Risk prediction model for non-alcoholic fatty liver and method for constructing the same

Abstract

The application relates to the field of medical health prediction, and specifically discloses a risk prediction model for non-alcoholic fatty liver and a construction method thereof, which comprises the following steps: acquiring multi-source medical data of a subject, calculating a heterogeneity index between different data sources, constructing an adaptive weighted fusion function, and generating a unified risk representation vector; constructing a reference framework containing multiple risk anchor points, comparing the risk representation vector with each risk anchor point, and determining a current risk level; extracting a core risk driving force through differential operation, constructing a nonlinear evolution model to predict a risk state transition probability; and based on historical prediction error feedback, dynamically adjusting the fusion weight and the model parameters, and generating a risk analysis report. Through heterogeneity quantification and adaptive fusion, the application solves the problems of low risk assessment accuracy and poor robustness caused by distribution differences of multi-source medical data, and realizes interpretable, dynamic and accurate prediction of the risk of non-alcoholic fatty liver.

Description

Technical Field

[0001] This invention belongs to the field of medical and health prediction, specifically a risk prediction model for non-alcoholic fatty liver disease and its construction method. Background Technology

[0002] Nonalcoholic fatty liver disease (NAFLD) is a metabolic liver disease characterized by excessive fat deposition in hepatocytes. Its spectrum includes simple fatty liver, nonalcoholic steatohepatitis, liver fibrosis, cirrhosis, and even hepatocellular carcinoma. It is closely associated with components of metabolic syndrome such as obesity, insulin resistance, type 2 diabetes, and dyslipidemia, and has become one of the most common chronic liver diseases worldwide. Currently, the diagnosis of NAFLD mainly relies on imaging examinations such as liver ultrasound and transient elastography, as well as serum liver enzyme level detection. In terms of risk assessment, existing technologies mostly employ threshold judgment methods based on single biomarkers or use regression models to linearly weight multiple risk factors, such as the fatty liver index and the hepatic steatosis index.

[0003] However, existing risk assessment methods do not fully consider the heterogeneous distribution differences among multi-source medical data. Specifically, data from different testing methods, time points, and medical institutions exhibit significant differences in characteristic distributions. For example, imaging data and serological data have different statistical characteristics, and fasting blood glucose and postprandial blood glucose have different dynamic fluctuation patterns. Existing methods typically treat all data sources equally or use fixed-weight fusion. When a data source has noise, missing data, or distribution bias, it directly affects the accuracy and stability of the risk assessment results. Therefore, how to achieve adaptive fusion while considering the heterogeneity of multi-source data and construct an accurate and robust risk prediction model for non-alcoholic fatty liver disease is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a risk prediction model for non-alcoholic fatty liver disease and its construction method, thereby resolving the issues of low accuracy and poor robustness in existing risk assessment methods due to the failure to consider the heterogeneous distribution differences among multi-source medical data.

[0005] This invention provides a method for constructing a risk prediction model for non-alcoholic fatty liver disease, comprising the following steps:

[0006] S1. Obtain multi-source medical data of the subjects, wherein the multi-source medical data includes at least two of the following: metabolic index sequences, lifestyle time-series data, and biomarker data; for each data source, extract its dynamic change characteristics and static baseline characteristics; calculate the heterogeneity index between different data sources, and construct an adaptive weighted fusion function based on the heterogeneity index to generate a unified risk characterization vector;

[0007] S2. Construct a reference framework containing multiple risk anchors, each risk anchor corresponding to a preset non-alcoholic fatty liver disease risk subtype; compare the risk representation vector with each risk anchor, calculate the multidimensional distance metric, and determine the subject's current risk level according to the minimum distance principle;

[0008] S3. Based on the risk characterization vector, extract the core risk drivers through differential operations. The core risk drivers include the cumulative amount of metabolic load, the level of inflammatory stress, and the rate of change of insulin resistance. Construct a nonlinear evolution model driven by the risk drivers to predict the probability of risk state transition of the subject within a preset time window.

[0009] S4. Based on historical prediction error feedback, dynamically adjust the fusion weights of the adaptive weighted fusion function and the model parameters of the nonlinear evolution model; generate a risk analysis report for the subjects based on the prediction results, the risk analysis report including the current risk level, future risk evolution trajectory, and the contribution ranking of key risk driving factors.

[0010] Preferably, the heterogeneity index The calculation method is as follows:

[0011]

[0012] in, and Indicates different data source indexes, and These represent data sources respectively. and data source Extracted feature vectors, Describing covariance, Represents variance. This represents the KL divergence between the feature distributions of two data sources. and Representing data sources respectively and data source The empirical probability distribution of the feature vectors, This is the scaling factor, and its value range is... ;

[0013] The adaptive weighted fusion function is constructed as follows:

[0014]

[0015] in, Indicates the total number of data sources. and Indicates the index of the data source. Represents a unified risk representation vector. Indicates the first Feature vectors of the data source Indicates the first Adaptive fusion weights of the data source Indicates the first The average heterogeneity index of the data source with all other data sources. This is the temperature coefficient, and its value range is... .

[0016] Preferably, the risk anchor point ,in Index representing risk anchor points, Represents the anchor center vector. Represents the covariance matrix. Indicates the effective radius of influence of the risk anchor point;

[0017] The multidimensional distance metric is a weighted combination of Mahalanobis distance and local density factor:

[0018]

[0019] in, Indicates the first The local data density of the region where each risk anchor point is located is calculated using the kernel density estimation method. This represents the maximum local density among all anchor points;

[0020] The current risk level is determined according to the following formula:

[0021]

[0022] like If the radius constraint of any anchor point is not met, it is determined to be an edge risk state.

[0023] Preferably, the extraction method for the core risk drivers is as follows:

[0024] The cumulative metabolic load Defined as:

[0025]

[0026] in, Indicates time, Represents the integral variable. This represents a time series of the insulin resistance index. Indicates blood sugar level, Indicates triglyceride levels, The attenuation factor has a value range of 1. ;

[0027] The level of inflammatory stress Defined as:

[0028]

[0029] in, and Indicates liver enzyme levels, Indicates C-reactive protein level, and It is an adjustable parameter. , ;

[0030] The rate of change of insulin resistance Defined as:

[0031]

[0032] in, This represents the time series of the insulin resistance index as assessed by the steady-state model. It represents the first derivative with respect to time.

[0033] Preferably, the nonlinear evolution model is as follows:

[0034]

[0035] in, express The risk state value at any given time, with a range of values ​​being: , This represents a cumulative risk auxiliary variable. The mean is variance is Gaussian white noise, , , , For model parameters, ;

[0036] The risk state transition probability is obtained by applying the above differential equation system within a time window. Numerical integration yields:

[0037]

[0038] in, The cumulative distribution function represents the standard normal distribution. , Representing state The corresponding mean and standard deviation, This represents the prediction time window, with a value range of [value missing]. The state , Selected from different states in low-risk steady state, medium-risk evolution state, and high-risk early warning state;

[0039] Wherein, the low-risk steady state corresponds to The intermediate risk evolution state corresponds to The high-risk warning state corresponds to .

[0040] Preferably, the adaptive calibration defines a prediction error function. :

[0041]

[0042] in, Indicates the length of the time window. This represents the predicted risk representation vector. This represents the risk representation vector based on actual observations. This represents the square of the Euclidean distance. Represents the regularization term, , Indicates the index of the model parameter. The regularization coefficient is . , Represents the set of fusion weights. , Represents the set of model parameters;

[0043] The dynamic update rule for the fusion weights is as follows:

[0044]

[0045] in, Indicates the index of the data source. Index representing the number of iterations and , Indicates the learning rate. , Indicates the momentum coefficient. , This indicates the direction of the last weight update.

[0046] Preferably, the contribution ranking of the key risk drivers is calculated in the following manner:

[0047] For any output result The output result For risk level or transition probability, the first Contribution of each risk driver Defined as:

[0048]

[0049] in, Index representing risk drivers, , Indicates the first Each input feature Indicates the output for the first... The expected gradient of each input feature is approximated using the Monte Carlo sampling method. This represents the standard deviation of the characteristic;

[0050] The final output list of contributions, sorted by score, is as follows:

[0051]

[0052] in, This represents the total number of risk drivers, and the contribution ranking is used to provide interpretable decision support information to clinicians.

[0053] The present invention also provides a risk prediction model for non-alcoholic fatty liver disease, which is constructed using the above-described method.

[0054] The present invention also provides a risk prediction device for non-alcoholic fatty liver disease, comprising:

[0055] The data acquisition module is configured to acquire multi-source medical data from the subjects.

[0056] The heterogeneity quantization module is configured to calculate the heterogeneity index between different data sources and perform adaptive weighted fusion.

[0057] The risk anchor comparison module is configured to compare the subject's risk representation vector with pre-constructed risk anchors to determine the current risk level;

[0058] The evolution prediction module is configured to predict the probability of future risk state transitions based on a nonlinear evolution model driven by risk drivers.

[0059] The adaptive calibration module is configured to dynamically adjust the fusion weights and model parameters based on historical prediction error feedback;

[0060] The output module is configured to generate and output risk analysis reports for the subjects.

[0061] Preferably, the pre-built risk anchors in the risk anchor comparison module are dynamically updated according to the following rules:

[0062] When the amount of new subject data reaches a preset threshold When this occurs, the risk anchor point is recalculated:

[0063]

[0064] in, Index representing risk anchor points, Indicates the original number One risk anchor point, Indicates the original number The number of data points covered by each anchor point This represents the set of data points that have newly fallen within the influence range of this anchor point. Indicates the number of new data points;

[0065] Simultaneously, detect whether there are new data-dense areas; if so, and the minimum distance from all existing anchor points is greater than a preset threshold. Then create a new risk anchor point:

[0066]

[0067] in, This represents the center vector of the newly detected dense region. This indicates the local density of the dense area. and For the preset threshold, , .

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] By introducing a heterogeneity index and an adaptive weighted fusion mechanism, intelligent integration of multi-source medical data is achieved. This mechanism can automatically identify distributional differences between different data sources and assign lower fusion weights to data sources with higher noise or lower quality, thereby effectively improving the robustness and accuracy of the risk representation vector and solving the technical problem that traditional equal-weighted fusion methods are susceptible to interference from anomalous data.

[0070] This invention achieves interpretable risk hierarchy positioning by constructing a risk anchor reference framework. Each risk anchor corresponds to a clinically significant non-alcoholic fatty liver disease risk subtype, enabling risk assessment results to move beyond abstract numerical scores and directly map to specific pathological features and clinical phenotypes. This significantly enhances the interpretability of the assessment results and provides clinicians with an intuitive reference for diagnostic decisions.

[0071] This invention constructs a nonlinear evolution model based on a differential equation system, which effectively characterizes the dynamic process of disease progression. The cumulative metabolic load, inflammatory stress level, and insulin resistance evolution rate introduced in the model correspond to the core pathological links in the occurrence, development, and deterioration of non-alcoholic fatty liver disease, respectively, and have clear biological significance, enabling more accurate capture of the intrinsic laws of disease evolution.

[0072] This invention achieves dynamic optimization of model parameters by establishing an adaptive calibration mechanism. This mechanism can adjust the fusion weights and model parameters in real time based on historical prediction error feedback, enabling the model to adapt to differences in data distribution under different population characteristics and medical environments, and exhibiting good generalization ability and adaptability. Attached Figure Description

[0073] Figure 1 This is the overall flowchart of the core method of the present invention;

[0074] Figure 2 This is a flowchart of the risk anchor point comparison process for the present invention. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0076] Example 1: Refer to Figure 1 The first embodiment of the present invention provides a method for constructing a risk prediction model for non-alcoholic fatty liver disease, including the following steps.

[0077] S1: Heterogeneous Quantification and Adaptive Fusion of Multi-Source Medical Data

[0078] This step aims to acquire multi-source medical data from subjects and generate a unified risk representation vector through heterogeneity quantification and adaptive weighted fusion, providing a standardized data foundation for subsequent risk prediction.

[0079] (a) Data collection

[0080] Based on a medical big data platform, health parameters of subjects from different dimensions are collected in real time. In this embodiment, the data dimensions include three levels: metabolic indicator layer ( Lifestyle layer ) and biological marker layer ( The specific data collected is represented as follows:

[0081]

[0082] in, Indicates the data dimension hierarchy. Indicates the first Data collection at different levels Indicates the first The first in the hierarchy Each detection point data vector. The data vector is a multidimensional data vector, which includes at least liver function indicators (ALT, AST), blood glucose and lipids (fasting blood glucose, triglycerides, total cholesterol), body mass index (BMI), imaging data (liver fat content), and genetic information (PNPLA3 genotype). This indicates the number of detection points in each layer, with the same number of detection points in each layer.

[0083] At the same time, based on the historical database, the same historical data set as the real-time subject data is extracted. This is used for subsequent long-term trend feature extraction.

[0084] (II) Feature Extraction

[0085] For the collected multi-source medical data, dynamic change features and static baseline features are extracted respectively. Specifically:

[0086] For short-time fluctuation characteristics, short-time Fourier transform is used for extraction:

[0087]

[0088] For long-term development trend characteristics, a long short-term memory network is used for extraction:

[0089]

[0090] For multimodal association features, a graph neural network is used for extraction:

[0091]

[0092] Based on the extracted features, a health data feature matrix is ​​constructed:

[0093]

[0094] (III) Calculation of Heterogeneity Index

[0095] The degree of heterogeneity among different data sources reflects the varying contributions of each data source to disease risk assessment. In this embodiment, for any two data sources... and Its heterogeneity index Defined as:

[0096]

[0097] in, and Indicates different data source indexes, and These represent data sources respectively. and data source Extracted feature vectors, Describing covariance, Represents variance. This represents the KL divergence between the feature distributions of two data sources. and Representing data sources respectively and data source The empirical probability distribution of the feature vectors, This is the scaling factor, and its value range is... .when When the heterogeneity index degenerates into a measure based solely on covariance; when At this time, the influence of KL divergence reaches its maximum.

[0098] (iv) Adaptive weighted fusion

[0099] Based on the calculated heterogeneity index, an adaptive weighted fusion function is constructed. The principle for determining the fusion weights is: the higher the heterogeneity of the data source compared to other data sources, the lower its weight, thereby reducing the interference of abnormal or noisy data sources on the fusion results.

[0100] The adaptive weighted fusion function is constructed as follows:

[0101]

[0102] in, This represents the total number of data sources (in this embodiment). ), and Indicates the index of the data source. Represents a unified risk representation vector. Indicates the first Feature vectors of the data source Indicates the first Adaptive fusion weights of the data source Indicates the first The average heterogeneity index of the data source with all other data sources. This is the temperature coefficient, and its value range is... .when When the weights are large, the distribution of fusion weights is steeper, and the weights of data sources with higher heterogeneity tend to approach zero; when When the weights are relatively small, the weights of each data source tend to be evenly distributed.

[0103] S2: Hierarchical localization based on risk anchor point comparison

[0104] Reference Figure 2 This step aims to compare the fused risk representation vector with the pre-constructed risk anchor points to determine the subject's current risk level.

[0105] (I) Construction of risk anchor points

[0106] Based on historical queue data, a density peak clustering algorithm is used to identify K cluster centers, each corresponding to a risk anchor point. In this embodiment, the risk anchor point... Represented as:

[0107]

[0108] in, Index representing risk anchor points, Represents the anchor center vector. Represents the covariance matrix. This represents the effective radius of influence of the risk anchor point. The effective radius of influence is determined based on the chi-square distribution critical value of the Mahalanobis distance. ,in The standard deviation is... Describing the degrees of freedom as Confidence level is The critical value of the chi-square distribution. In this embodiment, we take... .

[0109] (ii) Multidimensional distance metric

[0110] Risk representation vector of subjects When comparing with each risk anchor point, a weighted combination of Mahalanobis distance and local density factor is used as a multidimensional distance metric:

[0111]

[0112] in, Indicates the first The local data density of the region where each risk anchor point is located is calculated using the kernel density estimation method. Represents the maximum local density among all anchor points; local density factor ( The purpose of this is to appropriately reduce the distance metric for risk anchors located in high-density areas, making it easier for subjects to be classified into more representative anchor categories.

[0113] (III) Determination of Risk Level

[0114] The current risk level is determined according to the following formula:

[0115]

[0116] That is, the nearest risk anchor point that meets the radius constraint is selected as the subject's classification. If If the radius constraint of any anchor point is not met, it is determined to be a marginal risk state. At this time, an early warning for adding a new risk anchor point is triggered, indicating that the model needs to update the anchor point library to cover the new risk subtype.

[0117] S3: Extraction and Evolution Trajectory Prediction of Risk Drivers

[0118] This step aims to identify the core risk factors driving the progression of non-alcoholic fatty liver disease and to construct a nonlinear evolution model to predict the probability of risk state transitions.

[0119] (I) Extraction of core risk drivers

[0120] Based on the aforementioned risk representation vector, the following three types of core risk drivers are extracted through differential operations:

[0121] Cumulative metabolic load The cumulative effect characterizing a subject's historical metabolic burden is defined as:

[0122]

[0123] in, Indicates time, Represents the integral variable. This represents a time series of the insulin resistance index. Indicates blood sugar level, Indicates triglyceride levels, The attenuation factor has a value range of 1. The introduction of the decay factor reflects the "recency effect": the contribution of recent metabolic load to current risk is greater than that of long-term load.

[0124] Inflammatory stress level : Characterizes the level of activity in the liver's inflammatory response, defined as:

[0125]

[0126] in, and The ratio of liver enzyme levels (De Ritis ratio) is an important marker of liver damage. Indicates C-reactive protein level, and It is an adjustable parameter. , This definition uses the Sigmoid function form, which means that the inflammatory stress level is triggered when CRP exceeds a threshold. It then rose rapidly.

[0127] rate of change in insulin resistance : Characterizing the changing trend of insulin resistance, defined as:

[0128]

[0129] in, This represents the time series of the insulin resistance index as assessed by the steady-state model. This represents the first derivative with respect to time. A positive rate indicates that insulin resistance is worsening, while a negative rate indicates that it is improving.

[0130] (ii) Nonlinear evolution model

[0131] The risk status of the subjects was modeled as the following system of differential equations:

[0132]

[0133] in, express The risk state value at any given time, with a range of values ​​being: , This represents a risk accumulation auxiliary variable used to smooth out risk-driven short-term fluctuations. The mean is variance is Gaussian white noise is used to characterize unobservable random factors. , , , For model parameters, .

[0134] The design of this differential equation system embodies the following medical logic:

[0135] Cumulative metabolic load Positive driving force for the increase in risk status (coefficient) );

[0136] Inflammatory stress level Current risk status The product of these factors negatively influences risk changes, reflecting the nonlinear interaction between inflammation and disease progression (coefficient). );

[0137] rate of change in insulin resistance Driver risk accumulation auxiliary variable Changes (coefficients) ),and Its own speed attenuation.

[0138] (iii) Risk state transition probability

[0139] The risk state transition probability is obtained by applying the above differential equation system within a time window. Numerical integration yields:

[0140]

[0141] in, The cumulative distribution function represents the standard normal distribution. , Representing state The corresponding mean and standard deviation.

[0142] In this embodiment, the risk status is divided into three levels:

[0143] Low-risk steady state: corresponding This indicates that the liver is in good health and the risk of disease progression is low.

[0144] Medium-risk evolution state: corresponding This indicates that metabolic abnormalities have occurred, and the disease may be in a reversible stage.

[0145] High-risk warning status: corresponding This indicates that significant pathological changes have occurred in the liver, requiring clinical intervention.

[0146] State transition probability Indicates starting from the current state Over time After transitioning to state The probability of . In this embodiment, take . The duration is 1 month, consistent with the clinical follow-up period.

[0147] (iv) Model parameter optimization

[0148] The above model parameters , , , , , The following objective function was obtained through training:

[0149]

[0150] in, Represents the set of model parameters , Let represent the conditional probability of the true state of the nth training sample relative to the predicted state. This represents the L1 regularization term, used to improve model sparsity. This represents the square of the L2 regularization term, used to prevent overfitting. , The regularization coefficient is . .

[0151] S4: Adaptive calibration and report generation of prediction results

[0152] This step aims to dynamically adjust model parameters and fusion weights based on historical prediction error feedback, and generate an interpretable risk analysis report.

[0153] (a) Definition of prediction error function

[0154] Define the prediction error function as follows:

[0155]

[0156] in, Indicates the length of the time window. This represents the predicted risk representation vector. This represents the risk representation vector based on actual observations. This represents the square of the Euclidean distance. Represents the regularization term, , Indicates the index of the model parameter. The regularization coefficient is . .

[0157] (ii) Dynamic updates of fusion weights

[0158] The dynamic update of the fusion weights uses gradient descent with momentum:

[0159]

[0160] in, Indicates the index of the data source. Index representing the number of iterations and , Indicates the learning rate. , Indicates the momentum coefficient. , It indicates the direction of the last weight update; the introduction of the momentum term can accelerate convergence and reduce oscillations.

[0161] (iii) Dynamic updating of model parameters

[0162] The model parameters are dynamically updated using the Adaptive Moments Estimation (Adam) algorithm, which adaptively adjusts the learning rate for each parameter. The specific update rules are as follows:

[0163] First, calculate the first-order moment estimate and the second-order moment estimate of the gradient:

[0164]

[0165]

[0166] Then, deviation correction is performed:

[0167]

[0168] Last updated parameters:

[0169]

[0170] in, Represents the set of model parameters. , The attenuation rate is estimated by moments; in this embodiment, it is taken as... , , This represents the adaptive learning rate, which is taken in this embodiment. , It is the numerical stability constant. .

[0171] (iv) Ranking of the contribution of key risk drivers

[0172] To provide clinicians with interpretable decision support information, this embodiment calculates and ranks the contribution of each risk driver to the prediction results.

[0173] For any output result (The output results) (For risk level or transition probability), the first Contribution of each risk driver Defined as:

[0174]

[0175] in, Index representing risk drivers, , Indicates the first Each input feature Indicates the output for the first... The expected gradient of each input feature is approximated using the Monte Carlo sampling method. This represents the standard deviation of the feature. The definition means that the average sensitivity of the feature to the output multiplied by the natural variation of the feature; the larger the product, the greater the contribution of the feature to the prediction result.

[0176] The final output list of contributions, sorted by score, is as follows:

[0177]

[0178] in, This represents the total number of risk drivers.

[0179] (v) Generation of risk analysis report

[0180] Based on the calculation results of the above steps, a risk analysis report for the subject is generated. The report should include at least the following:

[0181] Current risk level: low-risk steady state, medium-risk evolving state, or high-risk early warning state;

[0182] Future Risk Evolution Trajectory: Predictions Based on Differential Equation Systems Change curve;

[0183] Ranking of key risk drivers by contribution: A list of risk factors arranged from highest to lowest contribution helps physicians identify key intervention targets;

[0184] Among them, the predicted risk representation vector The risk state value is obtained by forward integration of the nonlinear evolution model. The inverse transform is performed to obtain the following: , Let be the mapping function from the risk state space to the risk representation vector space, which is established by a regression model on the training set.

[0185] S5: Dynamic Update of Risk Anchor Points

[0186] In practical clinical applications, as subject data accumulates, existing risk anchors may no longer cover new risk subtypes. Therefore, this embodiment provides a dynamic updating mechanism for risk anchors.

[0187] When the amount of new subject data reaches a preset threshold When this occurs, the risk anchor point is recalculated:

[0188]

[0189] in, Index representing risk anchor points, Indicates the original number One risk anchor point, Indicates the original number The number of data points covered by each anchor point This represents the set of data points that have newly fallen within the influence range of this anchor point. This indicates the number of new data points.

[0190] Simultaneously, it detects whether new data-dense regions exist. If a new dense region exists, and the minimum distance between this region and all existing anchor points is greater than a preset threshold DthrDthr, then a new risk anchor point is created.

[0191]

[0192] in, This represents the center vector of the newly detected dense region. This indicates the local density of the dense area. and For the preset threshold, , In this embodiment, take ( (The average effective radius of existing anchor points). Take the median of the local density of all anchor points in the historical data.

[0193] Example 2: This example provides a risk prediction model for non-alcoholic fatty liver disease, which is constructed using the method described in Example 1. This model can be stored in a computer-readable storage medium or integrated into a medical information system as a software module.

[0194] Specifically, the model includes the following functional modules:

[0195] Data input interface: Receives multi-source medical data from the subjects;

[0196] Feature extraction engine: Extracts dynamic change features and static baseline features from the data;

[0197] Heterogeneity quantizer: Calculates the heterogeneity index between different data sources;

[0198] Adaptive fusion engine: performs weighted fusion based on heterogeneity index to generate risk representation vector;

[0199] Risk Anchor Matcher: Compares risk representation vectors with pre-constructed risk anchors;

[0200] Evolution predictor: Predicts the probability of risky state transitions based on differential equation systems;

[0201] Adaptive calibrator: dynamically adjusts model parameters based on prediction error feedback;

[0202] Report generator: Outputs interpretable risk analysis reports.

[0203] Example 3: This example provides a risk prediction device for non-alcoholic fatty liver disease, which includes the following modules:

[0204] Data Acquisition Module: Configured to acquire multi-source medical data from subjects. This module interfaces with the hospital's electronic medical record system, laboratory information system, and image archiving system to achieve automatic data collection.

[0205] Heterogeneity quantification module: Configured to calculate the heterogeneity index between different data sources and perform adaptive weighted fusion. The implementation of this module is described in detail in step S1 of Example 1.

[0206] Risk Anchor Comparison Module: This module is configured to compare the subject's risk representation vector with pre-constructed risk anchors to determine the current risk level. This module is also configured to dynamically update the risk anchors according to the rules described in step S5 of Example 1.

[0207] Evolution Prediction Module: Configured to predict the probability of future risk state transitions using a nonlinear evolution model driven by risk drivers. The implementation of this module is described in detail in step S3 of Example 1.

[0208] Adaptive calibration module: Configured to dynamically adjust fusion weights and model parameters based on historical prediction error feedback. The implementation of this module is described in detail in step S4 of Example 1.

[0209] Output module: Configured to generate and output risk analysis reports for subjects. The output formats of this module include, but are not limited to: structured reports in electronic medical record systems, visual charts on mobile devices, and data interfaces for use by clinical decision support systems.

[0210] Example 4: This example illustrates the beneficial effects of the above-described method, model, and apparatus.

[0211] First, by introducing a heterogeneity index and adaptive weighted fusion, this invention can effectively handle the distributional differences and noise interference among multi-source medical data. Compared with traditional equal-weighted fusion or single-modality-based analysis methods, this invention can automatically reduce the weight of abnormal or low-quality data sources, improving the robustness and accuracy of the risk representation vector.

[0212] Second, by constructing a risk anchor reference framework, this invention achieves interpretable risk hierarchy positioning. Each risk anchor corresponds to a specific risk subtype, making the risk assessment result not only a numerical score but also a clinically meaningful category label. Clinicians can quickly understand the risk characteristics and potential pathological mechanisms of a subject based on the risk anchor to which they belong.

[0213] Third, by constructing a nonlinear evolution model based on a system of differential equations, this invention can characterize the dynamic features of the progression of non-alcoholic fatty liver disease. The cumulative metabolic load, inflammatory stress level, and rate of insulin resistance evolution introduced into the model correspond to the core pathological links in the occurrence, development, and deterioration of the disease, respectively, and have clear biological significance.

[0214] Fourth, through the adaptive calibration mechanism, this invention can dynamically adjust the model parameters and fusion weights according to the actual prediction error, so that the model can adapt to the differences in data distribution among different populations and in different medical environments, and has good generalization ability.

[0215] Fifth, by calculating the contribution ranking of key risk drivers, this invention provides interpretable decision support information. Doctors can intuitively understand which risk factors have the greatest impact on current prediction results, thereby developing personalized intervention plans.

[0216] The embodiments of the present invention are given for the purposes of illustration and description. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any changes, modifications, substitutions and variations made by those skilled in the art to the above embodiments within the scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a risk prediction model for non-alcoholic fatty liver disease, characterized in that, Includes the following steps: S1. Obtain multi-source medical data of the subjects, wherein the multi-source medical data includes at least two of the following: metabolic index sequences, lifestyle time-series data, and biomarker data; for each data source, extract its dynamic change characteristics and static baseline characteristics respectively. Calculate the heterogeneity index among different data sources, and construct an adaptive weighted fusion function based on the heterogeneity index to generate a unified risk representation vector; S2. Construct a reference framework containing multiple risk anchors, each risk anchor corresponding to a preset non-alcoholic fatty liver risk subtype; The risk representation vector is compared with each risk anchor point to calculate the multidimensional distance metric, and the subject's current risk level is determined according to the minimum distance principle. S3. Based on the risk characterization vector, extract the core risk drivers through differential operations. The core risk drivers include the cumulative amount of metabolic load, the level of inflammatory stress, and the rate of change of insulin resistance. Construct a nonlinear evolution model driven by the risk drivers to predict the probability of risk state transition of the subject within a preset time window. S4. Based on historical prediction error feedback, dynamically adjust the fusion weights of the adaptive weighted fusion function and the model parameters of the nonlinear evolution model; generate a risk analysis report for the subjects based on the prediction results, the risk analysis report including the current risk level, future risk evolution trajectory, and the contribution ranking of key risk driving factors.

2. The method according to claim 1, characterized in that, The heterogeneity index The calculation method is as follows: in, and Indicates different data source indexes, and These represent data sources respectively. and data source Extracted feature vectors Describing covariance, Represents variance. This represents the KL divergence between the feature distributions of two data sources. and Representing data sources respectively and data source The empirical probability distribution of the feature vectors, This is the scaling factor, and its value range is... ; The adaptive weighted fusion function is constructed as follows: in, Indicates the total number of data sources. and Indicates the index of the data source. Represents a unified risk representation vector. Indicates the first Feature vectors of the data source Indicates the first Adaptive fusion weights of the data source Indicates the first The average heterogeneity index of the data source with all other data sources. This is the temperature coefficient, and its value range is... .

3. The method according to claim 1, characterized in that, The risk anchor point ,in Index representing risk anchor points, Represents the anchor center vector. Represents the covariance matrix. Indicates the effective radius of influence of the risk anchor point; The multidimensional distance metric is a weighted combination of Mahalanobis distance and local density factor: in, Indicates the first The local data density of the region where each risk anchor point is located is calculated using the kernel density estimation method. This represents the maximum local density among all anchor points; The current risk level is determined according to the following formula: like If the radius constraint of any anchor point is not met, it is determined to be an edge risk state.

4. The method according to claim 1, characterized in that, The core risk drivers are extracted using the following method: The cumulative metabolic load Defined as: in, Indicates time, Represents the integral variable. The time series of the insulin resistance index, Indicates blood sugar level, Indicates triglyceride levels, The attenuation factor has a value range of 1. ; The level of inflammatory stress Defined as: in, and Indicates liver enzyme levels, Indicates C-reactive protein level, and It is an adjustable parameter. , ; The rate of change of insulin resistance Defined as: in, This represents the time series of the insulin resistance index as assessed by the steady-state model. It represents the first derivative with respect to time.

5. The method according to claim 4, characterized in that, The specific nonlinear evolution model is as follows: in, express The risk state value at any given time, with a range of values ​​being: , This represents a cumulative risk auxiliary variable. The mean is variance is Gaussian white noise, , , , For model parameters, ; The risk state transition probability is obtained by applying the above differential equation system within a time window. Numerical integration yields: in, The cumulative distribution function represents the standard normal distribution. , Representing state The corresponding mean and standard deviation, This represents the prediction time window, with a value range of [value missing]. The state , Selected from different states in low-risk steady state, medium-risk evolution state, and high-risk early warning state; Wherein, the low-risk steady state corresponds to The intermediate risk evolution state corresponds to The high-risk warning state corresponds to .

6. The method according to claim 1, characterized in that, The adaptive calibration defines a prediction error function. : in, Indicates the length of the time window. This represents the predicted risk representation vector. This represents the risk representation vector based on actual observations. This represents the square of the Euclidean distance. Represents the regularization term. , Indicates the index of the model parameter. The regularization coefficient is . , Represents the set of fusion weights. , Represents the set of model parameters; The dynamic update rule for the fusion weights is as follows: in, Indicates the index of the data source. Index representing the number of iterations and , Indicates the learning rate. , Indicates the momentum coefficient. , This indicates the direction of the last weight update.

7. The method according to claim 1, characterized in that, The contribution ranking of the key risk drivers is calculated as follows: For any output result The output result For risk level or transition probability, the first Contribution of each risk driver Defined as: in, Index representing risk drivers, , Indicates the first Each input feature Indicates the output for the first... The expected gradient of each input feature is approximated using the Monte Carlo sampling method. This represents the standard deviation of the characteristic; The final output list of contributions, sorted by score, is as follows: in, This represents the total number of risk drivers, and the contribution ranking is used to provide interpretable decision support information to clinicians.

8. A risk prediction model for non-alcoholic fatty liver disease, characterized in that, It is constructed using the method described in any one of claims 1 to 7.

9. A risk prediction device for non-alcoholic fatty liver disease, characterized in that, include: The data acquisition module is configured to acquire multi-source medical data from the subjects. The heterogeneity quantization module is configured to calculate the heterogeneity index between different data sources and perform adaptive weighted fusion. The risk anchor comparison module is configured to compare the subject's risk representation vector with pre-constructed risk anchors to determine the current risk level; The evolution prediction module is configured to predict the probability of future risk state transitions based on a nonlinear evolution model driven by risk drivers. The adaptive calibration module is configured to dynamically adjust the fusion weights and model parameters based on historical prediction error feedback; The output module is configured to generate and output risk analysis reports for the subjects.

10. The risk prediction device for non-alcoholic fatty liver disease according to claim 9, characterized in that, The pre-built risk anchors in the risk anchor comparison module are dynamically updated according to the following rules: When the amount of new subject data reaches a preset threshold When this occurs, the risk anchor point is recalculated: in, Index representing risk anchor points, Indicates the original number One risk anchor point, Indicates the original number The number of data points covered by each anchor point This represents the set of data points that have newly fallen within the influence range of this anchor point. Indicates the number of new data points; Simultaneously, detect whether there are new data-dense areas; if so, and the minimum distance from all existing anchor points is greater than a preset threshold. Then create a new risk anchor point: in, This represents the center vector of the newly detected dense region. This indicates the local density of the dense area. and For the preset threshold, , .

Images