransferase, alkaline phosphatase, etc.) directly to predict AST, and the acquisition of input features also relies on professional biochemical tests, which cannot fundamentally reduce costs and time consumption; Second, existing models ignore the indirect influence of thyroid function on AST and fail to incorporate the thyroid-lipid metabolism-weight-blood pressure system into the AST prediction framework, resulting in insufficient predictive sensitivity for elevated AST related to thyroid dysfunction; Third, existing AI-assisted diagnostic systems are mostly black-box outputs, lacking interpretable feature attribution mechanisms, making it difficult to gain the trust of clinicians; Fourth, existing systems rarely consider multi-institutional data privacy compliance issues, making it difficult to promote and apply them in the context of federal medical data.

[0005] Therefore, there is an urgent need for an assessment system that can use non-invasive or low-cost indicators routinely collected by health check centers or outpatient clinics to quickly predict AST values ​​through artificial intelligence models, while also being interpretable and privacy compliant. Summary of the Invention

[0006] The purpose of this invention is to provide a numerical evaluation system for aspartate aminotransferase (AST) to solve the technical problems of long detection time, high cost, lack of AI fast regression prediction model and uninterpretable prediction results in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A numerical assessment system for aspartate aminotransferase (AST) includes interconnected modules for data acquisition and standardization, feature engineering and proxy variable construction, multi-path joint regression prediction, interpretable risk stratification output, and continuous learning and federated deployment.

[0009] The data acquisition and standardization module is used to receive triglyceride, total cholesterol, low-density lipoprotein, high-density lipoprotein, weight, body mass index, diastolic blood pressure and age and gender information of the subjects, and to perform cross-institutional numerical normalization processing on the indicators and output standardized feature vectors.

[0010] The feature engineering and proxy variable construction module is used to calculate the thyroid function proxy score based on the standardized feature vector, and use the proxy score as an intermediate latent variable to input to the multi-path joint regression prediction module.

[0011] The multi-path joint regression prediction module is used to extract heterogeneous features based on three independent pathways: lipid metabolism pathway, energy metabolism pathway, and hemodynamic pathway. After fusing with the thyroid function proxy score, it outputs the predicted value of aspartate aminotransferase (AST) and its uncertainty range.

[0012] The interpretable risk stratification output module is used to generate individualized risk levels and feature contribution reports based on feature attribution algorithms.

[0013] The continuous learning and federated deployment module is used to collaboratively update the global prediction model through a federated learning framework without the data leaving the organization.

[0014] Furthermore, the thyroid function proxy score is calculated according to the following formula:

[0015] TF-Score = α1 × TC norm +α2×LDL norm +α3×TG norm +α4×(1 / HDL norm )+α5×BMI norm +α6×DBP norm The meanings of the parameters in the formula are as follows: BMI norm DBP norm The values ​​are: total cholesterol, low-density lipoprotein, triglycerides, high-density lipoprotein, body mass index, and diastolic blood pressure, respectively, after Z-score standardization; α1 to α6 are weight coefficients obtained by regression fitting from the training dataset; a higher TF-Score indicates a more significant tendency for hypothyroidism, while a lower value indicates hyperthyroidism or a tendency for normal to hyperthyroidism.

[0016] Furthermore, the multi-path joint regression prediction module includes a hierarchical parallel sub-network structure:

[0017] The first layer consists of three parallel sub-networks: a lipid sub-network that processes triglycerides, total cholesterol, low-density lipoprotein, and high-density lipoprotein; an energy metabolism sub-network that processes body weight and body mass index; and a hemodynamic sub-network that processes diastolic blood pressure.

[0018] The second layer is a fusion layer, which concatenates the output feature vectors of the three sub-networks with the thyroid function proxy score, and extracts cross-path feature associations through a fully connected interactive layer.

[0019] The output layer is a Bayesian uncertainty output layer, which simultaneously outputs the mean predicted value μ of AST and the prediction uncertainty σ, forming a dual-format output of point estimate and confidence interval.

[0020] Furthermore, the normalization process of the data acquisition and standardization module includes the following steps:

[0021] Step 1: Perform Z-score normalization on all continuous features;

[0022] Step 2: Introduce linear deviation correction factors based on reference material calibration for different testing equipment to eliminate systematic offsets across institutions;

[0023] Step 3: Perform tail truncation on extreme values ​​of indicators such as triglycerides that exceed the threshold range to prevent single-point extreme values ​​from damaging the model's stability.

[0024] Furthermore, the training strategy for the multi-path joint regression prediction module includes:

[0025] The Huber loss function is used instead of the mean squared error loss to reduce the excessive influence of extremely high AST samples on training.

[0026] For samples with severe abnormal aspartate aminotransferase levels, focus regression loss was used for weighted upsampling to address the class imbalance problem.

[0027] A five-fold cross-validation scheme based on organizational structure is adopted to prevent data leakage within the same organization from causing inflated performance.

[0028] Furthermore, the interpretable risk stratification output module calculates the contribution of each input feature to the current subject's predicted aspartate aminotransferase (AST) value based on the SHAP algorithm, and generates a personalized report containing the following:

[0029] Positive and negative contributions of each feature;

[0030] Thyroid function proxies score and their corresponding thyroid function tendencies;

[0031] The four-level risk stratification results include a green normal level (predicted value below 40 U / L), a yellow mildly abnormal level (40 to 80 U / L), an orange moderately abnormal level (80 to 200 U / L), and a red severely abnormal level (above 200 U / L).

[0032] Automatic confirmation testing recommendations are triggered for orange and red level test subjects to achieve intelligent allocation of testing resources.

[0033] Furthermore, the continuous learning and federated deployment module adopts a horizontal federated learning architecture, including:

[0034] Each participating node calculates the model gradient on its local data and uploads only the gradient parameters to the federated aggregation server, not the original data.

[0035] The federated aggregation server uses a weighted federated average algorithm to aggregate the gradients of each node and distributes the updated global model to each node.

[0036] During the gradient uploading process, a differential privacy stochastic gradient descent algorithm is used to add noise to the gradient to prevent the original subject data from being inferred from the gradient information.

[0037] Furthermore, the continuous learning and federated deployment module also includes a model drift monitoring subunit. The model drift monitoring subunit continuously monitors the deviation between the local predicted value distribution and the historical benchmark distribution based on KL divergence. When the deviation exceeds a preset threshold, the model fine-tuning process is automatically triggered. When the amount of newly added local validation data exceeds a preset number, incremental learning update is triggered. The incremental learning completes the adaptive update of the model without retraining the full amount of data.

[0038] Furthermore, the system also includes an individualized reference interval dynamic calibration unit. The dynamic calibration unit automatically adjusts the normal reference threshold of aspartate aminotransferase (AST) based on the subject's age, gender, and body mass index, replacing the method of applying a fixed threshold uniformly to all subjects, thereby reducing the misjudgment rate caused by population heterogeneity.

[0039] Furthermore, the inference latency of a single subject record in the multi-path joint regression prediction module does not exceed 100 milliseconds. The system supports model inference on the local terminal or edge server of the physical examination center without relying on external networks. The model is compressed through knowledge distillation or parameter quantization to meet the computing power constraints of edge deployment.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The input features used in this invention are all derived from non-invasive or low-cost items routinely collected by health check centers, eliminating the need for additional venous blood sampling or thyroid-specific testing, which can significantly reduce testing costs.

[0042] The inference latency of a single record in this invention system does not exceed 100 milliseconds, and it can output AST risk prediction immediately after the physical examination, without having to wait for a biochemical testing process of 1 to 4 hours.

[0043] This invention introduces a thyroid function proxy score and a multi-path hierarchical fusion network, which, compared to the traditional single fully connected regression model, can effectively capture the combined influence of multi-dimensional features of thyroid-lipid metabolism-weight-blood pressure on AST values.

[0044] The system of this invention outputs the specific contribution of each input feature to the current subject's AST prediction value based on the SHAP algorithm, so that doctors can intuitively understand the basis of the prediction conclusion;

[0045] The system of this invention adopts a horizontal federated learning framework and a differential privacy gradient protection mechanism. Under the premise that the original patient data does not leave the institution, a high-performance prediction model can be collaboratively trained in a multi-center environment, which meets the requirements of HIPAA and domestic medical data security regulations.

[0046] This invention achieves precise allocation of testing resources by using a four-level risk stratification and intelligent triage mechanism to trigger confirmatory testing recommendations only for high-risk groups. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the data flow of an aspartate aminotransferase (AST) numerical evaluation system disclosed in an embodiment of this application. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some embodiments of this invention, but not all embodiments.

[0049] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0051] I. System Overall Architecture;

[0052] refer to Figure 1The data acquisition and standardization module is used to receive triglyceride, total cholesterol, low-density lipoprotein, high-density lipoprotein, weight, body mass index, diastolic blood pressure and age and gender information of the subjects, and to perform cross-institutional numerical normalization processing on the indicators and output standardized feature vectors.

[0053] The feature engineering and proxy variable construction module is used to calculate the thyroid function proxy score based on the standardized feature vector, and use the proxy score as an intermediate latent variable to input to the multi-path joint regression prediction module.

[0054] The multi-path joint regression prediction module is used to extract heterogeneous features based on three independent pathways: lipid metabolism pathway, energy metabolism pathway, and hemodynamic pathway. After fusing with the thyroid function proxy score, it outputs the predicted value of aspartate aminotransferase (AST) and its uncertainty range.

[0055] The interpretable risk stratification output module is used to generate individualized risk levels and feature contribution reports based on feature attribution algorithms.

[0056] The continuous learning and federated deployment module is used to collaboratively update the global prediction model through a federated learning framework without the data leaving the organization.

[0057] The data acquisition and standardization module takes the original indicator data of the examinee as input data and outputs standardized feature vectors as output data.

[0058] The feature engineering and proxy variable construction module takes standardized feature vectors as input and outputs TF-Score and augmented feature vectors as output.

[0059] The multi-path joint regression prediction module takes augmented feature vectors as input and outputs the AST prediction mean μ and uncertainty σ.

[0060] The interpretable risk stratification output module takes AST predicted values ​​plus SHAP calculations as input data and outputs a four-level risk report and feature attribution.

[0061] The Continuous Learning and Federated Deployment module takes local gradients from each institution as input data and outputs the updated global model as output data.

[0062] II. Data Acquisition and Standardization Module;

[0063] 1. Input feature set;

[0064] This invention selects the following eight categories of indicators as system input features. All of these indicators can be obtained during routine examinations at health check centers or outpatient clinics without the need for additional invasive procedures: triglycerides are collected using a fully automated biochemical analyzer (venous blood); total cholesterol is collected using a fully automated biochemical analyzer (venous blood); low-density lipoprotein (LDL) is collected using a fully automated biochemical analyzer (venous blood); high-density lipoprotein (HDL) is collected using a fully automated biochemical analyzer (venous blood); weight is collected using an electronic scale (non-invasive); body mass index (BMI) is collected using weight / height² (derived calculation, non-invasive); diastolic blood pressure is collected using an electronic blood pressure monitor (non-invasive); age and gender are registered through basic information.

[0065] It should be noted that the venous blood sampling for the four lipid tests is usually completed in the same batch as other routine physical examination blood tests. It is not a blood sampling item specifically added for this system, and therefore does not constitute an additional cost burden.

[0066] 2. Data standardization process;

[0067] The data acquisition and standardization module performs the following three standardization steps on the collected raw indicator data in sequence:

[0068] Step 1: Z-score standardization. Perform Z-score standardization on all continuous features. The formula is: x norm =(x-μ train ) / σ train , where μ train and σ train These are the mean and standard deviation of the feature in the training set, respectively, and are fixed as constants during model deployment to ensure normalization consistency during the inference phase.

[0069] Step Two: Introduce a linear deviation correction factor based on reference material calibration for different testing equipment to eliminate systematic bias across institutions. Due to systematic numerical biases between different brands of biochemical analyzers, the data acquisition and standardization module introduces a linear deviation correction factor based on reference material calibration. For each device, the correction factor is obtained through a comparative experiment with a standard reference material, using the formula: x corrected =k device ×x raw +b device Where k device and b device These are device-specific correction coefficients, entered by the administrator during system initialization.

[0070] Step 3: Extreme value truncation. Extreme values ​​in indicators such as triglycerides that exceed reasonable threshold ranges are truncated and replaced with corresponding truncation boundary values ​​to prevent single-point extreme values ​​from destructively affecting the model's inference results.

[0071] III. Feature Engineering and Proxy Variable Construction Module;

[0072] 1. Design background of the thyroid function proxies score;

[0073] Thyroid function is an important indirect factor affecting AST levels, and its influence on AST is mainly achieved through the following three biological pathways:

[0074] Lipid metabolism pathway: When thyroid hormone is insufficient, hepatocytes express less low-density lipoprotein receptor, leading to the accumulation of total cholesterol, low-density lipoprotein and triglycerides in the blood, which in turn drives the development of non-alcoholic fatty liver disease, causes inflammatory damage to hepatocytes, and increases AST.

[0075] Energy metabolism pathway: Thyroid hormone levels indirectly affect body weight and body mass index by regulating the body's basal metabolic rate. Hypothyroid patients have a decreased basal metabolic rate, weight gain, and an increased risk of fatty liver, which in turn affects AST.

[0076] Hemodynamic pathway: In hypothyroidism, peripheral vascular resistance increases, leading to elevated diastolic blood pressure; in hyperthyroidism, peripheral vasodilation occurs, resulting in decreased diastolic blood pressure. Changes in diastolic blood pressure reflect the state of thyroid function, and elevated diastolic blood pressure can lead to hepatic microcirculatory disturbances, thereby affecting AST (assay stress).

[0077] However, routine medical examinations at health check centers typically do not include thyroid function testing, making it impossible to directly characterize the aforementioned influencing pathways. Therefore, this invention proposes an innovative design of a "thyroid function proxy score," which indirectly quantifies thyroid function predisposition using available blood lipid, weight, and blood pressure indicators.

[0078] 2. The thyroid function proxy score is calculated according to the following formula;

[0079] TF-Score is calculated using the following weighted linear combination formula: TF-Score = α1 × TC norm +α2×LDL norm +α3×TG norm +α4×(1 / HDL norm )+α5×BMI norm +α6×DBP norm The meanings of the parameters in the formula are as follows: TC norm LDL norm TG norm HDL norm BMI norm DBP normThese are the dimensionless values ​​of total cholesterol, LDL cholesterol, triglycerides, HDL cholesterol, body mass index, and diastolic blood pressure after Z-score standardization; α1 to α6 are the weight coefficients obtained by fitting the training dataset through multiple linear regression; HDL-C is the reciprocal of (1 / HDL). norm This is because high-density lipoprotein has a protective effect on the cardiovascular system and liver, and is negatively correlated with the risk of hypothyroidism. Taking the reciprocal makes its change direction consistent with the other characteristics.

[0080] The biological significance of TF-Score is as follows: A high TF-Score suggests a possible tendency towards hypothyroidism, accompanied by elevated total cholesterol, LDL cholesterol, triglycerides, and BMI, as well as elevated diastolic blood pressure and low HDL cholesterol. Conversely, a low TF-Score suggests a possible tendency towards hyperthyroidism or euthyroidism, with the aforementioned indicators showing the opposite trend. TF-Score itself is not a direct surrogate indicator of thyroid function but rather serves as an intermediate latent variable in predictive models to enhance the models' ability to identify thyroid-related AST abnormalities.

[0081] 3. Weight coefficient training method;

[0082] The weighting coefficients α1 to α6 were obtained through the following steps: Using a historical dataset containing measured thyroid function indicators, TSH values ​​were used as the dependent variable, and TC values ​​were used as the weighting coefficients. norm LDL norm TG norm HDL norm BMI norm DBP norm Using α1 to α6 as independent variables, ridge regression is performed for fitting, and L2 regularization is used to prevent overfitting. The coefficients α1 to α6 are then fixed as model parameters. In institutions without historical TSH data, reference weight coefficients reported in the literature can be used for initialization, and then fine-tuned using local data.

[0083] IV. Multi-path joint regression prediction module;

[0084] 1. Network architecture design;

[0085] The multi-path joint regression prediction module of this invention adopts a hierarchical parallel sub-network structure, which consists of the following three layers:

[0086] The first parallel feature extraction layer consists of three parallel sub-networks: a lipid sub-network for processing triglycerides, total cholesterol, low-density lipoprotein, and high-density lipoprotein; an energy metabolism sub-network for processing body weight and body mass index; and a hemodynamic sub-network for processing diastolic blood pressure.

[0087] The lipid subnetwork corresponds to the lipid metabolism pathway: TC norm LDL norm TG norm HDL norm Using four blood lipid indicators as input, a combined blood lipid feature vector f is extracted through 2 to 3 layers of fully connected neurons. lipid This subnetwork is specifically designed to capture the degree of lipid metabolism disorder and its association with hepatic steatosis.

[0088] Energy metabolism subnetwork (corresponding energy metabolism pathway): based on Weight norm and BMI norm As input, feature vectors f of body weight and metabolic load are extracted through 1 to 2 layers of fully connected neurons. metabolism This subnetwork is specifically designed to capture the correlation between the body's energy metabolism status and liver metabolic load.

[0089] Hemodynamic subnetwork (corresponding to hemodynamic pathway): based on DBP norm As input, a blood pressure feature vector f is extracted through 1 to 2 layers of fully connected neurons. hemodynamic This subnetwork is specifically designed to capture the association features between hemodynamic state and liver microcirculation.

[0090] The three parallel sub-networks are trained independently to avoid mutual interference between features of different pathological pathways, while preserving the independent pathological significance of each dimension of features.

[0091] The second layer (cross-path feature fusion layer): combines the output feature vectors f from the three sub-networks. lipid f metabolism f hemodynamic The TF-Score is concatenated with the TF-Score along its dimensions to form a fused feature vector f. concat By learning cross-path feature associations through 2 to 3 fully connected interactive layers, the synergistic effect features between various pathological pathways are extracted.

[0092] Output Layer (Bayesian Uncertainty Output Layer): A Bayesian deep learning output layer is used to output the mean μ of the AST predictions and the prediction uncertainty σ. The mean μ is used as the point estimate prediction of the AST, and the uncertainty σ is used to construct a 95% confidence interval [μ-1.96σ,μ+1.96σ]. This provides clinicians with a quantifiable measure of the reliability of the prediction results, helping them distinguish between high-confidence and low-confidence predictions, and thus make more prudent clinical decisions.

[0093] 2. Model training strategy;

[0094] The model training process employs the following three key strategies:

[0095] Strategy 1: Huber Loss Function. The Huber loss function replaces the traditional mean squared error (MSE) loss function. The Huber loss function is equivalent to MSE when the prediction error is small, but becomes a linear loss when the prediction error exceeds a threshold δ. Its advantage lies in the fact that clinical pathological conditions such as acute hepatitis and drug-induced liver injury can cause the AST (Average Stress Tolerance) to rise sharply to tens of times the normal value. If MSE is used, the high error terms of these extreme value samples will dominate gradient updates, leading to a decrease in the model's prediction accuracy within the normal range. The Huber loss effectively reduces the excessive influence of extremely high-value samples on training, improving the model's prediction accuracy within the normal and mildly abnormal ranges.

[0096] Strategy 2: Focus Regression Loss Weighted for Imbalanced Samples: Cases with severe AST abnormalities account for a very low percentage of the physical examination population, constituting a serious sample imbalance problem. This invention draws on the FocalLoss concept from the field of object detection and proposes a focus regression loss method. This method weights and upsamples the loss of samples with high AST values, with the weights being positively correlated with the magnitude of the prediction error. This allows the model to fully learn the characteristic patterns of severely abnormal samples, improving its ability to identify high-risk individuals.

[0097] Strategy 3: Five-fold cross-validation by institution: During the model evaluation phase, a five-fold cross-validation scheme by institution is adopted to ensure that the test subject data of the same institution will not appear in both the training set and the validation set at the same time. This prevents the validation performance from being artificially high due to the similarity of test subject features within the same institution, and ensures that the model evaluation results truly reflect the cross-institution generalization performance.

[0098] 3. Performance evaluation indicators;

[0099] This invention sets the following performance evaluation indicators for the multi-path joint regression prediction module as admission criteria for system deployment:

[0100] Mean absolute error; target value ≤ 8 U / L; average deviation between predicted and measured values ​​does not exceed 8 U / L, meeting the accuracy requirements for clinical risk stratification;

[0101] Coefficient of determination; target value ≥ 0.70; the model can explain more than 70% of the AST numerical variation and has practical predictive value;

[0102] Abnormal detection rate; target value ≥ 85%; recall rate not less than 85% for the clinically most important AST abnormality threshold;

[0103] False positive rate; target value ≤15%; the proportion of normal subjects being misjudged as abnormal does not exceed 15%, and alarm fatigue is controlled;

[0104] Single record inference latency; target value <100ms; meets the performance requirements of real-time feedback at the physical examination site and edge deployment scenarios.

[0105] V. Explainable Risk Stratification Output Module; The Explainable Risk Stratification Output Module calculates the contribution of each input feature to the predicted value of the current subject's aspartate aminotransferase (AST) based on the SHAP algorithm, and generates an individualized report containing the following: the positive and negative contributions of each feature; the thyroid function proxy score and its corresponding thyroid function tendency indication;

[0106] 1. SHAP feature attribution mechanism;

[0107] This invention employs the SHAP algorithm to perform feature attribution analysis on the output of the prediction model. SHAP is based on the Shapley value principle in game theory, quantifying the impact of each feature on the prediction result by calculating the marginal contribution of each feature in all combinations of feature subsets.

[0108] For each subject, the interpretable risk stratification output module calculates the SHAP contribution of eight input features to the current AST predicted value μ. Each contribution has a sign: a positive value indicates that the feature increases the predicted value, and a negative value indicates that the feature decreases the predicted value. The sum of the SHAP contributions of each feature is equal to the difference between the predicted value μ and the baseline value.

[0109] The SHAP output is presented to doctors as a visual waterfall chart with accompanying text descriptions, such as: "This subject's predicted AST value is 63 U / L, higher than the upper limit of normal. The main contributing factors are: elevated triglycerides (+12.3 U / L), elevated BMI (+8.1 U / L), and elevated diastolic blood pressure (+4.2 U / L); high-density lipoprotein provides a protective correction to the predicted value (-3.5 U / L). TF-Score=0.73, suggesting a mild tendency towards hypothyroidism. Further confirmation with thyroid-stimulating hormone testing is recommended."

[0110] 2. Four-level risk stratification and intelligent triage;

[0111] Based on the AST predicted mean μ, the risk stratification output module can be interpreted to classify subjects into the following four risk levels and provide corresponding clinical recommendations:

[0112] Risk level: Normal; marked in green; AST < 40; no special treatment required, regular check-ups and follow-ups are recommended;

[0113] Risk level: Mildly abnormal; marked in yellow; AST (40~80); It is recommended to complete a full set of biochemical tests, and pay attention to fatty liver and thyroid function;

[0114] Risk level: Moderately abnormal; marked in orange; AST (80~200); It is recommended to complete the AST biochemical verification test on the same day, follow up at the outpatient clinic, and assess whether fatty liver or hypothyroidism exists;

[0115] Risk level: Severely abnormal; marked in red; AST>200; suggests possible acute liver injury, requiring immediate AST confirmation test to assess whether emergency treatment or referral is needed.

[0116] The core logic of the intelligent triage mechanism is that the system only proactively triggers the "suggest completing AST biochemical verification on the same day" prompt for orange and red-level examinees, rather than recommending high-cost biochemical testing for all examinees. This mechanism allows medical examination centers to concentrate limited biochemical testing resources on high-risk groups, effectively controlling the false negative rate while avoiding unnecessary testing costs and psychological burdens on low-risk groups.

[0117] 3. Individualized reference interval dynamic calibration;

[0118] The traditional AST normal reference range is based on statistical data of the general population, which may have systematic biases for specific subgroups, leading to misjudgments.

[0119] The individualized reference interval dynamic calibration unit of this invention dynamically adjusts the normal reference threshold of AST based on the subject's age, gender, and BMI. Specifically, during model training, the training data is stratified by age (20-39 years, 40-59 years, and over 60 years), gender (male / female), and BMI (<24, 24-28, >28). The 95th percentile of the normal AST range in each subgroup is calculated as the dynamic reference upper limit for that subgroup. During inference, the corresponding reference upper limit threshold is used based on the subject's subgroup, replacing the uniform fixed threshold for risk stratification judgment.

[0120] VI. Continuous Learning and Federal Deployment Module;

[0121] 1. Horizontal federated learning architecture;

[0122] Each participating node calculates the model gradient on its local data and uploads only the gradient parameters, not the original data, to the federated aggregation server. The federated aggregation server uses a weighted federated averaging algorithm to aggregate the gradients of each node and distributes the updated global model to each node. During the gradient upload process, a differential privacy stochastic gradient descent algorithm is used to add noise to the gradient to prevent the original subject data from being inferred from the gradient information.

[0123] The specific collaboration process includes the following steps:

[0124] S1. The federated aggregation server distributes the current global model parameters to all participating nodes;

[0125] S2. Each node performs forward and backward propagation on local patient data and calculates the local gradient;

[0126] S3. Each node applies the differential privacy stochastic gradient descent algorithm to its local gradient, adding calibrated Gaussian noise to the gradient so that the noise size satisfies the (ε,δ) differential privacy budget constraint, preventing the original patient data from being inferred from the gradient information.

[0127] S4. Each node uploads the noisy gradient parameters to the federated aggregation server;

[0128] S5. The federated aggregation server uses a weighted federated average or federated near-end gradient algorithm to aggregate the gradients of each node. The weights are usually proportional to the amount of local data on each node.

[0129] S6. The aggregation server distributes the updated global model parameters to each node, completing a round of federated updates;

[0130] S7. Repeat the above steps until the model converges.

[0131] Through the aforementioned horizontal federated learning framework, this system can improve the generalization performance of the model by utilizing subject data from different regions and population distributions, even without sharing data among multiple centers. This effectively overcomes the problems of insufficient training data in single centers and population distribution bias.

[0132] 2. Model drift monitoring and incremental learning;

[0133] The continuous learning and federated deployment module also includes a model drift monitoring subunit. The model drift monitoring subunit continuously monitors the deviation between the local predicted value distribution and the historical benchmark distribution based on KL divergence. When the deviation exceeds a preset threshold, the model fine-tuning process is automatically triggered. When the amount of newly added local validation data exceeds a preset number, incremental learning update is triggered. The incremental learning completes the adaptive update of the model without retraining the full amount of data.

[0134] 3. Lightweight edge deployment;

[0135] To enable the system to run independently on local terminals or edge servers in the health check center, without relying on external network connections, this invention performs the following lightweight processing on the multi-path joint regression prediction model:

[0136] Knowledge distillation: Using the soft-label output of the complete network to guide the training of a lightweight small network, while reducing the number of parameters by 60% to 80%, the prediction accuracy of the student model is maintained at more than 95% of that of the teacher model.

[0137] Parameter quantization: The model weights are quantized from 32-bit floating-point precision to 8-bit integer precision, compressing the model size to one-quarter of its original size without significantly losing prediction accuracy, thus greatly reducing the memory and computing resources required for inference.

[0138] Structural pruning: Pruning the connection weights in the model whose contribution is below a threshold removes redundant parameters, further reducing the model size and improving inference speed.

[0139] After the above lightweighting process, the system's single-record inference latency on ordinary medical terminals can reach less than 100 milliseconds. The system supports model inference on local terminals or edge servers in the physical examination center without relying on external networks. The model is compressed through knowledge distillation or parameter quantization to meet the computing power constraints of edge deployment.

[0140] Example 1: Standard application scenario in a health checkup center;

[0141] A tertiary-level Class A hospital's physical examination center has adopted the aspartate aminotransferase (AST) numerical assessment system described in this invention. The system is integrated into the physical examination report information system, and the specific business process is as follows:

[0142] A 42-year-old male subject completed a routine physical examination at a health checkup center. The data collected are as follows: TG = 2.8 mmol / L (high), TC = 6.2 mmol / L (high), LDL-C = 4.1 mmol / L (high), HDL-C = 0.9 mmol / L (low), weight = 88 kg, BMI = 30.2 kg / m². 2 (Obesity), diastolic blood pressure = 92 mmHg (high), age 42, male.

[0143] The M1 module performs Z-score standardization and device calibration on the above data, outputting a standardized feature vector. The feature engineering and proxy variable construction module calculates a TF-Score of 0.81, indicating a mild tendency towards hypothyroidism. The multi-path joint regression prediction module uses three-path parallel sub-network fusion inference to output an AST predicted value of μ=71U / L, with a confidence interval of [58U / L, 84U / L], and an inference time of approximately 35 milliseconds. The continuous learning and federated deployment module outputs a yellow risk level (mildly abnormal), and the SHAP attribution report shows that the main contributing factors are: high triglycerides (+14.2U / L), high BMI (+10.3U / L), high LDL cholesterol (+6.8U / L), and a protective correction for HDL cholesterol (-4.1U / L). The system also outputs a TF-Score thyroid warning, suggesting further TSH testing. The system does not trigger orange or red alerts, and the examinee's medical examination report includes a yellow risk label and a recommendation for follow-up examination.

[0144] The subject subsequently underwent AST biochemical verification testing on the same day, with an actual value of 68 U / L. The absolute error between this value and the system's predicted value of 71 U / L was 3 U / L, falling within the target range of MAE ≤ 8 U / L. Simultaneously, the TSH test result was 8.2 mIU / L (high), confirming hypothyroidism, consistent with the conclusion indicated by the TF-Score.

[0145] Example 2: Outpatient rapid assessment scenario;

[0146] At a community health service center outpatient clinic, a patient presented with fatigue and poor appetite, making it impossible to immediately obtain AST biochemical test results. The outpatient doctor used the system to input the patient's most recent physical examination blood lipid data (TG=3.5mmol / L, TC=7.1mmol / L, LDL-C=4.8mmol / L, HDL-C=0.8mmol / L), current weight (92kg, BMI=32.1), and diastolic blood pressure (98mmHg). Within approximately 50 milliseconds, the system output an AST predicted value μ=134U / L (confidence interval [108U / L, 160U / L]), with a risk level of orange (moderately abnormal), and recommended completing the AST confirmatory test on the same day.

[0147] Based on this, the doctor prioritized the patient's urgent liver function test, which ultimately yielded an AST level of 118 U / L, leading to a diagnosis of non-alcoholic steatohepatitis complicated by hypothyroidism. The system completed a risk pre-assessment while awaiting the biochemical results, providing valuable early information for the doctor's diagnostic and treatment decisions.

[0148] Example 3: Multi-center Federated Learning Scenario;

[0149] Three medical examination institutions (Institution A, B, and C) located in different regions jointly participated in the federated learning collaborative training of this system. Each institution installed a local federated node client, with the following data volumes: Institution A 8,000 cases, Institution B 5,200 cases, and Institution C 3,400 cases, totaling 16,600 training data cases. However, the data of each institution was not shared with each other.

[0150] After 20 rounds of federated updates, the global model achieved an average MAE of 6.8 U / L on the independent test sets of the three institutions, which is better than single-institution training (MAE of 8.3 U / L for institution A, 9.1 U / L for institution B, and 11.2 U / L for institution C). This demonstrates that federated learning effectively utilizes the complementarity of multi-center data, improves the model's cross-institution generalization performance, and ensures that no original patient data is transferred across institutions, thus meeting the requirements for data sovereignty protection.

Claims

1. A numerical evaluation system for aspartate aminotransferase (AST), characterized in that, The system includes interconnected modules for data acquisition and standardization, feature engineering and proxy variable construction, multi-path joint regression prediction, interpretable risk stratification output, and continuous learning and federated deployment. The data acquisition and standardization module is used to receive triglyceride, total cholesterol, low-density lipoprotein, high-density lipoprotein, weight, body mass index, diastolic blood pressure and age and gender information of the subjects, and to perform cross-institutional numerical normalization processing on the indicators and output standardized feature vectors. The feature engineering and proxy variable construction module is used to calculate the thyroid function proxy score based on the standardized feature vector, and use the proxy score as an intermediate latent variable to input to the multi-path joint regression prediction module. The multi-path joint regression prediction module is used to extract heterogeneous features based on three independent pathways: lipid metabolism pathway, energy metabolism pathway, and hemodynamic pathway. After fusing with the thyroid function proxy score, it outputs the predicted value of aspartate aminotransferase (AST) and its uncertainty range. The interpretable risk stratification output module is used to generate individualized risk levels and feature contribution reports based on feature attribution algorithms. The continuous learning and federated deployment module is used to collaboratively update the global prediction model through a federated learning framework without the data leaving the organization.

2. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The thyroid function proxy score is calculated using the following formula: TF-Score = α1 × TC norm +α2×LDL norm +α3×TG norm +α4×(1 / HDL norm )+α5×BMI norm +α6×DBP norm The meanings of the parameters in the formula are as follows: BMI norm DBP norm The values ​​are: total cholesterol, low-density lipoprotein, triglycerides, high-density lipoprotein, body mass index, and diastolic blood pressure, respectively, after Z-score standardization; α1 to α6 are weight coefficients obtained by regression fitting from the training dataset; a higher TF-Score indicates a more significant tendency for hypothyroidism, while a lower value indicates hyperthyroidism or a tendency for normal to hyperthyroidism.

3. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The multi-path joint regression prediction module includes a hierarchical parallel sub-network structure: The first layer consists of three parallel sub-networks: a lipid sub-network that processes triglycerides, total cholesterol, low-density lipoprotein, and high-density lipoprotein; an energy metabolism sub-network that processes body weight and body mass index; and a hemodynamic sub-network that processes diastolic blood pressure. The second layer is a fusion layer, which concatenates the output feature vectors of the three sub-networks with the thyroid function proxy score, and extracts cross-path feature associations through a fully connected interactive layer. The output layer is a Bayesian uncertainty output layer, which simultaneously outputs the mean predicted value μ of AST and the prediction uncertainty σ, forming a dual-format output of point estimate and confidence interval.

4. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The normalization process of the data acquisition and standardization module includes the following steps: Step 1: Perform Z-score normalization on all continuous features; Step 2: Introduce linear deviation correction factors based on reference material calibration for different testing equipment to eliminate systematic offsets across institutions; Step 3: Perform tail truncation on extreme values ​​of indicators such as triglycerides that exceed the threshold range to prevent single-point extreme values ​​from damaging the model's stability.

5. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The training strategy for the multi-path joint regression prediction module includes: The Huber loss function is used instead of the mean squared error loss to reduce the excessive influence of extremely high AST samples on training. For samples with severe abnormal aspartate aminotransferase levels, focus regression loss was used for weighted upsampling to address the class imbalance problem. A five-fold cross-validation scheme based on organizational structure is adopted to prevent data leakage within the same organization from causing inflated performance.

6. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The interpretable risk stratification output module calculates the contribution of each input feature to the current subject's predicted aspartate aminotransferase (AST) value based on the SHAP algorithm, and generates a personalized report containing the following: Positive and negative contributions of each feature; Thyroid function proxy score and its corresponding thyroid function tendency indication; The four-level risk stratification results include a green normal level (predicted value below 40 U / L), a yellow mildly abnormal level (40 to 80 U / L), an orange moderately abnormal level (80 to 200 U / L), and a red severely abnormal level (above 200 U / L). Automatic confirmation testing recommendations are triggered for orange and red-level test subjects to achieve intelligent allocation of testing resources.

7. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The continuous learning and federated deployment module adopts a horizontal federated learning architecture, including: Each participating node calculates the model gradient on its local data and uploads only the gradient parameters to the federated aggregation server, not the original data. The federated aggregation server uses a weighted federated average algorithm to aggregate the gradients of each node and distributes the updated global model to each node. During the gradient uploading process, a differential privacy stochastic gradient descent algorithm is used to add noise to the gradient to prevent the original subject data from being inferred from the gradient information.

8. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The continuous learning and federated deployment module also includes a model drift monitoring subunit. The model drift monitoring subunit continuously monitors the deviation between the local predicted value distribution and the historical benchmark distribution based on KL divergence. When the deviation exceeds a preset threshold, the model fine-tuning process is automatically triggered. When the amount of newly added local validation data exceeds a preset number, incremental learning update is triggered. The incremental learning completes the adaptive update of the model without retraining the full amount of data.

9. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The system also includes an individualized reference interval dynamic calibration unit. The dynamic calibration unit automatically adjusts the normal reference threshold of aspartate aminotransferase (AST) based on the subject's age, gender, and body mass index, instead of applying a fixed threshold uniformly to all subjects, thereby reducing the misjudgment rate caused by population heterogeneity.

10. The aspartate aminotransferase numerical assessment system according to claim 1, characterized in that: The inference latency of a single subject record in the multi-path joint regression prediction module does not exceed 100 milliseconds. The system supports model inference on the local terminal or edge server of the physical examination center without relying on external networks. The model is compressed through knowledge distillation or parameter quantization to meet the computing power constraints of edge deployment.

Images