Coronary artery calcification early warning system for type 2 diabetes patients

Abstract

The invention discloses a coronary artery calcification early warning system for type 2 diabetes patients, and relates to the technical field of medical detection. A data acquisition module is used for acquiring continuous physiological parameter data of a user; the risk modeling module is combined with coronary artery calcification evolution characteristics in historical clinical samples to construct a multi-parameter dynamic association model; an index weight calculation unit generates a risk influence factor vector based on a sensitivity analysis result of the physiological indexes on risk prediction; the machine learning analysis module performs iterative training on the prediction model by adopting an integrated learning algorithm, and performs prediction updating by utilizing a risk influence factor vector; the early warning trigger module dynamically generates a graded early warning signal according to the grading trend and a set threshold value; the weak item positioning module carries out contribution degree analysis and anomaly recognition on the key risk indexes and automatically generates personalized intervention suggestions; according to the invention, early recognition and dynamic early warning of coronary artery calcification progress can be realized, and the method is suitable for intelligent early warning management scenes of chronic disease cardiovascular risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical detection, in particular to an early warning system for coronary calcification of type 2 diabetes patients. BACKGROUND

[0002] Coronary artery calcification (CAC) is one of the early indicators of the occurrence and development of coronary heart disease, and its degree is closely related to the risk of cardiovascular events. Especially in type 2 diabetes patients, due to their metabolic abnormalities and chronic low-grade inflammatory state, the incidence of coronary calcification is significantly higher than that of the general population, and the progression is faster. Clinical studies have shown that even in the asymptomatic stage, coronary calcification may silently form and develop in diabetes patients, increasing the risk of sudden cardiovascular events.

[0003] In the prior art, the degree of coronary calcification is mainly quantitatively evaluated by coronary CT (such as CAC score), but this method relies on high-cost equipment, has radiation risk, is not suitable for frequent detection, and cannot realize continuous dynamic monitoring, making it difficult to identify high-risk patients in a timely manner.

[0004] In addition, the existing early warning system is based on static data for single-point risk assessment, ignoring the dynamic evolution process of coronary calcification risk of type 2 diabetes patients under different physiological states (such as blood glucose fluctuations and changes in inflammatory levels), resulting in delayed early warning and low precision. SUMMARY

[0005] The purpose of the present application is to provide an early warning system for coronary calcification of type 2 diabetes patients to solve the problems in the background art.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: an early warning system for coronary calcification of type 2 diabetes patients, comprising:

[0007] A data acquisition module for acquiring continuous physiological parameter data of a target user;

[0008] A risk modeling module for constructing a multi-parameter dynamic correlation model based on the acquired continuous physiological parameter data and the coronary calcification evolution characteristics of type 2 diabetes patients in the historical clinical sample library, for reflecting the time evolution trend of coronary calcification;

[0009] An index weight calculation unit for assigning each parameter with a corresponding weight based on the statistical analysis results of the sensitivity of different indicators to coronary calcification, and outputting a risk influence factor vector;

[0010] A machine learning analysis module for iteratively training the multi-parameter dynamic correlation model based on an ensemble learning algorithm, and predicting and updating high-risk events using the risk influence factor vector.

[0011] an early warning trigger module configured to generate an early warning signal and push the early warning signal to a user terminal or a doctor terminal when a coronary calcification trend in the prediction result breaks through a preset risk threshold;

[0012] a weak point positioning module configured to determine an index parameter group that contributes most to the risk in the prediction result, thereby guiding an intervention direction.

[0013] Preferably, the continuous physiological parameter data includes blood glucose fluctuation curve, heart rate variability, blood pressure, inflammatory marker level, glycosylated hemoglobin, and past CAC score.

[0014] Preferably, the risk modeling module includes:

[0015] The continuous physiological parameter data obtained by the data acquisition module is preprocessed;

[0016] A plurality of physiological parameter evolution sequences of known type 2 diabetes patients and their corresponding coronary calcification progression trajectories are extracted from a historical clinical sample library, the samples are clustered by using time series similarity measurement, and a typical calcification evolution mode is extracted;

[0017] The preprocessed continuous physiological parameter data and the calcification evolution mode are combined, and a long short-term memory network is used to build a time series prediction model to dynamically fit the coronary calcification score trend in the future period.

[0018] Preferably, the index weight calculation unit includes:

[0019] Based on the corresponding relationship between the known coronary calcification score and the continuous physiological parameter data in the historical clinical sample, the sensitivity score of each index to the coronary calcification prediction output at different score stages is calculated and normalized; and the normalized index weights are combined in the original parameter arrangement order to form a multi-dimensional risk influence factor vector.

[0020] Preferably, the machine learning analysis module includes:

[0021] The preprocessed continuous physiological parameter data and the historical coronary calcification score sample are divided, and a plurality of basic learners are trained respectively to form an initial model pool;

[0022] Based on the output results of the basic learners, a strong learner model is constructed by using a weighted voting mechanism to nonlinearly fuse the prediction output results of the plurality of learners, and the optimal fusion weight is determined through cross-validation.

[0023] Preferably, the risk influence factor vector is used to predict and update the high-risk event, including:

[0024] The continuous physiological parameters are item-by-item weighted with the risk influence factor vector to construct a risk weighted input matrix, wherein each physiological parameter value is multiplied by its corresponding weight to serve as a model input feature;

[0025] The risk weighted input matrix is input into the trained time series prediction model to output an updated coronary calcification score prediction value through model forward propagation, and a risk evolution sequence in the time window is generated in combination with a score change trend and a historical trend chart;

[0026] According to the current prediction value and its change speed, a set risk judgment threshold value is compared, and when the prediction score reaches or exceeds the high risk threshold value, the potential high risk event is marked;

[0027] For the potential high risk event, the input parameters and the prediction results in this state are fed back to the multi-parameter dynamic correlation model as training samples to update the multi-parameter dynamic correlation model parameters.

[0028] Preferably, the early warning trigger module comprises:

[0029] The coronary calcification score sequence in the future time window output by the prediction model is obtained, the absolute value and the time derivative of the score are calculated, the growth rate, trend shape and fluctuation amplitude of the score are evaluated, and a risk trend score index is constructed accordingly;

[0030] The risk trend score index is compared with a plurality of pre-set early warning trigger conditions, the early warning trigger conditions including: the prediction score exceeding the static high risk threshold value, the score growth rate exceeding the dynamic rate threshold value per unit time, and the score curve showing short-term acceleration steep increase and other nonlinear evolution characteristics;

[0031] Based on the risk intensity, evolution trend type and historical personal risk curve in the comparison result, a hierarchical early warning signal is automatically generated and divided into high risk, medium risk or attention level.

[0032] Preferably, the weak item positioning module comprises: sensitivity analysis of the weighted physiological parameters input into the time series prediction model, calculation of the risk contribution degree of each physiological parameter in the current prediction output to generate a feature contribution list; according to the weight order in the feature contribution list, the first k key parameters with the greatest impact on the calcification score prediction value in the current prediction time window are screened out to form a high influence factor set, and the key parameters include blood glucose variation, heart rate variability and CRP level.

[0033] In the above technical solution, the present application provides technical effects and advantages:

[0034] 1. The coronary calcification early warning system provided by the application aims at the clinical problems that type 2 diabetes patients are prone to have occult arterial calcification, the progress is fast, and the problems are difficult to identify early. An intelligent early warning framework is constructed based on multi-dimensional continuous physiological parameters, combining machine learning, dynamic modeling and risk interpretation mechanism. The system can realize dynamic fitting of coronary calcification score trend, abnormal trend identification and multi-grade risk early warning by integrating time series prediction model, risk influence factor weighting mechanism and multi-source clinical data learning method, and significantly improves the prospective identification ability of potential cardiovascular events.

[0035] 2. The application introduces a weak item positioning module. The physiological indicators with the largest risk contribution in the current prediction result are identified by a model explainability analysis algorithm, individual analysis is performed in combination with historical baseline deviation, and targeted intervention suggestions are automatically generated, thereby establishing an intelligent closed-loop process integrating prediction, explanation and intervention. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0037] Figure 1 The system module flowchart of the present application. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0039] Embodiment, please refer to Figure 1 The coronary calcification early warning system for type 2 diabetes patients described in the present embodiment comprises:

[0040] A data acquisition module is used to acquire continuous physiological parameter data of a target user.

[0041] A risk modeling module is used to construct a multi-parameter dynamic correlation model according to the acquired continuous physiological parameter data, in combination with the coronary calcification evolution characteristics of type 2 diabetes patients in the historical clinical sample library, for reflecting the time evolution trend of coronary calcification.

[0042] An index weight calculation unit is configured to assign each parameter with a corresponding weight based on statistical analysis results of sensitivity of different indexes to coronary calcification, and output a risk influence factor vector.

[0043] A machine learning analysis module is configured to iteratively train the multi-parameter dynamic correlation model based on an ensemble learning algorithm, and predict and update high-risk events by using the risk influence factor vector.

[0044] An early warning triggering module is configured to generate an early warning signal and push it to a user terminal or a doctor terminal when a coronary calcification trend in the prediction result breaks through a preset risk threshold.

[0045] A weak point positioning module is configured to determine a parameter group of indexes that contributes most to the risk in the prediction result, thereby guiding the intervention direction.

[0046] The data acquisition module described in the application constitutes a basic data input unit of the system, and is designed to continuously, dynamically and structurally acquire key physiological indexes of a target user.

[0047] Blood glucose fluctuation curve data: the module records the blood glucose level change of the user in real time by integrating a continuous glucose monitoring device (CGM) interface, and the acquisition frequency can reach once every 5 minutes, thereby generating a complete time-blood glucose sequence. The acquisition module not only records static blood glucose values, but also calculates and outputs blood glucose fluctuation amplitude, intraday coefficient of variation (i.e. the ratio of standard deviation to average value in a certain time period), and high and low blood glucose exposure time, thereby providing support for metabolic influence analysis of coronary calcification.

[0048] Heart rate variability parameters (HRV): by using a wearable smart electrocardiogram device, the module acquires RR interval data of the user in a resting state and a mild activity state, and then calculates standard indicators of heart rate variability, such as standard deviation (SDNN), average RR interval, low / high frequency ratio, etc. Research shows that changes in sympathetic / parasympathetic nerve tension are closely related to the degree of cardiovascular sclerosis, and therefore such parameters can be used as one of the indirect physiological signals for inferring the progress of coronary calcification.

[0049] Dynamic blood pressure parameters: by interfacing with a 24-hour dynamic blood pressure monitoring device, the module records systolic pressure, diastolic pressure and pulse pressure, and identifies and labels the morning blood pressure rise amplitude and night non-dipper blood pressure state. The data in this part is used to reflect the risk of decreased vascular wall compliance related to calcification in type 2 diabetes patients.

[0050] Inflammatory biomarker levels: Concentrations of chronic inflammatory biomarkers, including high-sensitivity C-reactive protein (hs-CRP), interleukin-6 (IL-6), and tumor necrosis factor-α (TNF-α), were collected periodically by connecting to medical testing databases or home sampling devices. Chronic inflammation is considered an important inducing mechanism for atherosclerosis and its calcification process. The data acquisition module has a dynamic trend analysis function for these indicators to identify signals that exacerbate the risk of inflammation-related coronary artery calcification.

[0051] Glycated Hemoglobin (HbA1c) Data: This module periodically acquires the target user's HbA1c test values ​​via medical device interface or user upload. HbA1c is a key indicator reflecting average blood glucose control over the past 8 to 12 weeks, and its elevation is significantly correlated with the incidence of cardiovascular disease. This module collects this parameter and performs cross-validation with continuous blood glucose data to improve the accuracy of glucose metabolism abnormality risk assessment.

[0052] Previous coronary artery calcium scoring data (CAC score): The data acquisition module supports extracting the target user's coronary artery calcium scoring data, such as the Agatston score and calcification volume score, from previous physical examination reports, electronic medical records, or DICOM imaging systems, and using it as a label value or initial risk baseline input in the prediction model. For users without any initial CT records, the system defaults to an initial score of zero, but will dynamically update it through model predictions.

[0053] To ensure structured management and consistent processing of the aforementioned multi-source data, the data acquisition module employs a unified timestamp synchronization mechanism. This involves attaching standardized UTC timestamps to all data during the acquisition process and constructing a corresponding time-series database based on the acquisition frequency of physiological parameters. Furthermore, the module includes data integrity detection and outlier removal mechanisms. For example, it automatically identifies and marks entries with abnormal acquisition time intervals or values ​​exceeding physiological ranges to prevent interference with subsequent modeling results.

[0054] The multidimensional physiological indicator set generated by the data acquisition module, after preliminary preprocessing (such as normalization, difference imputation, outlier smoothing, etc.), will be passed as input variables to the risk modeling module. The collected data will eventually form a set of multidimensional vectors, with each time point containing multiple indicators such as blood glucose, heart rate variability, blood pressure, and inflammatory status, forming a sample vector sequence for training or inference.

[0055] The risk modeling module described in this invention is the core technical unit of the system used to establish a predictive model for the evolution trend of coronary artery calcification. The goal of this module is to construct a time-series prediction model with dynamic learning capabilities by deeply modeling continuous physiological parameter data of the target user and combining it with the actual progression trajectory of coronary artery calcification in type 2 diabetic patients in historical clinical samples. This enables trend prediction of coronary artery calcification scores over a future period, thus assisting in early warning functions.

[0056] Before proceeding to the modeling stage, the continuous physiological parameter data acquired by the data acquisition module must first undergo comprehensive preprocessing to improve the model training quality and the structural regularity of the input data. This step includes, but is not limited to:

[0057] Time alignment and completion: Since the sampling frequencies of different parameters may vary (e.g., blood glucose every 5 minutes, heart rate variability every 1 minute), the system first synchronizes and aligns the data based on a unified time axis. For data with missing time periods, forward padding, interpolation algorithms, or moving averages are used to complete the data.

[0058] Outlier identification and removal: By setting a threshold within a reasonable physiological range, data entries that show mutations or exceed the normal physiological range are marked and removed to prevent extreme values ​​from misleading the model.

[0059] Normalization: To avoid the impact of different physiological indicators on the model training results due to different units, all input parameters are standardized. Z-Score normalization is often used, which involves subtracting the mean from the data of each indicator and dividing by its standard deviation, so that the distribution mean is 0 and the standard deviation is 1.

[0060] Feature derivation: Based on the original physiological data, the system can also generate a set of derived features, such as the blood glucose variation coefficient, the maximum / minimum blood pressure fluctuation range within 24 hours, and the low-frequency / high-frequency ratio of heart rate variability, to improve the model's ability to express dynamic trends.

[0061] After the above steps, a set of high-quality, structured multidimensional physiological parameter time series data is generated, which serves as the basis for model input.

[0062] To enhance the model's ability to identify trends in coronary artery calcification, the risk modeling module further incorporates evolutionary information from a historical clinical sample database to construct a representative reference trajectory template.

[0063] The system accesses long-term follow-up data of type 2 diabetes patients provided by medical research institutions. The samples contain continuous physiological parameter sequences over multiple time periods and corresponding coronary artery calcification scores (such as the Agatston score). Each set of sample records constitutes a "physiological parameter-score trajectory" pair.

[0064] To identify groups of samples with similar developmental trends, the system employs Dynamic Time Warping (DTW) to calculate the similarity of time-series trajectories among different patients. DTW can pair and compare physiological evolution processes of different lengths and rhythms, allowing samples with similar trends but different time alignments to be grouped together.

[0065] Based on the DTW similarity matrix, the system uses hierarchical clustering or density clustering algorithms to group the samples, obtaining several representative coronary artery calcification evolution patterns. Each pattern represents a typical calcification development path, such as "steady-state," "slowly rising," or "rapidly progressing." Corresponding physiological parameter change feature sequences and scoring evolution curves are extracted for each pattern and used as prior knowledge for modeling.

[0066] This step enables the model to have "analogical cognition" ability during the prediction process, that is, to determine which typical evolutionary trajectory the target user's current state is closer to, thereby improving the accuracy and interpretability of the prediction.

[0067] After obtaining the preprocessed physiological parameter sequences of the target users and typical evolutionary templates, the risk modeling module enters the deep modeling stage, using a Long Short-Term Memory (LSTM) network to predict the time trend of coronary artery calcification scores.

[0068] LSTM is an improved recurrent neural network architecture with the ability to remember long-term temporal dependencies, making it particularly suitable for processing continuous physiological signals. The model's input layer receives a sequence of multidimensional physiological parameter vectors, and the hidden layers contain multiple LSTM units, each with forget gates, input gates, and output gates to control information retention, updating, and output. The output layer is a regression structure that outputs predicted coronary artery calcification scores for a future time window.

[0069] The system utilizes input and rating sequence pairs from a historical sample database for supervised training. The objective function is the mean squared error between the predicted and actual ratings. To prevent overfitting, L2 regularization and Dropout mechanisms are introduced. After training, the model will be used for rating prediction inference for the target user.

[0070] After inputting recent continuous physiological parameter data of the target user, the model outputs predicted coronary artery calcium scores for several future time points based on historically learned evolutionary patterns and parameter correlations, generating a score trend curve. The system further uses a trend analysis algorithm to determine the rate of score change; if it exceeds a risk threshold, an early warning is triggered.

[0071] The indicator weight calculation unit described in this invention is one of the key components for achieving refined modeling of early coronary artery calcification risk. Its main function is to quantify the sensitivity of each physiological indicator to the predicted calcification score based on the known mapping relationship between coronary artery calcification scores and continuous physiological parameter data from historical clinical sample data. Based on this, it constructs a set of normalized indicator weights to form a complete risk impact factor vector, which is then used as input variables to the risk modeling module, enabling the model to selectively process multidimensional features.

[0072] In the process of calculating indicator weights, historical case data with known labels are first retrieved from the sample database. Each set of sample data should contain at least the following two types of information:

[0073] A set of time-series physiological parameter data (such as blood glucose change curves, heart rate variability sequences, blood pressure records, fluctuations in inflammatory markers, etc.) is denoted as parameter matrix X;

[0074] The coronary artery calcium score (Agatston score or equivalent score) recorded synchronously with this parameter data is denoted as the target variable Y.

[0075] To assess the predictive impact of each parameter on the calcification score, the system incorporates sensitivity analysis to quantify the contribution of each parameter to the prediction results at different scoring stages. This analysis can be implemented using one or more of the following methods:

[0076] The absolute value method of multiple linear regression coefficients estimates the linear relationship between the target scoring variable and each input feature by training a standard multiple linear regression model. The absolute values ​​of the regression coefficients of each feature output by the model can be used as a preliminary reference for sensitivity scores.

[0077] Random Forest Feature Importance Method: The random forest regression model is used to train the samples. By using the average information gain brought by a certain feature when the tree structure splits at the node, the "average contribution" of each feature is calculated, and its importance score is obtained.

[0078] SHAP (Shapley Additive exPlanations): When performing interpretive analysis on nonlinear models, the SHAP method is used to calculate a weighted average of the importance of each feature in each predicted sample, thereby obtaining a global sensitivity score, which has high interpretability.

[0079] For example, for the i-th physiological indicator, its average SHAP value calculated across all historical samples can be denoted as Si, which reflects the average influence of the indicator on the calcification scoring model.

[0080] To ensure that the sensitivity scores of all indicators can be compared in the same metric space and are applicable to subsequent model fusion, the original sensitivity scores need to be normalized so that they are distributed in a uniform interval of [0,1] and sum to 1, thus having statistical stability.

[0081] Commonly used normalization methods are as follows:

[0082] Max-Min Normalization: All sensitivity scores are processed as follows: For each index score Si, its normalized score... , where Smin and Smax are the minimum and maximum values ​​among all sensitivity scores, respectively.

[0083] Softmax normalization: For scenarios that emphasize relative weight differences, the Softmax function can be used: For each metric score Si, its normalized weight is: , where j is the index of all indicators.

[0084] After normalization, the sensitivity weights of all indicators will be limited to a uniform range, forming a stable and continuous weight distribution. This normalization operation not only improves the adaptability of data during the model input stage, but also provides a mathematical basis for the system to perform dynamic weight adjustments in the future.

[0085] After the normalized sensitivity weights of all indicators are determined, the system will recombine the normalized weights of each indicator into a structured vector, denoted as the risk impact factor vector WS, according to the order in which the indicators correspond in the original data collection sequence.

[0086] Assume the total number of physiological parameters involved in the modeling is n, which are respectively The risk impact factor vector is then expressed as: ,in These correspond to the normalized sensitivity weights for each indicator.

[0087] This risk impact factor vector, as one of the output data of the module, will be fed into the risk modeling module together with the user's real-time physiological parameter input. It will be used to dynamically fit and predict the future trend of coronary artery calcification score, forming a systematic and multi-level early warning mechanism.

[0088] The machine learning analysis module of this invention uses ensemble learning as its core concept. Through the combined training of multiple basic learners, it improves the prediction accuracy and generalization ability of the model under high-dimensional, multi-temporal physiological parameter data. Its construction process is as follows:

[0089] This module first receives input data from the data acquisition module and the indicator weight calculation unit, specifically including:

[0090] A continuous physiological parameter matrix that has been normalized, outlier-handled, and feature-derived;

[0091] Historical coronary calcification scoring data (e.g., Agatston score) were synchronized with this.

[0092] The data is divided into samples according to time slices. Each training sample contains a continuous physiological parameter sequence and corresponding score results within a time window.

[0093] To enhance model diversity and prediction stability, the system uses multiple base learners trained in parallel. Optional learners include, but are not limited to:

[0094] Gradient Boosting Regression Tree (GBRT);

[0095] Random Forest Regressor (RF);

[0096] Support Vector Regression Machine (SVR);

[0097] Multilayer perceptron (MLP);

[0098] Long Short-Term Memory Network (LSTM);

[0099] Each learner is trained independently on the sample data, generating its own set of prediction functions to form the initial model pool. Some models are constructed using feature subsets or sample subsets to enhance the diversity of model structure, thereby improving the subsequent ensemble results.

[0100] The system employs either a weighted voting mechanism or a stacking ensemble mechanism to construct strong learners. In the weighted voting mode, the predicted output of each base learner is assigned a weight based on its performance in cross-validation. Let Pi be the predicted output of each learner and Wi be its weight; then the final ensemble prediction result P is the weighted average of all Pi and Wi, i.e.: ,in .

[0101] In stacked mode, the predictions of all base learners are used as inputs to train a meta-learner (such as linear regression or XGBoost) for secondary learning, further improving the fitting ability. This mechanism effectively integrates the nonlinear response capabilities of different models to physiological indicators, improving stability and robustness on diverse sample populations.

[0102] To improve the system's sensitivity to individual disease progression and its predictive foresight, this invention further introduces risk factor vectors to participate in model input reconstruction, enhances the model's ability to identify fluctuations in key indicators, and constructs a mechanism for identifying high-risk events and providing model feedback.

[0103] Risk impact factor vector output by the indicator weight calculation unit With the user's current continuous physiological parameter sequence Perform item-by-item weighting to construct the weighted input matrix X′, i.e.:

[0104] For the input vector at each time t After weighting, it becomes This processing enhances the expressive power of high-weight indicators in the predictive model, improving the model's response accuracy to individualized risk drivers. The system supports the rolling construction of weighted matrices, which can form time-series input matrices for deep predictive models.

[0105] The constructed risk-weighted input matrix is ​​fed into a trained time-series prediction model (such as LSTM or an ensemble model), and the predicted output, i.e., the coronary artery calcification score sequence at several future time points, is obtained through forward propagation. Combined with historical score trajectories, the system forms a complete time-window risk evolution map, recording the score growth rate, volatility, and trend tendency.

[0106] The system sets two types of judgment criteria to identify potentially high-risk events:

[0107] Scoring threshold judgment: When the predicted score value Pt is greater than or equal to the system's preset high-risk threshold T1 (e.g., Agatston score > 300), the system marks it as "high-risk status";

[0108] Growth rate judgment: If the score growth rate dP / dt exceeds the threshold T2 within a continuous time window (e.g., the score growth rate per unit time exceeds 20%), the system will also mark it as "rapid progress and high risk".

[0109] If either condition is met, a high-risk trend is determined to exist within the current time window, triggering the early warning module to take action.

[0110] Once a high-risk event is identified, the system uses the event's input parameters (weighted physiological parameters) and the model's output score together as new training samples, feeding them back into the original multi-parameter dynamic association model and executing the following update strategy:

[0111] If the model is a deep neural network structure, perform incremental training and fine-tune the key weight layers;

[0112] If the model is an ensemble structure, the model structure is optimized by updating some of the basic learners or retraining the meta-learners;

[0113] The system records feedback samples and results to form an "abnormal event backtracking library," which is used for subsequent optimization of sampling strategies and correction of risk thresholds.

[0114] This feedback mechanism ensures that the model can learn and optimize continuously, and has the ability to cope with disease evolution, individual differences and dynamic changes over time.

[0115] In the early warning system for coronary artery calcification described in this invention, the warning triggering module is used to generate a risk warning signal in a timely manner based on preset risk threshold conditions and individualized trend characteristics when the prediction model output shows abnormal changes in the trend of coronary artery calcification.

[0116] First, the system obtains the coronary artery calcium score sequence within future time windows output by a time series prediction model (such as LSTM or ensemble regression model). This sequence represents the trend prediction result of the system's coronary artery calcium score for the target user at several future time points, denoted as:

[0117] Rating prediction sequence: in, Indicates the predicted time point The predictive value of coronary artery calciumization score.

[0118] To determine whether the scoring sequence exhibits an abnormal trend or a rapid upward tendency, the system further calculates the following indicators for the sequence:

[0119] SRI (Score Absolute Value): The predicted endpoint value is used to determine whether the current score has reached the set static high-risk threshold.

[0120] Rate of Rating (RRP): Calculates the rate of increase of ratings per unit time within a time window. Its expression is: RRP = [ – P(t1)] / [ – t1]; reflects the overall rate of change in coronary artery calcium score within the prediction period.

[0121] The derivative sequence of the rating is obtained by calculating the incremental difference between consecutive time points on the P-sequence. .

[0122] AI-powered acceleration score change analysis: Further analysis of the rate of change of the derivative sequence is performed to calculate the acceleration of the score growth rate. It is used to identify whether the scoring curve shows an abnormally steep rise in a short period of time.

[0123] Volatility FM: Records the maximum volatility range in the scoring sequence, i.e., the difference between the maximum and minimum values. .

[0124] The above indicators together constitute the risk trend scoring indicator set, denoted as RTS = {SRI, RRP, AI, FM}, which is used for subsequent comparison with preset trigger conditions.

[0125] The early warning triggering module has multiple pre-set sets of trigger conditions for comparing various trend scoring indicators in the aforementioned RTS. If any trigger condition is met, the system marks it as a potential risk trend and initiates the early warning signal generation process. These conditions include, but are not limited to, the following three categories:

[0126] If the endpoint predictive value of the scoring sequence is greater than or equal to the static high-risk threshold, a high-risk warning is triggered directly. The static high-risk threshold can be set according to clinical criteria (e.g., an Agatston score ≥ 300 is considered a high risk of calcification).

[0127] If the rate of increase (RRP) of the score is greater than or equal to the rate threshold, it indicates that the score is rising rapidly in the short term. The system identifies this as "rapidly progressive calcification" and triggers a warning of medium or higher level even if the current score has not yet reached the static high-risk threshold.

[0128] If an acceleration AI ≥ acceleration threshold is detected in the scoring derivative sequence and the duration exceeds the set interval Δt, the system determines it as a short-term rapid progression trend; such nonlinear curve features are of great clinical significance for identifying sudden lesions.

[0129] All comparison operations are completed automatically by the system algorithm module. If any indicator in RTS exceeds the corresponding threshold, the system marks that time period as an abnormal evolution window and records the relevant indicators and triggering conditions.

[0130] To improve the clinical interpretability and processing efficiency of early warning information, the system further generates multi-level early warning signals based on comparison results, combined with users' historical risk records and trend classifications. The system defaults to the following three levels:

[0131] High-risk level: Meets the static high-risk threshold, or is triggered by both scoring rate and acceleration, or the predicted score is on an upward trend with a large growth slope for multiple consecutive time points. The system will push a "high-risk warning," and immediate intervention or review is recommended.

[0132] Medium risk level: Only the rate threshold or acceleration threshold is met, or the score fluctuation range FM is large but does not show a continuous upward trend. The system will push a "medium risk warning" and recommend that you check or evaluate lifestyle interventions within one week.

[0133] Level requiring attention: If the score changes slowly but shows a short-term rebound trend, or if the model confidence decreases (i.e., the model uncertainty increases), the system will push a "pay attention prompt" to remind the user to continue monitoring.

[0134] In addition, the system combines an individual's historical scoring trajectory (such as the scoring trend over the past 6 months) and uses an incremental trend fitting method (such as linear / exponential fitting) to compare the current trend and assess whether the change is an abnormal deviation. If the scoring curve deviates from the historical average slope by more than a certain percentage (such as 50%), the warning level can be further upgraded.

[0135] All early warning information can be pushed to users' mobile terminals, home monitoring device interfaces, or doctors' remote monitoring platforms. The push content includes: the current risk level and its meaning; the corresponding predicted score and trend chart; the main trigger indicators and their changes; and system-suggested measures or appointment suggestions.

[0136] In this invention, the weakness localization module is used to determine the key physiological parameter that has the greatest impact on the predicted risk level based on the coronary artery calcification score prediction results output by the model. This is the "individual risk contribution factor" identified by the system, thereby achieving an accurate interpretation of potentially dangerous physiological states and automatically linking clinical intervention recommendations to support users or doctors in carrying out targeted health management.

[0137] This module first receives a risk-weighted physiological parameter dataset as input. This dataset is formed after being processed by the aforementioned "indicator weight calculation unit" and represents the user's weighted physiological characteristics within a current time window, including but not limited to time-series indicators such as blood glucose fluctuation curve, heart rate variability (HRV), and high-sensitivity C-reactive protein (CRP).

[0138] The system integrates interpretable machine learning algorithms to perform feature contribution analysis on physiological parameters input to time series prediction models (such as LSTM), that is, to assess the degree of risk impact of each parameter in the current prediction result. Implementation methods include:

[0139] SHAP value method: Based on game theory, it decomposes the model output into the marginal contribution of each input feature, thereby quantitatively measuring the impact of each indicator on the predicted score;

[0140] LIME method: Perform local perturbation analysis on the model and reflect the local sensitivity of input features to output by fitting a linear surrogate model;

[0141] Feature importance ranking algorithms: such as tree model feature importance ranking methods for regression problems, which measure the information gain of features in sample splitting.

[0142] The calculation results are output in vector form, forming a list of feature contributions for the current time window, denoted as: ;in, This represents the i-th physiological indicator. This indicates its contribution to the current score prediction. The larger the value, the more significant the impact of the indicator on the predicted risk score.

[0143] After obtaining the feature contribution list, the system sorts them from high to low contribution and selects the top k key parameters that contribute the most to the risk score within the current prediction time window to form a set of high-impact factors. Here, k is the number of parameters that the system sets or can adjust. By default, indicators that are closely related to the coronary calcification process in previous studies, such as blood glucose variability (GVI), heart rate variability (HRV), and high-sensitivity C-reactive protein (CRP) levels, are given priority.

[0144] The goal of this step is to compress the sources of risk from a high-dimensional indicator space into a small number of key factors, which will facilitate the subsequent identification of the source of the anomaly and the formulation of intervention strategies.

[0145] The system further performs individualized time-series analysis and compares each key parameter in the aforementioned set of high-impact factors with historical baselines. The operational process includes:

[0146] Historical baseline calculation: Retrieve the mean and standard deviation of this parameter for the target user over a certain period of time (e.g., 3 months or 6 months) to construct an individual historical reference interval, denoted as:

[0147] Baseline Mean Standard deviation Corresponding to the normal fluctuation range Where α is the fluctuation tolerance coefficient, which is usually taken as 1 or 1.5.

[0148] Current state comparison: The average parameter value within the current prediction period. Its historical baseline Compare them, and if there is a significant deviation (e.g.: If ᵢ, where β is the determination coefficient, then the system determines that the parameter is an "abnormal impact index".

[0149] Tag Generation: An "anomaly impact tag" is attached to each identified critical anomaly indicator, and features such as offset direction (upward or downward), offset magnitude, and duration are recorded to form a structured output record: E = { (Type, Deviation Amount, Duration)}; This record serves as an important explanatory output of the individual risk evolution process, supporting clinical follow-up.

[0150] To achieve a linkage between early warning and personalized health management, this module generates specific personalized intervention recommendations based on key risk indicators identified by abnormal impact markers, in conjunction with an intervention knowledge base. The knowledge base can be derived from: clinical medical guidelines (such as diabetes and cardiovascular risk management guidelines published by the ADA and ESC); structured physician prescription templates (intervention recommendation rule sets in electronic medical record systems); medical order knowledge graphs; or chronic disease management databases.

[0151] For example, if the abnormal item is "increased blood glucose variability", the intervention suggestion can be "it is recommended to optimize insulin dosage distribution or adjust meal structure".

[0152] If the system identifies "elevated CRP levels accompanied by a rapid increase in the score", it can indicate "there is an inflammation-related cardiovascular risk, and it is recommended to re-examine and evaluate anti-inflammatory intervention".

[0153] The final intervention recommendations are pushed to user terminals (such as apps or wearable device interfaces) or doctor platforms through the system interface. The recommendations include: identified high-risk key indicators; the direction of abnormal indicators and deviations compared to historical data; explanatory text; and recommended health management measures or examination suggestions. The system also allows doctors to manually revise the recommendations, or users to mark the implementation status of the recommendations, forming a closed-loop health management record and continuously improving the effectiveness of system interventions.

[0154] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An early warning system for coronary artery calcification in patients with type 2 diabetes, characterized in that: include: The data acquisition module is used to acquire continuous physiological parameter data of the target user; The risk modeling module is used to construct a multi-parameter dynamic correlation model based on the collected continuous physiological parameter data and the evolution characteristics of coronary artery calcification in patients with type 2 diabetes in the historical clinical sample database, so as to reflect the time evolution trend of coronary artery calcification. The indicator weight calculation unit is used to assign corresponding weights to each parameter based on the statistical analysis results of the sensitivity of different indicators to coronary artery calcification, and output the risk impact factor vector. The machine learning analysis module uses an ensemble learning algorithm to iteratively train a multi-parameter dynamic correlation model and uses risk impact factor vectors to predict and update high-risk events. The early warning trigger module generates an early warning signal and pushes it to the user terminal or doctor's terminal when the trend of coronary artery calcification in the prediction results exceeds the preset risk threshold. The weakness identification module is used to determine the group of indicator parameters that contributes the most to the risk in the prediction results, thereby guiding the direction of intervention.

2. The early warning system for coronary artery calcification in patients with type 2 diabetes as described in claim 1, characterized in that: The continuous physiological parameter data include blood glucose fluctuation curves, heart rate variability, blood pressure, inflammatory marker levels, glycated hemoglobin, and previous CAC scores.

3. The early warning system for coronary artery calcification in patients with type 2 diabetes as described in claim 1, characterized in that: The risk modeling module includes: The continuous physiological parameter data obtained by the data acquisition module are preprocessed; We extracted physiological parameter evolution sequences and corresponding coronary artery calcification progression trajectories from multiple groups of known type 2 diabetes patients from a historical clinical sample database. We then clustered the samples using time series similarity metrics to extract typical calcification evolution patterns. By combining preprocessed continuous physiological parameter data and calcification evolution patterns, a time series prediction model was constructed using a long short-term memory network to dynamically fit the trend of coronary artery calcification scores over a future period.

4. The early warning system for coronary artery calcification in patients with type 2 diabetes as described in claim 3, characterized in that: The indicator weight calculation unit includes: Based on the correspondence between known coronary artery calcification scores and continuous physiological parameter data in historical clinical samples, the sensitivity scores of each indicator to the coronary artery calcification prediction output at different scoring stages were calculated and normalized. The weights of each indicator after normalization were combined in the order of the original parameters to form a multidimensional risk influence factor vector.

5. The early warning system for coronary artery calcification in patients with type 2 diabetes as described in claim 4, characterized in that: The machine learning analysis module includes: The preprocessed continuous physiological parameter data and historical coronary artery calcification score samples were divided and trained separately through multiple basic learners to form an initial model pool. Based on the output of the base learner, a strong learner model is constructed using a weighted voting mechanism. The prediction outputs of multiple learners are nonlinearly fused, and the optimal fusion weights are determined through cross-validation.

6. The early warning system for coronary artery calcification in patients with type 2 diabetes as described in claim 5, characterized in that: This includes using risk impact factor vectors to predict and update high-risk events, including: The continuous physiological parameters and the risk impact factor vector are weighted item by item to construct a risk-weighted input matrix, where each physiological parameter value is multiplied by its corresponding weight and used as the model input feature. The risk-weighted input matrix is ​​input into the trained time series prediction model. The updated coronary artery calcification score prediction value is output through forward propagation of the model. Combined with the score change trend and historical trend map, a risk evolution sequence within the time window is generated. Based on the current predicted value and its rate of change, compare it with the set risk assessment threshold. When the predicted score reaches or exceeds the high-risk threshold, it is marked as a potential high-risk event. For potentially high-risk events, the input parameters and prediction results in this state are used as training samples and fed back to the multi-parameter dynamic association model to update the parameters of the multi-parameter dynamic association model.

7. The early warning system for coronary artery calcification in patients with type 2 diabetes as described in claim 6, characterized in that: The aforementioned early warning triggering module includes: Obtain the coronary artery calcification score sequence within the future time window output by the prediction model, calculate the absolute value and time derivative of the score, evaluate the growth rate, trend pattern and fluctuation amplitude of the score, and construct a risk trend score index accordingly. The risk trend scoring index is compared with multiple pre-set early warning triggering conditions, including: the predicted score exceeds the static high-risk threshold, the score growth rate per unit time exceeds the dynamic rate threshold, and the score curve shows nonlinear evolution characteristics such as a short-term acceleration surge. Based on the risk intensity, evolution trend type and historical individual risk curve in the comparison results, a graded early warning signal is automatically generated, which is divided into high risk, medium risk or level requiring attention.

8. The early warning system for coronary artery calcification in patients with type 2 diabetes as described in claim 7, characterized in that: The weakness localization module includes: performing sensitivity analysis on the weighted physiological parameters input to the time series prediction model, calculating the risk contribution of each physiological parameter in the current prediction output, and generating a feature contribution list; sorting the features contribution list by weight, selecting the top k key parameters that have the greatest impact on the predicted calcification score within the current prediction time window, and forming a set of high-impact factors, wherein the key parameters include blood glucose variability, heart rate variability, and CRP level.

Images