Group employee chronic disease digital mapping and closed-loop management method

Abstract

This invention discloses a digital mapping and closed-loop management method for chronic diseases among group employees, belonging to the field of smart healthcare technology. It addresses the problem of being unable to distinguish between the persistence and transient nature of abnormal indicators due to cross-timescale feature misalignment between retinal images and arteriosclerosis data. The method includes: training a cross-modal association network with physiological coupling constraints to construct a statistical distribution of group arteriosclerosis parameters; for employees with only a single retinal scan, using this network to estimate their arteriosclerosis parameters and quantify cross-timescale uncertainty, determining risk persistence based on group distribution deviation, and analyzing the weights of pathological attribution factors through an attribution network; generating combined intervention plans by combining risk and attribution information; executing closed-loop tracking and adaptively adjusting the plan based on effects and attribution bias triggers; and performing correlation analysis with operational data such as corporate sick leave rates to achieve enterprise-level performance evaluation and global correction.

Description

Technical Field

[0001] This invention belongs to the field of smart healthcare and health management information technology, specifically involving a digital mapping and management method that uses retinal scan images and arteriosclerosis detection data, combined with traditional Chinese medicine constitution identification, to continuously identify chronic disease risks, attribute pathological causes, and implement closed-loop health interventions for group employees. Background Technology

[0002] With the increasing demand for corporate health management, chronic disease risk assessment and intervention systems based on artificial intelligence and big data analysis have become a research hotspot. Retinal fundus images contain rich information on microcirculation and vascular health and have been widely used for early screening of chronic diseases such as hypertension and diabetes. Meanwhile, non-invasive parameters for detecting arteriosclerosis, such as pulse wave velocity (PWV) and ankle-brachial index (ABI), are internationally recognized quantitative indicators for assessing cardiovascular function and the degree of arteriosclerosis. Traditional Chinese medicine constitution identification, through methods such as tongue examination and medical history taking, determines an individual's constitution type from a holistic perspective, playing an important role in personalized health management.

[0003] Existing technologies have attempted to integrate and analyze health data from multiple modalities, such as retinal images, arteriosclerosis indicators, and traditional Chinese medicine constitution data, to more comprehensively assess individual health risks and develop intervention strategies. However, in actual health management scenarios for corporate employees, a key technical issue is often overlooked: feature misalignment across time scales. The specific manifestation and root cause of this problem lies in the fact that retinal scans are typically collected only during annual corporate physical examinations, reflecting only 1-2 discrete snapshots of vascular morphology per year, resulting in an extremely low sampling frequency. In contrast, arteriosclerosis parameters (such as PWV and ABI) can be collected frequently and quasi-continuously, monthly or even daily, using portable devices or integrated health monitoring systems, reflecting dynamic and continuous changes in vascular function. The sampling frequencies of the two differ by orders of magnitude, and their time scales are completely different. When existing technologies directly concatenate or fuse feature vectors acquired at different time scales and feed them into a risk assessment model, the model will be unable to effectively distinguish whether a particular abnormality in retinal vascular morphology (e.g., increased vascular tortuosity) represents a persistent, progressive pathological change or merely a transient physiological fluctuation caused by short-term physiological factors (such as excessive fatigue or transient blood pressure fluctuations). This misalignment of data features due to time scale discrepancies can lead to the system underreporting of employees with persistently high risk or over-intervening in employees with transient abnormalities, severely reducing the accuracy and efficiency of corporate health management.

[0004] Therefore, there is an urgent need for a digital mapping and closed-loop management method for chronic diseases that can overcome the aforementioned defects in cross-timescale feature misalignment and accurately distinguish whether the abnormalities of employees with only single, discrete retinal scan data are persistent high risks or transient fluctuations. Summary of the Invention

[0005] The purpose of this invention is to address the problem of cross-timescale feature misalignment in the aforementioned background technology by providing a digital mapping and closed-loop management method for chronic diseases for group employees. This method can utilize existing information from employees with only single retinal scan data, and through morphological inference and cross-timescale uncertainty quantification, accurately distinguish whether abnormalities in their health indicators are due to persistent high risk or transient fluctuations. It also provides pathological attribution analysis, ultimately forming a complete closed loop from risk assessment and precise intervention to enterprise performance evaluation.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for digital mapping and closed-loop management of chronic diseases for group employees, comprising the following steps: Step S1, Basic Model Construction. Obtain the retinal vascular morphology parameters and arteriosclerosis parameters of a reference employee group within the group, which possesses retinal scan images and synchronous arteriosclerosis detection data. Train a cross-modal association network to establish a mapping function from the retinal vascular morphology parameters to the arteriosclerosis parameters. The training loss function includes a physiological coupling constraint loss term based on a pre-defined physiological regression relationship between the retinal vascular morphology parameters and the arteriosclerosis parameters. Simultaneously, based on the arteriosclerosis detection data of all employees within the group stratified by demographic characteristics, construct a statistical distribution of group arteriosclerosis parameters. The retinal vascular morphology parameters include at least the vessel diameter ratio and vessel tortuosity.

[0007] Step S2, Risk Persistence Assessment and Attribution. Retinal scans and Traditional Chinese Medicine (TCM) constitution identification analysis are performed on each employee within the group to obtain retinal vascular morphology parameters and TCM constitution type determination results. For employees without simultaneous arteriosclerosis detection data, the mapping function is used to convert their retinal vascular morphology parameters into presumed arteriosclerosis parameters, and the cross-timescale uncertainty index during the prescribing process is output. The deviation of the presumed arteriosclerosis parameters from the statistical distribution of the group's arteriosclerosis parameters in its respective stratum is calculated, generating risk persistence assessment results and outputting a risk persistence index and risk persistence label. The presumed arteriosclerosis parameters, retinal vascular morphology parameters, and TCM constitution type determination results are input into a pre-trained attribution analysis network, outputting the contribution weights of each attribution factor. The attribution factors include at least two of the following: vascular endothelial function degeneration factors, vascular smooth muscle proliferation factors, and microcirculation inflammatory factors.

[0008] Step S3: Intervention plan generation. Using the risk persistence index, the risk persistence label, and the contribution weights of the attribution factors as inputs, employees are divided into different intervention levels. For each intervention level, a combined intervention plan is generated, which includes any two or all three of the following: lifestyle intervention plan, traditional Chinese medicine intervention plan, and medical-grade body shaping intervention plan.

[0009] Step S4: Closed-loop execution and effect tracking. The combined intervention plan is pushed to employees, and retinal vascular morphology parameters are collected for post-intervention re-examination. These re-examined retinal vascular morphology parameters are re-entered into step S2 to update the contribution weights of the risk persistence index and the attribution factor. The change in the risk persistence index before and after the intervention is calculated as an intervention effect evaluation index. When the intervention effect evaluation index is lower than a preset effect threshold, the intervention plan is adjusted, and step S3 is re-executed.

[0010] Step S5, Model Correction and Performance Evaluation. The improvement rate of the group risk persistence index for all employees before and after the statistical intervention is calculated. A causal correlation analysis is performed between this improvement rate and the group's corporate operational data to generate a corporate health management performance evaluation report. Simultaneously, the re-verified accuracy rate is fed back to step S1. When the accuracy rate falls below a preset accuracy threshold, an update to the statistical distribution of the group's arteriosclerosis parameters or retraining of the cross-modal association network is triggered. The corporate operational data includes at least sick leave rate data and employee satisfaction data.

[0011] In the above technical solution, the preferred training method for the cross-modal association network in step S1 is as follows: The vascular diameter ratio and vascular tortuosity of the reference employee group are used as input features, and the synchronously acquired pulse wave conduction velocity and ankle-brachial index are used as the first and second output labels, respectively. A dual-output deep regression network containing a shared feature extraction layer, a first task-specific output layer, and a second task-specific output layer is trained. The shared feature extraction layer is used to learn the common potential representation between retinal vascular morphology features and arteriosclerosis parameters. The physiological coupling constraint loss term is: ,

[0012] in, The morphological feature representation mapping function output by the shared feature extraction layer. This is a functional feature representation mapping function constructed based on a pre-defined medical regression relationship between pulse wave conduction velocity and vascular endothelial function index. The total training loss function is: ,

[0013] in, This is the error term for pulse wave conduction velocity prediction. This is the ankle-brachial index prediction error term. These are the preset coupling constraint weights. The core function of this physiological coupling constraint loss term is to force the mapping relationship learned by the network during training to conform to known physiological laws by introducing a consistency constraint between morphological feature representation and functional feature representation in the latent space, thus avoiding mapping outputs that violate medical common sense that may be generated by a purely data-driven model.

[0014] In the above technical solution, the preferred method for constructing the statistical distribution of the population arteriosclerosis parameters in step S1 is to stratify by age group and gender, calculate the sample mean and sample standard deviation of pulse wave velocity and ankle-brachial index for each stratum, and store them as a stratified normal distribution parameter table. The deviation in step S2 is the cumulative probability value of the presumed arteriosclerosis parameter in the normal distribution of its respective age and gender strata. This cumulative probability value reflects the relative position of the presumed value in the population—the lower the value, the greater the deviation of the presumed value from the population norm, and the higher the significance of the anomaly. The cross-timescale uncertainty index in step S2 is a quantified presumed confidence level based on the difference between the input features of the mapping function (i.e., discrete sampling time points of vessel diameter ratio and vessel tortuosity) and the continuous timescale corresponding to the presumed arteriosclerosis parameters. This uncertainty index is fundamentally different from general model cognitive uncertainty or data random uncertainty: it does not reflect insufficient model training or data noise, but specifically quantifies the inherent information loss in the cross-timescale operation of inferring dynamic functional state from static morphological snapshots. When the uncertainty index exceeds a preset threshold, the system will trigger a prompt to arrange an actual arteriosclerosis test for the employee to eliminate the presumed risk.

[0015] In the above technical solution, the preferred method for determining the risk persistence label in step S2 is as follows: A first probability threshold and a second probability threshold are set, where the first probability threshold is less than the second probability threshold; when the cumulative probability value is lower than the first probability threshold, a persistent high-risk label is output; when the cumulative probability value is higher than the second probability threshold, a transient fluctuation label is output; and when the cumulative probability value is between the two, an uncertain label is output. The training method for the attribution analysis network in step S2 is as follows: A reference group with synchronous retinal scan data, arteriosclerosis detection data, and clinical diagnostic labels is obtained. Retinal vascular morphology parameters, measured arteriosclerosis parameters, and TCM constitution type determination results are used as inputs, and the attribution factors of clinical diagnosis are used as output labels for training, enabling the attribution analysis network to learn the mapping relationship from multimodal inputs to attribution factor weights. The intervention levels described in step S3 are divided as follows: using the risk persistence index as the first dimension and the prediction confidence of the cross-modal association network as the second dimension, a four-quadrant hierarchical structure is constructed, including a low-risk, high-confidence health maintenance level, a low-risk, low-confidence enhanced monitoring level, a high-risk, high-confidence proactive intervention level, and a high-risk, low-confidence priority confirmation level. For employees in the proactive intervention level and the priority confirmation level, the corresponding intervention strategy is further queried from a preset attribution-intervention mapping table, taking into account the contribution weight of the attribution factors.

[0016] In the above technical solution, the specific content of the combined intervention program is as follows: the lifestyle intervention program includes a structured weight loss camp plan, which includes a diet prescription, an exercise prescription, and a behavioral cognitive intervention module; the traditional Chinese medicine intervention program includes a traditional Chinese medicine acupuncture program and a traditional Chinese medicine conditioning program matched based on the traditional Chinese medicine constitution type determination results; the medical-grade body shaping intervention program includes a cryolipolysis program matched after indication assessment based on the estimated arteriosclerosis parameters, the traditional Chinese medicine constitution type determination results, and the contribution weights of the attribution factors. The attribution-intervention mapping table uses the combination of attribution factor weights as an index field and associates corresponding recommended intervention strategies; when the weight of the vascular endothelial function degeneration factor exceeds a preset weight threshold, the mapping table associates and recommends exercise and nutritional intervention programs aimed at improving vascular endothelial function; when the weight of the vascular smooth muscle proliferation factor exceeds a preset weight threshold, the mapping table associates and recommends a traditional Chinese medicine conditioning program aimed at regulating vascular tone. The setting of this mapping table transforms the abstract attribution analysis results into specific, executable intervention strategies, realizing the automatic connection from etiological inference to targeted intervention.

[0017] In the above technical solution, the indication assessment for the cryolipolysis treatment includes: obtaining the employee's body fat percentage, waist-to-hip ratio, the presumed arteriosclerosis parameter, and the weight of the vascular endothelial degeneration factor; when the body fat percentage exceeds a preset body fat threshold, the waist-to-hip ratio exceeds a preset waist-to-hip ratio threshold, the presumed arteriosclerosis parameter is in a preset non-high-risk range, and the weight of the vascular endothelial degeneration factor is lower than a preset indication weight threshold, the employee is determined to be a candidate for cryolipolysis; based on the employee's body fat distribution characteristics and TCM constitution type, the treatment site and treatment parameters for cryolipolysis are determined. Incorporating attribution factor weights into the indication assessment is a key optimization point of this solution: when the weight of the vascular endothelial degeneration factor is high, it indicates that the employee's main problem is endothelial structural damage rather than simple fat accumulation. In this case, cryolipolysis may not be the optimal choice, and the system excludes such employees from the indication range accordingly.

[0018] In the above technical solution, the intervention effect evaluation index in step S4 is calculated as follows: the risk persistence index before intervention is recorded as the baseline index, and the risk persistence index obtained after post-intervention review is recorded as the current index. The difference or ratio between the baseline index and the current index is calculated as the intervention effect evaluation index. Simultaneously, the updated contribution weights of the attribution factors are compared with the contribution weights of the attribution factors before intervention, and the weight offset of each attribution factor is calculated. The weight offset can be calculated using Euclidean distance or Manhattan distance. When the intervention effect evaluation index is lower than the preset effect threshold, or when the weight offset of any attribution factor exceeds the preset offset threshold, the intervention plan is adjusted, and step S3 is re-executed: based on the updated attribution factor weights, the adjusted intervention strategy is queried from the attribution-intervention mapping table, and the combined intervention plan is updated. The significance of this dual-trigger mechanism is that even if the macro-risk index improves, if the deep-seated attribution structure changes adversely (for example, endothelial function improves but smooth muscle hyperplasia problem worsens), the system can still keenly capture and dynamically adjust the intervention strategy, avoiding the intervention direction deviation that may be caused by simply relying on the final effect indicator.

[0019] In the above technical solution, the causal association analysis in step S5 adopts a difference-in-differences model: employees participating in the health management project within the group are used as the treatment group, and employees who do not participate or delay participation within the group are used as the control group; the changes in sick leave rate and employee satisfaction before and after the intervention are calculated for the treatment group and the control group, respectively; using difference-in-differences estimation, the first causal effect estimate of the improvement rate of the group risk persistence index on the change in sick leave rate and the second causal effect estimate of the change in employee satisfaction are calculated. The causal association analysis also includes mediation effect analysis based on attribution factors: using the improvement of the group weight of each attribution factor as the mediating variable, the path-specific causal effect of the improvement rate of the group risk persistence index on the changes in sick leave rate and employee satisfaction through each attribution factor path is calculated. The value of this mediation effect analysis lies in the fact that it not only answers the question of whether the health intervention improved the enterprise's operating indicators, but also further reveals which pathological improvement path (improvement of endothelial function, inhibition of smooth muscle proliferation, or relief of inflammation) the health intervention affects the enterprise's operating indicators through, thereby providing attribution path-level decision support for the precise allocation of enterprise health resources. The corporate health management performance evaluation report also includes: employee distribution and trends at each intervention level, participation and completion rates of various combined intervention programs, mean and variance of risk persistence index improvement at each intervention level, statistics on the improvement of group weights of each attribution factor, causal effect attribution contribution of each attribution path, and corporate health investment return score calculated by weighting the first causal effect estimate and the second causal effect estimate.

[0020] The above technical solution also includes: a preset TCM constitution-cryoliposuction parameter mapping table. The mapping table uses TCM constitution type as an index field, and each index field is associated with a corresponding suitable treatment site code and recommended treatment energy range for cryoliposuction. When generating the medical-grade body sculpting intervention plan, the mapping table is queried based on the employee's TCM constitution type determination result to obtain the corresponding suitable treatment site code and recommended treatment energy range. The recommended treatment energy range is adjusted according to the weight of the vascular endothelial function degeneration factor—the higher the weight of the vascular endothelial function degeneration factor, the lower the upper limit of the recommended treatment energy range. This linkage mechanism is based on the following medical understanding: vascular endothelial function degeneration means a decline in microcirculation regulation capacity and a weakened tolerance and repair capacity of tissues to cryotherapy damage; therefore, a more conservative treatment energy is needed to avoid adverse reactions.

[0021] Based on the above technical solutions, this invention also provides a federated continuous learning extension scheme: Multiple group enterprises deploy local risk persistence discrimination models; each enterprise uses retinal scan data, traditional Chinese medicine constitution identification data, and arteriosclerosis monitoring data of its local employees to execute steps S1 to S5, encrypting and uploading the updated model parameters of the cross-modal association network, the updated parameters of the statistical distribution of the population arteriosclerosis parameters, the updated model parameters of the attribution analysis network, and the updated parameters of the attribution-intervention mapping table to a central parameter server; before uploading, the updated model parameters are subject to random noise added via a differential privacy mechanism to meet employee health requirements. Kang Data's privacy protection requirements; the central parameter server uses a federated averaging algorithm to aggregate the parameter updates uploaded by each enterprise, generating a global cross-modal association network, a global attribution analysis network, a global population atherosclerosis parameter statistical distribution, and a global attribution-intervention mapping table, and distributes them to each enterprise's local model; when a new enterprise connects for the first time, the Meta-Learning Model-Independent Algorithm (MAML) is used, with the aggregated global cross-modal association network parameters and global attribution analysis network parameters as meta-initial parameters, to perform rapid adaptation with a small number of local paired data of the new enterprise, generating an enterprise-specific cross-modal association network and attribution analysis network adapted to the characteristics of the new enterprise's employee group.

[0022] Compared with the prior art, the present invention has the following beneficial effects.

[0023] First, it effectively overcomes the problem of feature misalignment across time scales. Existing technologies directly stitch together retinal morphological features and arteriosclerosis functional features in the feature space. Since these two reflect physiological states at completely different time scales, this stitching produces misalignment artifacts due to semantic misalignment. This invention achieves the conversion from morphological to functional inference through a cross-modal mapping network and specifically introduces a cross-time scale uncertainty index to quantify the credibility of this inference operation. This endows a single morphological snapshot, which originally lacked temporal dimension information, with a temporal dimension for functional inference, fundamentally resolving the misalignment contradiction.

[0024] Secondly, it achieves a deeper understanding from risk warning to pathological attribution. Existing technologies typically stop at outputting a comprehensive risk score, lacking the ability to determine the underlying pathological causes of the risk. This invention, through an attribution analysis network, further infers the contribution weights of various pathological attribution factors (vascular endothelial function degeneration, smooth muscle hyperplasia, and microcirculatory inflammation) from estimated arteriosclerosis parameters, retinal morphological characteristics, and traditional Chinese medicine constitution types. This deepens the assessment from assessing the level of risk to understanding its origin, providing a direct decision-making basis for subsequently developing highly pathology-specific intervention programs.

[0025] Third, a dynamic adaptive closed loop of attribution-intervention is constructed. This invention introduces attribution factor weight shift as an independent trigger condition for intervention program adjustment for the first time, complementing the traditional program adjustment mechanism based solely on the final effect indicator. Even if the macro-level risk index improves, if the deep-seated attribution structure undergoes adverse changes, the system can still keenly capture and dynamically adjust the intervention strategy, achieving truly intelligent navigation of the intervention process.

[0026] Fourth, it achieves a complete closed loop from enterprise-level performance evaluation to global model calibration. This invention not only quantifies the causal contribution of health interventions to enterprise operational indicators through difference-in-differences models and mediation effect analysis, but also feeds back the accuracy of the re-verified judgment to the basic model construction steps, triggering updates to the population statistical distribution or retraining of the network, thus forming a long-term self-optimizing closed loop covering individual interventions and system evolution. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the following describes the specific implementation of this invention in detail using an employee health management scenario of a large group enterprise as an example. Figure 1 As shown. It should be noted that the following description is only used to explain the technical principles and implementation process of the present invention, and is not intended to limit the scope of protection of the present invention.

[0029] I. System Deployment and Basic Model Construction (Step S1) Within a large corporate group, a central processing system was built for the digital mapping and closed-loop management of employees' chronic diseases. This system achieves data interoperability with the company's human resources management system, medical examination center database, and employee health application app.

[0030] First, employees who had previously undergone non-mydriatic fundus photography and arteriosclerosis detection were selected from the group's health checkup database as a reference group. Retinal scan images of this reference group were preprocessed, and an arteriovenous vascular tree structure was automatically extracted using a deep convolutional neural network-based vascular segmentation model. Retinal vascular morphology parameters for each employee, including arterial diameter ratio (AVR) and vascular tortuosity, were calculated. Simultaneously, arteriosclerosis parameters, including pulse wave velocity (PWV) and ankle-brachial index (ABI), were retrieved from the health checkup database for this reference group.

[0031] Using the vessel diameter ratio and vessel tortuosity of the aforementioned reference employee group as input features, and their corresponding PWV and ABI as the first and second output labels, respectively, a cross-modal association network is trained. This network employs a dual-output deep regression network architecture comprising a shared feature extraction layer, a first task-specific output layer, and a second task-specific output layer. The shared feature extraction layer, composed of multiple stacked fully connected layers, is responsible for learning common latent representations related to arteriosclerosis parameters from retinal vessel morphology features. The two task-specific output layers receive the output of the shared feature extraction layer and each outputs PWV and ABI predictions through dedicated fully connected layers.

[0032] One of the key innovations of this approach lies in the design of the training loss function. In addition to the standard PWV prediction error term and ABI prediction error term, a physiological coupling constraint loss term is introduced into the loss function. This term constructs a functional feature representation mapping function H(·) based on the regression relationship between PWV and the vascular endothelial function index established in internationally published cardiovascular epidemiological studies (such as the Framingham Heart Study). During training, by minimizing the difference between the morphological feature representation G(·) output by the shared feature extraction layer and the functional feature representation H(·), the morphological latent representation learned by the network is forced to maintain consistency with the physiological representation derived from the true functional parameters in the feature space. This constraint mechanism effectively prevents the model from producing mapping outputs that violate known physiological laws when training data is limited, thus improving the medical reliability of the inference results.

[0033] After the above model training is completed, the arteriosclerosis detection data of all employees in the group are further collected and stratified by age group (each ten years is a group) and gender. For each group, the sample mean and sample standard deviation of PWV and ABI are calculated, and a stratified normal distribution parameter table is constructed and stored in the system database as a group reference benchmark for subsequent judgment of the degree of deviation of individual employees.

[0034] II. Digital mapping and risk persistence assessment for individual employees (Step S2) When an employee participates in the annual physical examination and only completes the fundus photography and traditional Chinese medicine constitution identification, but does not undergo arteriosclerosis testing for some reason, the system will execute the following process for that employee.

[0035] The system extracts vessel diameter ratio and vessel tortuosity from the employee's retinal scan image. It then calls the mapping function of the pre-trained cross-modal association network from step S1 to convert these morphological feature parameters into estimated arteriosclerosis parameters (estimated PWV and estimated ABI). During this estimation process, the system simultaneously quantifies and outputs a cross-timescale uncertainty index. This index can be quantified using the Monte Carlo Dropout method: a Dropout layer is enabled during the inference phase, and the same input feature is randomly forward-propagated multiple times. The standard deviation of the multiple prediction results is calculated, and this standard deviation serves as a measure of cross-timescale uncertainty. Its physical meaning is that since the input is a discrete-time snapshot of the morphology, while the output is a quasi-continuous timescale functional state, the input features themselves do not carry sufficient temporal resolution to completely determine the output functional state. The resulting uncertainty is quantified as this index. When this uncertainty index exceeds a preset threshold, the system automatically sends a reminder to the employee and their health manager, suggesting that the employee undergo actual arteriosclerosis testing to eliminate estimation errors.

[0036] Subsequently, the system retrieves the statistical distribution parameters of arteriosclerosis parameters for the employee's age and gender stratification. Taking the estimated PWV value as an example, assuming the mean PWV of the employee's stratum is μ and the standard deviation is σ, the system calculates the cumulative probability value of the estimated PWV value within the normal distribution of that stratum. This cumulative probability value reflects the degree of deviation of the estimated value from the homogeneous group: the lower the cumulative probability, the greater the deviation of the estimated value from the group's normality, the higher the significance of the abnormality, and the greater the likelihood that it is a persistent pathological change rather than a transient fluctuation. Based on a preset first probability threshold and a second probability threshold, the system compares the cumulative probability value with both to determine the risk persistence label: below the first threshold, it outputs a "persistent high risk" label; above the second threshold, it outputs a "transient fluctuation" label; and between the two, it outputs an "uncertain" label, while simultaneously outputting a quantified risk persistence index.

[0037] After completing the risk persistence assessment, the system enters the attribution analysis phase. Estimated arteriosclerosis parameters, original retinal vascular morphology parameters, and TCM constitution type determination results are input into a pre-trained attribution analysis network. The training data for this attribution analysis network comes from a reference group that has undergone comprehensive clinical diagnosis and has a clear pathological attribution. The network outputs the normalized contribution weights of each attribution factor, including vascular endothelial function degeneration factors, vascular smooth muscle proliferation factors, and microcirculatory inflammatory factors. Through the Softmax output layer, the sum of the weights of these three factors is always 1, facilitating the direct interpretation of the relative importance of each pathological pathway.

[0038] III. Intervention Plan Generation and Delivery (Step S3) The system takes the risk persistence index, risk persistence label and attribution factor contribution weight output in step S2 as comprehensive inputs, and uses the risk persistence index as the first dimension and the prediction confidence of the cross-modal association network (which can be converted from uncertainty indicators) as the second dimension to classify employees into health maintenance level, enhanced monitoring level, active intervention level or priority confirmation level.

[0039] For employees categorized into the active intervention or priority confirmation level, the system further invokes the attribution-intervention mapping table, using the attribution factor weight combination as an index to query the corresponding recommended intervention strategy. This mapping table pre-determines the correspondence between attribution factor weight combinations and intervention strategies: for example, when the weight of vascular endothelial function degeneration factors exceeds a preset threshold, it recommends aerobic exercise and specific nutritional supplementation programs aimed at improving vascular endothelial function; when the weight of vascular smooth muscle proliferation factors exceeds a preset threshold, it recommends traditional Chinese medicine acupuncture or specific herbal remedies aimed at regulating vascular tone; when the weight of microcirculation inflammatory factors exceeds a preset threshold, it recommends lifestyle adjustments and traditional Chinese medicine heat-clearing and dampness-removing programs aimed at anti-inflammation and improving microcirculation.

[0040] Based on the above query results, the system automatically generates a combined intervention plan, which includes at least two of the following: a lifestyle intervention plan, a traditional Chinese medicine intervention plan, and a medical-grade body shaping intervention plan. The lifestyle intervention plan includes a structured weight loss program (including a diet prescription, an exercise prescription, and a behavioral cognitive intervention module); the traditional Chinese medicine intervention plan includes a traditional Chinese medicine acupuncture and herbal medicine treatment plan matched based on the results of traditional Chinese medicine constitution type determination; and the medical-grade body shaping intervention plan includes a cryolipolysis plan matched after indication assessment.

[0041] For the indication assessment of cryolipolysis, the system adopts the following comprehensive judgment logic: It obtains the employee's body fat percentage, waist-to-hip ratio, estimated arteriosclerosis parameters, and vascular endothelial degeneration factor weights. An employee is considered a candidate for cryolipolysis only if their body fat percentage exceeds a preset threshold, their waist-to-hip ratio exceeds a preset threshold, their estimated arteriosclerosis parameters are within a preset non-high-risk range, and their vascular endothelial degeneration factor weight is below a preset indication weight threshold. The significance of including the vascular endothelial degeneration factor weight in the assessment is that an excessively high weight indicates structural damage to the employee's vascular endothelium and a decline in microcirculation regulation and repair capabilities; cryolipolysis in this case may increase the risk of adverse reactions, thus excluding such employees. For employees identified as indication candidates, the specific treatment site and parameters are determined based on their body fat distribution characteristics and TCM constitution type. The determination of treatment parameters also referenced a pre-defined TCM constitution-cryolipolysis parameter mapping table. This mapping table uses TCM constitution type as an index, records the corresponding appropriate treatment site code and recommended treatment energy range, and dynamically adjusts the upper limit of the energy range according to the weight of vascular endothelial function degeneration factor.

[0042] IV. Closed-loop execution, effect tracking, and adaptive adjustment (Step S4) The system pushes the generated combined intervention plan to employees through the enterprise health application app and sets the intervention period (e.g., three months). After the intervention period ends, employees go to the medical examination center to complete a follow-up examination, and the system collects the retinal vascular morphological characteristic parameters of their follow-up examination.

[0043] The system re-enters the re-examined retinal vascular morphology parameters into step S2, updating the employee's presumed arteriosclerosis parameters, cross-timescale uncertainty index, risk persistence index, and risk persistence label. Simultaneously, the re-examined data is input into the attribution analysis network to obtain the updated attribution factor contribution weights.

[0044] The system records the pre-intervention risk persistence index as the baseline index and the post-intervention risk persistence index as the current index, calculating the difference or ratio between the two as an indicator of intervention effectiveness. Simultaneously, it calculates the offset of each attribution factor weight before and after the intervention. The offset of each attribution factor weight can be calculated using Euclidean distance or Manhattan distance, i.e., calculating the distance between the updated attribution factor weight vector and the pre-intervention attribution factor weight vector. A preset offset threshold is set; when the weight offset of any attribution factor exceeds this threshold, it is determined that the attribution structure has changed significantly.

[0045] The system performs a dual evaluation of intervention effectiveness indicators and attribution factor offsets. If the intervention effectiveness indicators are below a preset threshold, it indicates that the intervention program has failed to effectively improve the employee's persistent risk status. If the weight offset of any attribution factor exceeds a preset offset threshold, it indicates that the employee's pathological attribution structure has changed significantly—even if the macro-risk index is improving, a previously minor pathological pathway may be becoming the primary concern. When either of these conditions is triggered, the system automatically re-executes step S3, prioritizing the query of the adjusted intervention strategy from the attribution-intervention mapping table based on the updated attribution factor weights, updating the combined intervention program, and pushing it to the employee. This dual-trigger mechanism ensures that the intervention strategy can respond promptly to deep-seated changes in the employee's physical condition, rather than simply waiting for macro-indicators to deteriorate before taking action.

[0046] V. Enterprise Performance Evaluation and Global Model Correction (Step S5) In the group's management backend, the system uses a preset time window (such as quarterly or annually) to statistically analyze the improvement rate of the group risk persistence index for all employees participating in the health management project. A difference-in-differences model is used for causal association analysis: employees participating in the project are treated, while employees who did not participate or delayed participation for various reasons are treated as the control group. The changes in sick leave rate and employee satisfaction scores before and after the intervention are calculated for both groups. Difference-in-differences estimation is used to eliminate time trends and selection biases, and the net causal effect of the health management project on the company's operational indicators is calculated.

[0047] Building upon this, the system further performs a mediation effect analysis based on attribution factors. Using the improvement in the group weight of each attribution factor as the mediating variable, it calculates the path-specific causal effect of the improvement rate of the group risk persistence index on the changes in sick leave rate and employee satisfaction through each attribution factor path. This analysis reveals which pathological pathway the intervention measures ultimately improve to affect corporate operating indicators, providing refined attribution path-level guidance for enterprises to optimize the allocation of health resources.

[0048] The system summarizes the above analysis results to generate an enterprise health management performance evaluation report. The report includes the distribution and changing trends of employees at each intervention level, the participation and completion rates of various combined intervention programs, the mean and variance of the improvement of the risk persistence index at each level, the statistics of the improvement of the group weight of each attribution factor, the causal effect attribution contribution of each attribution path, and the enterprise health investment return score calculated based on the causal effect estimate.

[0049] Meanwhile, the system continuously tracks the accuracy rate of all re-examination employees, i.e., whether employees labeled "persistently high risk" actually show worse functional indicator progress in the re-examination. When the accuracy rate drops below the preset accuracy rate threshold, the system automatically triggers the model calibration process, using newly accumulated reference employee data to trigger the update of the statistical distribution of population arteriosclerosis parameters or the retraining of cross-modal association networks, realizing a long-term self-optimization closed loop from enterprise-level performance feedback to system-level model evolution.

[0050] VI. Cross-enterprise Federated Continuous Learning Deployment When this solution is deployed across multiple group enterprises, a federated continuous learning architecture can be used. Each enterprise deploys a complete risk persistence discrimination model locally, independently executing all the above steps using local employee detection data. Each enterprise uploads its local model parameter updates (not the original employee health data) with random noise added via a differential privacy mechanism, then encrypts and uploads them to the central parameter server. The central parameter server aggregates the parameter updates from each enterprise using a federated averaging algorithm, generating a global cross-modal association network, a global attribution analysis network, a global population atherosclerosis parameter statistical distribution, and a global attribution-intervention mapping table, and distributes these to each enterprise to update its local model. When a new enterprise first connects to the federated learning network, the Meta-Learning Model-Independent Algorithm (MAML) is used, with the aggregated global model parameters as initial meta-parameters, to perform rapid small-sample adaptation on the new enterprise's limited local paired data, generating a specific model adapted to the characteristics of the new enterprise's employee population. This architecture achieves continuous evolution of model performance while protecting the data privacy of employees in each enterprise.

[0051] Through the detailed description of the above specific embodiments, those skilled in the art can clearly understand the structure, connection relationship, working principle, and specific implementation process of the various technical features of the present invention. The present invention uses a cross-modal association network and physiological coupling constraint mechanism to infer discrete retinal morphological features as continuous arteriosclerosis functional parameters; it quantifies the reliability of the inference operation through cross-timescale uncertainty indicators; it achieves risk persistence discrimination and pathological attribution through population statistical distribution deviation and attribution analysis networks; it achieves dynamic adaptive adjustment of intervention strategies through attribution-intervention mapping tables and attribution bias triggering mechanisms; it achieves causal evaluation of enterprise performance through difference-in-differences models and mediation effect analysis; and finally, it achieves long-term self-optimization across the entire system through discrimination accuracy feedback and federated continuous learning, thus completely solving the core technical problem of cross-timescale feature misalignment.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A digital mapping and closed-loop management method for chronic diseases among group employees, designed to address the problem in group employee health management where cross-timescale feature misalignment exists between retinal scan images and arteriosclerosis detection data, making it impossible to distinguish between persistent high risk and transient fluctuations of abnormal indicators in a single test. Its characteristics are: Includes the following steps: Step S1, Basic model construction; The retinal vascular morphology parameters and arteriosclerosis parameters of a reference employee group within the group, which possesses retinal scan images and synchronous arteriosclerosis detection data, are obtained. A cross-modal association network is trained to establish a mapping function from the retinal vascular morphology parameters to the arteriosclerosis parameters. The training loss function includes a physiological coupling constraint loss term based on a pre-defined physiological regression relationship between the retinal vascular morphology parameters and the arteriosclerosis parameters. Simultaneously, based on the arteriosclerosis detection data of all employees within the group stratified by demographic characteristics, a statistical distribution of group arteriosclerosis parameters is constructed. The retinal vascular morphology parameters include at least the vessel diameter ratio and vessel tortuosity. Step S2, Risk persistence assessment and attribution; Retinal scans and TCM constitution identification analysis are performed on each employee within the group to obtain retinal vascular morphology parameters and TCM constitution type determination results. For employees without synchronous arteriosclerosis detection data, the mapping function is used to convert their retinal vascular morphology parameters into presumed arteriosclerosis parameters, and the cross-timescale uncertainty index in the prescribing process is output. The deviation of the presumed arteriosclerosis parameters from the statistical distribution of the group's arteriosclerosis parameters in its respective stratum is calculated to generate risk persistence discrimination results and output risk persistence index and risk persistence label. The estimated arteriosclerosis parameters, retinal vascular morphology parameters, and TCM constitution type determination results are input into a pre-trained attribution analysis network, and the contribution weights of each attribution factor are output. The attribution factors include at least two of the following: vascular endothelial function degeneration factors, vascular smooth muscle proliferation factors, and microcirculation inflammatory factors. Step S3: Intervention plan generation; Using the risk persistence index, the risk persistence label, and the contribution weight of the attribution factor as input, employees are divided into different intervention levels. For each intervention level, a combined intervention plan is generated, which includes any two or all three of the following: lifestyle intervention plan, traditional Chinese medicine intervention plan, and medical-grade body shaping intervention plan. Step S4: Closed-loop execution and effect tracking; The combined intervention plan is pushed to employees and retinal vascular morphology parameters are collected for post-intervention re-examination. The re-examined retinal vascular morphology parameters are re-entered into step S2 to update the contribution weights of the risk persistence index and the attribution factor. The change in the risk persistence index before and after the intervention is calculated as the intervention effect evaluation index. When the intervention effect evaluation index is lower than the preset effect threshold, the intervention plan is adjusted and step S3 is executed again. Step S5, Model calibration and performance evaluation; The improvement rate of the group risk persistence index of all employees before and after the statistical intervention is analyzed. The improvement rate of the group risk persistence index is then used to perform a causal correlation analysis with the group's corporate operation data to generate a corporate health management performance evaluation report. At the same time, the accuracy rate of the re-examination is fed back to step S1. When the accuracy rate is lower than the preset accuracy rate threshold, the update of the statistical distribution of the group arteriosclerosis parameters or the retraining of the cross-modal association network is triggered. The corporate operation data includes at least sick leave rate data and employee satisfaction data.

2. The method for digital mapping and closed-loop management of chronic diseases for group employees according to claim 1, characterized in that, The training method for the cross-modal correlation network mentioned in step S1 is as follows: Using the vessel diameter ratio and vessel tortuosity of the reference employee group as input features, and the synchronously acquired pulse wave velocity (PWV) and ankle-brachial index (ABI) as the first and second output labels, a dual-output deep regression network containing a shared feature extraction layer, a first task-specific output layer, and a second task-specific output layer is trained. The shared feature extraction layer is used to learn the common potential representation between retinal vessel morphology features and arteriosclerosis parameters. The physiological coupling constraint loss term is: ， in, H is the morphological feature representation mapping function output by the shared feature extraction layer. It is a functional feature representation mapping function constructed based on the preset medical regression relationship between pulse wave conduction velocity and vascular endothelial function index; The total training loss function is: ， in, This is the error term for pulse wave conduction velocity prediction. This is the ankle-brachial index prediction error term. These are the preset coupling constraint weight coefficients.

3. The method for digital mapping and closed-loop management of chronic diseases for group employees according to claim 2, characterized in that, The method for constructing the statistical distribution of the group arteriosclerosis parameters in step S1 is as follows: stratify by age group and gender, calculate the sample mean and sample standard deviation of pulse wave conduction velocity and ankle-brachial index for each stratum, and store them as a stratified normal distribution parameter table. The degree of deviation mentioned in step S2 is the cumulative probability value of the estimated arteriosclerosis parameter in the normal distribution of its age and sex stratification; The cross-timescale uncertainty index in step S2 is: the estimated confidence level quantified based on the difference between the discrete sampling time points of the vessel diameter ratio and vessel tortuosity, which are the input features of the mapping function, and the continuous time scale corresponding to the estimated arteriosclerosis parameters; when the cross-timescale uncertainty index exceeds the preset uncertainty threshold, a prompt is triggered to arrange an actual arteriosclerosis test for the employee.

4. The method for digital mapping and closed-loop management of chronic diseases for group employees as described in claim 3, characterized in that, The method for determining the risk persistence label in step S2 is as follows: a first probability threshold and a second probability threshold are set, wherein the first probability threshold is less than the second probability threshold; when the cumulative probability value is lower than the first probability threshold, a persistent high-risk label is output; when the cumulative probability value is higher than the second probability threshold, a transient fluctuation label is output; when the cumulative probability value is between the first probability threshold and the second probability threshold, an uncertain label is output. The training method of the attribution analysis network in step S2 is as follows: a reference group with synchronous retinal scan data, arteriosclerosis detection data and clinical diagnostic labels is obtained. The retinal vascular morphology parameters, measured arteriosclerosis parameters and TCM constitution type determination results are used as inputs, and the attribution factors of clinical diagnosis are used as output labels for training, so that the attribution analysis network learns the mapping relationship from multimodal input to attribution factor weights. The intervention levels described in step S3 are divided as follows: using the risk persistence index as the first dimension and the prediction confidence of the cross-modal association network as the second dimension, a four-quadrant hierarchical structure is constructed, including: low-risk high-confidence health maintenance level, low-risk low-confidence enhanced monitoring level, high-risk high-confidence active intervention level, and high-risk low-confidence priority confirmation level. For employees in the active intervention level and the priority confirmation level, the corresponding intervention strategy is further queried from the preset attribution-intervention mapping table by combining the contribution weight of the attribution factors.

5. The method for digital mapping and closed-loop management of chronic diseases for group employees according to claim 4, characterized in that, The combined intervention plan mentioned in step S3 is as follows: The lifestyle intervention program includes a structured weight loss camp program, which comprises dietary prescriptions, exercise prescriptions, and behavioral cognitive intervention modules. The TCM intervention program includes a TCM acupuncture program and a TCM herbal conditioning program that are matched based on the TCM constitution type determination results. The medical-grade body shaping intervention program includes a cryolipolysis program matched after indication assessment based on the estimated arteriosclerosis parameters, the TCM constitution type determination results, and the contribution weight of the attribution factors. The attribution-intervention mapping table uses the weight combination of attribution factors as an index field, and the associated records contain corresponding recommended intervention strategies. When the weight of the vascular endothelial function degeneration factor exceeds a preset weight threshold, the mapping table recommends exercise and nutritional intervention programs aimed at improving vascular endothelial function. When the weight of the vascular smooth muscle proliferation factor exceeds a preset weight threshold, the mapping table recommends traditional Chinese medicine conditioning programs aimed at regulating vascular tone.

6. The method for digital mapping and closed-loop management of chronic diseases for group employees according to claim 5, characterized in that, The indication assessment for the cryolipolysis program includes: obtaining the employee's body fat percentage, waist-to-hip ratio, the presumed arteriosclerosis parameter, and the weight of the vascular endothelial degeneration factor; when the body fat percentage exceeds a preset body fat threshold, the waist-to-hip ratio exceeds a preset waist-to-hip ratio threshold, the presumed arteriosclerosis parameter is in a preset non-high-risk range, and the weight of the vascular endothelial degeneration factor is lower than a preset indication weight threshold, the employee is determined to be a candidate for cryolipolysis; based on the employee's body fat distribution characteristics and TCM constitution type, the treatment site and treatment parameters for cryolipolysis are determined.

7. The method for digital mapping and closed-loop management of chronic diseases for group employees according to claim 1, characterized in that, The intervention effect evaluation indicators mentioned in step S4 are calculated as follows: The risk persistence index before intervention is recorded as the baseline index, and the risk persistence index obtained after intervention is recorded as the current index. The difference or ratio between the baseline index and the current index is calculated as the evaluation index of the intervention effect. The updated contribution weights of the attribution factors are compared with the contribution weights of the attribution factors before the intervention, and the weight offset of each attribution factor is calculated; the weight offset is calculated using Euclidean distance or Manhattan distance. When the intervention effect evaluation index is lower than the preset effect threshold, or when the weight offset of any attribution factor exceeds the preset offset threshold, the intervention plan adjustment is triggered, and step S3 is re-executed: based on the updated attribution factor weights, the adjusted intervention strategy is queried from the attribution-intervention mapping table and the combined intervention plan is updated.

8. The method for digital mapping and closed-loop management of chronic diseases for group employees according to claim 1, characterized in that, The causal association analysis in step S5 uses a difference-in-differences model: employees who participate in the health management project within the group are used as the treatment group, and employees who do not participate or delay participation within the group are used as the control group; the changes in sick leave rate and employee satisfaction before and after the intervention are calculated for the treatment group and the control group, respectively; the difference-in-differences estimation is used to calculate the first causal effect estimate of the improvement rate of the group risk persistence index on the change in sick leave rate, and the second causal effect estimate of the change in employee satisfaction. The causal association analysis also includes mediation effect analysis based on attribution factors: using the improvement of the group weight of each attribution factor as the mediating variable, the path-specific causal effect of the improvement rate of the group risk persistence index on the change in sick leave rate and the change in employee satisfaction through each attribution factor path is calculated. The corporate health management performance evaluation report also includes: employee distribution and trends at each intervention level, participation and completion rates of various combined intervention programs, mean and variance of risk persistence index improvement at each intervention level, statistics on the improvement of group weights of each attribution factor, causal effect attribution contribution of each attribution path, and corporate health investment return score calculated by weighting the first causal effect estimate and the second causal effect estimate.

9. The method according to claim 1, characterized in that, Also includes: A preset TCM constitution-cryoliposuction parameter mapping table is provided. The mapping table uses TCM constitution type as the index field. Each index field is associated with a corresponding cryoliposuction suitable treatment site code and recommended treatment energy range. When generating the medical-grade body shaping intervention plan, the mapping table is queried based on the employee's TCM constitution type determination result to obtain the corresponding suitable treatment site code and recommended treatment energy range; The recommended treatment energy range is adjusted according to the weight of the vascular endothelial function degeneration factor; when the weight of the vascular endothelial function degeneration factor is higher, the upper limit of the recommended treatment energy range is correspondingly reduced; the medical-grade shaping intervention program includes at least one parameter among treatment site, probe model, cooling temperature and treatment duration.

10. The method for digital mapping and closed-loop management of chronic diseases for group employees according to any one of claims 1 to 9, characterized in that, It also includes federal continuous learning steps: Each of the group companies deployed a local risk persistence discrimination model; each company used local employees' retinal scan data, traditional Chinese medicine constitution identification data and arteriosclerosis monitoring data to perform steps S1 to S5, and encrypted and uploaded the updated values ​​of the model parameters of the cross-modal association network, the updated values ​​of the statistical distribution of the population arteriosclerosis parameters, the updated values ​​of the model parameters of the attribution analysis network, and the updated values ​​of the attribution-intervention mapping table to the central parameter server. Before uploading, random noise is added to the model parameters via a differential privacy mechanism to meet the privacy protection requirements of employee health data. The central parameter server uses a federated averaging algorithm to aggregate the parameter updates uploaded by each enterprise, generate a global cross-modal association network, a global attribution analysis network, a global population atherosclerosis parameter statistical distribution, and a global attribution-intervention mapping table, and distributes them to the local models of each enterprise. When a new enterprise is first connected, the Meta-Learning Model-Independent Algorithm (MAML) is used. The aggregated global cross-modal association network parameters and global attribution analysis network parameters are used as initial meta-parameters. The small number of local paired data of the new enterprise are used for rapid adaptation with few samples to generate an enterprise-specific cross-modal association network and attribution analysis network that are adapted to the characteristics of the new enterprise's employee group.

Images