A myopia occurrence risk assessment method based on multi-source data

Abstract

The application provides a myopia occurrence risk assessment method based on multi-source data, and relates to the technical field of myopia prevention and control, comprising: collecting physiological parameters and dynamic behavior data of a user through a multi-sensor, and preprocessing and labeling; applying a sliding window algorithm to detect short-term trends of behavior, generate a myopia risk index change process, identify false risk reduction intervals and delay window periods, mark myopia risk index rebound events and cumulative rebound risk peaks; when the behavior continues to meet the standard, starting a physiological parameter verification process, generating a closed-loop intervention suggestion based on the rebound warning point and collecting intervention feedback; comparing the score change with the measured physiological data direction, pausing the process and generating a fuse log when a conflict event is triggered; extracting abnormal data of prevention and control compliance to update a causal graph, nonlinearly adjusting model weights and sensitivity coefficients according to comprehensive feedback indicators, and iteratively optimizing the model until the warning is removed to output a report.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of myopia prevention and control, in particular to a myopia occurrence risk assessment method based on multi-source data. BACKGROUND

[0002] Myopia problem shows a significant growth trend worldwide, especially in the youth group, and has become a major public health challenge. Accurate assessment of individual myopia occurrence risk and timely intervention are crucial for preventing and controlling myopia development. In recent years, with the development of wearable devices and sensing technology, it has become more convenient to obtain user daily behavior and physiological parameter data, which provides the possibility for developing dynamic risk assessment methods.

[0003] Current myopia risk assessment techniques mainly rely on static questionnaires to collect historical information or use single type device data to build models. These schemes usually establish fixed weight scoring rules based on statistical analysis or use simple regression models for one-time prediction of risk. Some methods try to integrate behavior and physiological data, analyze time series changes to identify risk trends and provide intervention prompts.

[0004] Existing methods have obvious limitations: there is a lack of effective identification mechanism for the authenticity and continuity of user behavior improvement, which leads to evaluation results being easily disturbed by temporary changes; there is a lack of robust processing strategy for potential conflicts between multi-source heterogeneous data; and the response differences of different users to intervention measures and the possible error influence of measurement devices are not fully considered. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the deficiencies of the prior art, the present application provides a myopia occurrence risk assessment method based on multi-source data to solve the problems of lacking effective identification mechanism for the authenticity and continuity of user behavior improvement, leading to evaluation results being easily disturbed by temporary changes; lacking robust processing strategy for potential conflicts between multi-source heterogeneous data; and failing to fully consider the response differences of different users to intervention measures and the possible error influence of measurement devices.

[0007] (II) Technical solutions

[0008] To achieve the above purpose, the present application is implemented by the following technical solutions: a myopia occurrence risk assessment method based on multi-source data, 1. comprising the following steps:

[0009] S1 collects physiological parameters and dynamic behavior data of the user through multiple sensors, and generates behavior labels and physiological labels after preprocessing;

[0010] S2 maps the behavior label and the physiological label to establish a causal relationship, constructs a causal graph, aligns the physiological parameter data and the dynamic behavior data of different collection frequencies on the time axis through an interpolation and redundancy compression method, and constructs a time series data sequence; in the time series data sequence, a multi-dimensional vector composed of the behavior label and the physiological label is generated for each time node, and stored in a time series database;

[0011] S3 applies a sliding window algorithm to the time series data to detect a short-term improvement trend of user behavior, constructs a historical baseline based on a weighted mean of physiological parameters and dynamic behavior data of a user in a fixed period, constructs a myopia risk score model in combination with the historical baseline, and generates a myopia risk index, generates a myopia risk index change process according to user historical and current behavior data and physiological parameter change trend, the myopia risk index change process includes a score rising interval, a score falling interval, and a score stable interval; when a time period in which the myopia risk index decreases in the score falling interval but the change amplitude of the physiological parameter data does not reach a set physiological improvement threshold is identified, it is marked as a delay window period; in the delay window period, if the myopia risk index is detected to rise again and exceed the previous baseline value, it is marked as a myopia risk index rebound event, and the score falling interval is marked as a false risk falling interval; a cumulative rebound risk peak is marked at a myopia risk index rebound peak point;

[0012] S4 when detecting that the user behavior continuously meets a preset standard condition, starting a physiological parameter verification process, if no positive change of the physiological parameter data is observed or a rebound event is triggered, generating a closed-loop intervention suggestion based on a final rebound risk warning point, generating intervention feedback information by collecting dynamic behavior data after executing the closed-loop intervention suggestion;

[0013] S5 when the myopia risk index change process and the change direction of the actually measured physiological parameter data are opposite, or the difference between the dynamic behavior data collected by different sensors exceeds a tolerance range, a conflict event is triggered; after the conflict event is triggered, the current myopia risk index calculation process is suspended, the data processing window is locked, and a fuse log report is generated, and a device calibration reminder is sent, the device calibration reminder is generated based on device usage frequency and calibration history record;

[0014] S6, in view of the behavior deviation characteristics of part of the users in the intervention execution process, extracting the prevention and control compliance abnormal user data, updating the multi-dimensional vector and the causal graph, generating a comprehensive feedback index based on the closed-loop intervention suggestion execution record, the cumulative rebound risk peak value and the device calibration information; according to the comprehensive feedback index, nonlinearly adjusting the mapping weight of the behavior label to the physiological label of the myopia risk scoring model; using the adjusted myopia risk scoring model to regenerate the myopia risk index, if the condition of the final rebound risk warning point is still met, repeating steps S3-S6 until the warning is lifted; when the myopia risk scoring model reaches the threshold value, synchronizing to all running instances; and then regenerating the closed-loop intervention suggestion based on the final rebound risk warning point.

[0015] Preferably, in the process of collecting physiological parameters and dynamic behavior data of the user by multiple sensors in step S1, first, the device for measuring physiological parameters is subjected to a standardized self-checking operation, which includes zero-point calibration, response detection of data acquisition channels and measurement stability test automatically performed after the device is started, to ensure that each measurement result is within the set repeatability error range: the axial length is set to ±0.05 mm, and the diopter is set to within ±0.25 D, to ensure measurement accuracy. Physiological parameter data includes axial length, diopter, corneal curvature, interpupillary distance and visual acuity grade, which are measured and obtained by special devices such as axial length measuring instrument, autorefractor, corneal topography instrument, interpupillary distance ruler and standard visual acuity chart. At the same time, real-time online state detection is performed on all sensors for collecting dynamic behavior data, such as illumination sensor, distance sensor and posture recognition camera, before uploading the data. The detection process includes device power-on response test, data acquisition frequency check and signal-to-noise ratio analysis, to ensure the effectiveness and continuity of the uploaded data. After obtaining the raw data, pre-processing is performed on various types of data, and abnormal detection rules based on historical models are used to eliminate mutant values and error samples. At the same time, interpolation algorithms such as linear interpolation or spline interpolation are used to fill in the missing data, to ensure the time sequence integrity of the data. The pre-processing operation includes time series fusion processing of multiple time period data, and in the case of a downward trend in the score and no improvement in the physiological parameter data, a time window label is generated. The time window label includes event occurrence time, event type, confidence value, deviation amount of behavior label relative to historical baseline, deviation amount of physiological label relative to historical baseline, and time sequence continuity identification information. The pre-processed data is labeled according to the pre-defined behavior classification system and physiological change type, wherein the behavior label is generated according to the threshold range of eye use time and outdoor activity time, including long-time close-range eye use, high-frequency illumination variation and low environmental light exposure. Long-time close-range eye use refers to eye use time ≥ 2 hours / day; low outdoor exposure refers to outdoor activity time ≤ 30 minutes / day; the illumination frequency variation threshold is set to more than 10 changes per hour; the physiological label is generated according to the change trend of physiological parameters such as axial length growth rate and visual acuity decline amplitude, including rapid axial length growth and negative diopter change; all labels are accompanied by time stamp, confidence value and source device number. The confidence value is obtained by real-time detection of three core state indicators by multiple sensors, which are: power-on response time with a weight of 0.3, linearly deducted score when exceeding 3 seconds; sampling rate stability with a weight of 0.4, attenuated in proportion when fluctuating more than ±5%; signal-to-noise ratio with a weight of 0.3, stepwise weight reduction when less than 25 dB; the three scores are summed to generate a confidence value in the interval [0, 1], and are applied in stages: confidence value ≥ 0.8 is high confidence, the data can be directly used; confidence value 0.6-0.8 is medium confidence, interpolation compensation is required; confidence value < 0.6 is low confidence, automatically trigger device calibration.

[0016] Preferably, the step S2 of mapping the behavior label and the physiological label into a causal relationship first updates the weight of the causal path in the causal graph based on the public medical database or the database, the public medical database including PubMed literature database or clinical guideline database, the incremental update including identifying newly added causal paths and combining path source confidence level for fusion evaluation, for optimizing the false risk decline interval related to the cumulative rebound risk peak node; and using user historical data to optimize and adjust the weight of the core causal path between the behavior node and the physiological node, and then calling the pre-defined causal graph structure framework, the graph including the behavior node, the physiological node and the intermediate variable node, the behavior node corresponding to the user behavior mode such as long-time reading and short-time outdoor activity, the intermediate variable node being configured as a physiological response delay for a delay window period, the input edge weight of the node being = behavior improvement amplitude x time attenuation factor, and the output edge weight being a physiological lag coefficient; when the node activation value is greater than a threshold value, triggering a delay window period mark, for modeling the non-direct contact between behavior and physiological change such as eye fatigue and circadian rhythm disorder, a special physiological node type is set in the causal graph as a cumulative rebound risk peak node, the node storing the physiological parameter data of the nearsightedness risk index peak time stamp in the rebound event, such as axial length or diopter, and establishing a reverse correlation path with the behavior improvement label, for identifying the physiological deterioration state after the false risk decline; in this graph framework, the path weight is updated incrementally by introducing an external authoritative database, the incremental update process including path credibility calculation and new path identification, wherein the credibility calculation is completed according to the path frequency and sample consistency index; the data alignment adopts the method of interpolation and redundancy compression, the interpolation is used to reconstruct the behavior change curve under low sampling frequency, and the redundancy compression is used to eliminate the time error introduced by repeated sampling, finally all data are aligned to a unified time axis, so that the key behavior and physiological change nodes correspond accurately; in the process of constructing time series data, a multi-dimensional vector including original behavior data, physiological parameter value, behavior label, physiological label, time weight index and risk score gradient feature is generated for each time node, and stored in the time series database.

[0017] Preferably, in S3, a sliding window algorithm is used to dynamically identify short-term improvement trends of user behavior in the constructed time series data, and a historical baseline is constructed based on the weighted average of the user's physiological parameters and dynamic behavior data in the past 7 days, wherein the weight decreases by a time decay coefficient of 0.85, and then a myopia risk score model is constructed based on the historical baseline to generate a risk score, with a sliding window step of 1 hour for the past 24 hours, the behavior label data in each scanning window is labeled and the difference with the previous historical behavior baseline is calculated, the difference is combined with the trend direction to generate a behavior improvement indicator through the Euclidean distance, and if the behavior improvement indicator appears positive trend for 3 consecutive windows, it is marked as a potential improvement section; then the dynamic behavior data in the improvement section is paired with the corresponding physiological label to analyze whether it drives the synchronous change of physiological parameters; based on the pairing result, the myopia risk score model is started, the current behavior characteristic value and the physiological trend gradient weight are fused, the risk score is output, and the score change process curve is constructed based on the historical score trajectory; the score change process is divided into score rising interval, score falling interval and score stable interval, the boundary judgment standard of score change rate and fluctuation amplitude is set, the score change rate threshold is set to 0.05 points / hour, and the fluctuation amplitude threshold is ±0.1 points, combined with the mapping consistency of behavior and physiological indicators, the special stage of risk score decline but physiological parameter improvement is not obvious is identified; when it is detected that the score has decreased but has not reached the physiological improvement threshold in this stage, it is marked as a delay window period, the physiological improvement threshold is that the axial length improvement threshold is decreased by ≥0.05 mm / month or ≥0.2 mm / quarter; the refractive power improvement threshold is positive change ≥0.25 D / quarter; the corneal curvature improvement threshold is stable change trend ≤±0.05 mm; the pupil distance fluctuation threshold is day-to-day fluctuation amplitude ≤±0.5 mm; the vision level is improved by 1 level or more; if the physiological data slope does not reach the improvement threshold, the risk score correction is not triggered; the delay window period is set to 1 hour / day; it is observed whether there is a score rebound in the delay window period, if the score rises again and exceeds the previous baseline value by 0.2 points, it is judged that a rebound event has occurred, and the previous score falling interval is marked as a false risk falling interval, if the behavior label shows that the eye use time is reduced in the score falling interval, but the physiological parameter such as axial length growth rate does not slow down, the direction inconsistency lasts for more than 2 hours, and the physiological parameter fluctuation exceeds the threshold, such as refractive power single-day change >0.5 D, indicating measurement abnormality or external interference, it is directly marked as a false risk decline; further, the highest point at the end of the score rebound is marked to record the cumulative rebound risk peak.

[0018] Preferably, the marking condition of the cumulative rebound risk peak value is satisfied if any of the following conditions is met: the myopia risk index reaches the historical highest value after rebounding, the myopia risk index after rebounding enters the high risk level interval and exceeds the high risk level threshold, the myopia risk index after rebounding is more than 90 points, or the myopia risk index after rebounding is in the medium-high risk level and the rising rate exceeds 0.5 points / hour. If the peak value marking condition is met, the cumulative rebound risk peak value is marked at the myopia risk index rebound peak point.

[0019] Preferably, when the myopia risk index rebound curve appears a local maximum value and the peak value marking condition is met, the point is recorded as the cumulative rebound risk peak value. The determination method of the local maximum value is that the time sequence of the myopia risk index is denoted as I risk (t), and the local maximum value at time point t needs to meet the neighborhood extremum condition and the baseline exceeding condition at the same time. The neighborhood extremum condition is I risk (t) > I risk (t-Δt) and I risk (t) > I risk (t+Δt), where Δt is a preset time step, which is 1 hour by default and consistent with the sliding window step. The baseline exceeding condition is I risk (t) ≥ B baseline +ΔT, where B baseline is the dynamic weighted average value of the myopia risk index at 24 hours before the start time of the delay window period, and the weight is calculated according to the time decay coefficient 0.85. ΔT is a preset exceeding threshold, which is 0.2 points by default. The value is based on the minimum significant rising amplitude of rebound events in historical data. If the data at t±Δt is missing, it is filled by linear interpolation.

[0020] Preferably, the start signal of the physiological parameter data verification process is triggered when it is detected that the user behavior continuously meets the preset standard condition; first entering the initial stage, in which continuous physiological parameter data sets are synchronously collected by multiple sensors, including original data collection records of axial length, diopter, corneal curvature, interpupillary distance, and visual acuity grade; next entering the delay stage, in which the stable state of the user behavior label is continuously tracked, and the historical risk score trend is combined to compare the fluctuation of the score interval in the last three times and screen the micro-change characteristics of the physiological parameters; then entering the maintenance stage, which requires the user behavior to maintain the preset behavior standard condition for at least N observation periods, N is 3 consecutive periods, each period is 1 day, i.e. 3 days of continuous maintenance, and the slope calculation and threshold verification of the physiological parameter data are performed; if the change direction of the physiological parameter data is opposite or the slope does not reach the set physiological improvement threshold in any stage, the process continues without triggering the risk score correction; if the conditions of the final rebound risk warning point are still met, repeat steps S3-S6 until the warning is lifted; the warning lifting needs to be verified by double verification: the risk score fluctuation standard deviation is less than or equal to 0.5 points for 3 consecutive observation periods; and the change direction of the physiological parameters is consistent with the preset improvement direction and the amplitude meets the standard, i.e. the axial length is shortened by more than 0.02 mm per period. If the verification fails, return to step S3 to reiterate the model; when the myopia risk score model reaches the threshold, synchronize to all running instances; when the myopia risk score model has a dynamic risk score fluctuation standard deviation of less than or equal to 0.5 points for 3 consecutive iterations, and the false risk reduction interval misjudgment rate is reduced to less than or equal to 5%, it is determined that the threshold is reached; after the final rebound risk warning point is lifted, output the risk score report and closed-loop intervention suggestion; the closed-loop intervention suggestion content includes: increasing the user's daily outdoor activity time by not less than 2 hours, optimizing the reading distance to keep it at 30 to 40 cm, keeping the viewing distance of the tablet / mobile phone at not less than 40 cm, adjusting the light level of the environment where the user is to make it reach the appropriate eye use illuminance of 300 to 500 lux; after the intervention suggestion is generated, it is immediately pushed to the user terminal and the generation time, suggestion level, execution priority, and related risk peak source node are recorded; the intervention suggestion content is matched with the key path node in the causal graph and dynamically adjusted after comparing with the historical intervention feedback effect; after receiving the intervention suggestion, the user's execution process is recorded by real-time collection of dynamic behavior data by multiple sensors to form intervention feedback information; the intervention feedback information needs to include closed-loop intervention suggestion execution record, dynamic behavior data change after intervention, physiological parameter data detection result, and user subjective feeling log, in which the subjective feeling log obtains the user's description of visual clarity, eye fatigue degree, and eye comfort through the interactive terminal and performs text labeling processing.

[0021] Preferably, after the execution of the closed-loop intervention recommendation, an intervention feedback collection task is automatically started; first, through a data comparison mechanism, it is confirmed whether there is an execution record of the closed-loop intervention recommendation, which is a prerequisite for execution compliance judgment; if no relevant execution record is found, the user is directly determined as a prevention and control compliance abnormal user; if there is an execution record, it enters the data analysis stage, and the change amplitude of the user risk score is monitored within the specified observation period, and compared with the set risk score change threshold; if the risk score change amplitude after the execution of the intervention recommendation is less than the set threshold of 0.1 points or the change direction of the myopia risk index or physiological parameter data after the execution of the intervention recommendation is opposite to the preset improvement direction, and the change amplitude is more than 5% or the dynamic behavior data presents change characteristics including but not limited to abnormal fluctuations of eye use behavior or abnormal rearrangement of daily work and rest, it is marked as a compliance abnormal user; next, the change direction of the user's physiological parameter data is determined, if the change direction is opposite to the expected improvement direction, or the change amplitude is less than the set physiological response threshold, the physiological response threshold is set to be the change of the axial length of the eye less than 0.02mm / cycle and the change of the blood pressure 3mmHg / cycle, that is, it is recorded as a compliance deviation event; in addition, if the change direction of the risk score or the physiological parameter data is opposite to the preset improvement direction and the amplitude exceeds the reverse change judgment threshold, the reverse change judgment threshold is the reverse trend of the risk score or the physiological parameter change amplitude exceeding 0.3 points or 0.05mm, which constitutes a serious compliance abnormal event; further detect the change of the dynamic behavior data of the user during the intervention, if there is abnormal fluctuation of eye use behavior or significant rearrangement of daily work and rest time, it is also included in the compliance deviation feature range; all identified prevention and control compliance abnormal users are added with corresponding labels in the user behavior record, and used for updating the causal path weight in the subsequent score model adjustment; the re-collection process of the dynamic behavior data includes real-time behavior flow obtained from multiple sensors, wherein each behavior flow contains behavior type, execution time period, duration, behavior intensity score and confidence value; by comparing these feedback dynamic behavior data with the baseline data before execution, the intervention effect evaluation index value is obtained; the evaluation result is mapped with the original closed-loop recommendation, which is used to judge the intervention effect and the user response state; if the intervention effect evaluation is low or negative, the corresponding causal path node is marked for correction and is prepared to enter the score model updating process.

[0022] Preferably, the conflict event recognition process is triggered when a significant deviation or conflict is detected between the risk score change process and the measured physiological parameter data change direction; this process first compares the direction of the score curve change trend with the trend direction of the physiological parameter time series, if there is a direction inconsistency between the two and the trend angle deviation exceeds the threshold set as an angle of 30 degrees, it constitutes a potential conflict event; secondly, the dynamic behavior data collected by multiple devices is compared for the same behavior event recognition result, if there is a difference in key behavior type judgment and the difference exceeds the set tolerance range, the tolerance range is set as 10% of the behavior type judgment difference, which also constitutes a conflict event trigger condition; on the basis of potential conflict event recognition, a conflict score function is called to quantitatively evaluate the current data state; the conflict score function considers the residual distribution characteristics between the predicted value and the measured value, the consistency score index between the recognition results of each collection device, and the parameters of the historical data fluctuation range, to calculate the total conflict value; if the total conflict value exceeds the pre-set threshold value 0.7, it is determined as a formal conflict event and the current risk score process is immediately suspended; after suspension, the current time series data processing window is locked to avoid interference of the score result on the subsequent model; at the same time, a fuse log report is generated to record the detailed information of the conflict event, including the conflict type identifier, the trigger timestamp, the trigger condition description, the involved sensor number and the conflict score value; the fuse log is uploaded to the score model master node and used for model health assessment; device calibration reminders are pushed to the user terminal, the reminder content is generated by the joint of device usage frequency and historical calibration records, including detailed steps of instructing the user to perform operation level action verification on the specified device, identifying the device number and comparing the historical calibration data; after the device calibration is completed, the action effectiveness is verified, and according to the verification result, the risk score process is restored or the device is marked as abnormal for subsequent processing.

[0023] Preferably, in S5, the determination of the conflict event includes using a multi-model parallel prediction method to predict the change direction of the physiological parameter data, and triggering the conflict event determination when there is an angle deviation between the predicted physiological trend direction and the actual measured direction, and the absolute value of the difference exceeds a threshold, or the recognition results of different devices on the same behavior event differ by more than a set tolerance; the conflict event recognition uses a conflict score function to calculate a total conflict value based on the difference distribution characteristics between the predicted value and the measured value, the consistency score index between devices, and the historical data fluctuation range; when the total conflict value exceeds a set confidence threshold, the fuse operation is triggered; the myopia risk index calculation process includes: based on the causal graph of behavior labels and physiological labels, the path of the myopia risk index is developed, a dynamic score curve is generated combined with time series data, the score trend is evaluated, and a myopia risk index rebound node is determined, until the myopia risk index report output or closed-loop intervention suggestion is triggered; the causal graph constructs a causal relationship path based on user behavior, physiological response, and score fluctuation node; the fuse log report includes conflict type identification, trigger timestamp, trigger condition description, involved sensor number, and conflict score value; the device calibration reminder includes pushing device-level precise calibration process instructions based on device usage statistics, prompting the user to perform operation-level action verification, device number identification, and historical calibration record comparison verification steps.

[0024] Preferably, in S5, to further improve the accuracy of risk change judgment and the intervention effectiveness of group behavior trend response, the risk score change sequence of all users is input into a density-based spatial clustering algorithm together with the corresponding behavior label and physiological label sequence, the preset minimum sample number threshold is 50, and the distance threshold is the comprehensive weighted distance of the behavior dimension and the physiological dimension in the multi-dimensional Euclidean space, which is less than 0.75; in actual operation, the user data is preliminarily classified by introducing a predefined user risk level standard, which divides the risk score interval into four levels: low risk (0-20), medium risk (21-50), high risk (51-80), and extremely high risk (above 81), each level corresponds to a group of reference behavior patterns and physiological indicator typical values; then the common behavior trajectory features and physiological parameter change paths of each cluster group are extracted, the structural mapping between behavior causes and physiological results is established, and the risk propagation path graph of the cluster group is automatically generated; the importance of the nodes and paths in the graph is analyzed, and according to the preset importance scoring mechanism, the paths with a frequency greater than 5% and the nodes with an association degree greater than 0.6 are significant factors; the high-frequency cause behavior nodes and susceptible physiological nodes are marked; finally, a group-level intervention suggestion list is generated, each suggestion in the list is arranged from high to low according to the predefined risk mitigation effect level, and each suggestion is equipped with a standardized execution plan, including daily behavior target value, visual achievement progress bar, and score interval description of the suggested target population, to ensure that the intervention suggestions have unified executability when promoted to the group.

[0025] Preferably, after the intervention feedback information and physiological parameter detection data are uploaded to the central processing node in S6, the comprehensive feedback index generation phase is entered; first, the dynamic behavior data and physiological parameter data collected in each period are uniformly time-aligned to construct a cross-modal data association matrix; each feedback record is classified according to the intervention suggestion number, and the behavior improvement score, physiological response score and compliance confidence index in the corresponding feedback period are calculated; the behavior improvement score is determined according to the frequency, duration and intensity change amplitude of the user's execution of the suggestion, for example, daily behavior target values such as outdoor activities ≥ 60 minutes and close-range eye use ≤ 2 hours; the physiological response score is quantitatively evaluated based on the eye axis change rate, diopter change direction and pupil distance fluctuation level of the user before and after the intervention; the compliance confidence index is obtained by comparing the time tags in the suggestion execution record with the consistency of the behavior log data, including three dimensions of time accuracy of suggestion execution, consistency of behavior duration and device acquisition confidence level; the score results are integrated into a comprehensive feedback index, which is stored in a structured record format and includes fields such as suggestion number, user identification, intervention type label, execution time period, feedback period number and score value; the comprehensive feedback index also includes the cosine similarity between the behavior trend vector and the physiological change slope vector, which is used to represent whether the behavior change is consistent with the expected direction; if the similarity is less than a set minimum threshold, the record is marked as a feedback abnormal item and used for subsequent weight calibration.

[0026] Preferably, after the comprehensive feedback index is generated, the edge weights between the behavior nodes and the physiological result nodes in the causal graph are updated; first, all records with a comprehensive feedback index score higher than a set influence threshold in the intervention suggestion effective period are selected as valid samples; then the corresponding behavior node sequence, intervention suggestion type and physiological change result triplets in these samples are extracted and mapped to the causal graph paths; for each path, the weighted average score of the behavior improvement score and the physiological response score in the feedback index is calculated as an edge weight adjustment factor; if the score is significantly higher than the upper limit of the confidence range of the original edge weight, the weight value of the path edge is increased and marked as a positive feedback path; if the score is significantly lower than the lower limit of the edge weight confidence range, the weight value of the corresponding edge is reduced and marked as a negative feedback path; if the path weight change trend is the same for two consecutive evaluation periods, it is marked as a trend stable path to improve the causal prediction stability of the subsequent risk score model; in addition, the participation frequency and influence strength of the behavior nodes are evaluated; if a behavior node is repeatedly triggered in multiple feedback samples and the corresponding physiological change slope values are all significantly negative, the node is evaluated as a high-risk behavior node; the high-risk behavior node is automatically added to the user's personalized risk behavior list and given a higher priority and intervention intensity in subsequent intervention suggestion generation.

[0027] Preferably, after completing the update of the causal graph, the weight synchronization process of the myopia risk scoring model is triggered according to the latest path edge weight and feedback indicators; the synchronization process first extracts the behavior risk factor set adopted by the current myopia risk scoring model and its current weight vector; then the newly generated causal path weight is mapped into the risk factor space, and the factor weight adjustment value is calculated according to the mapping relationship; if the proportion of a certain risk factor in the new weight distribution increases by more than a set threshold, it means that its influence has increased, and the relative weight of this factor in the myopia risk scoring model is automatically increased; on the contrary, if its proportion decreases significantly, its reference contribution in the model is correspondingly reduced; after synchronization is completed, a round of fine-tuning operation based on feedback indicators is performed on each layer parameter in the myopia risk scoring model; the fine-tuning process is based on the latest batch of effective samples, and a dynamic weight resampling mechanism is used to construct a training subset; during the training process, the multi-factor residual error reduction method is used to adjust the parameters of the myopia risk scoring model, ensuring that each iteration optimizes towards the direction of minimizing the scoring error; after fine-tuning is completed, a new version of the myopia risk scoring model is generated and a new version backtesting process is performed, which includes using the last three periods of non-training data as the validation set to test the scoring stability and error range of the new myopia risk scoring model under different user populations and behavior types; if the scoring stability indicator of the myopia risk scoring model backtesting is not less than the previous version, and the error reduction rate reaches the set optimization threshold, the current myopia risk scoring model version is confirmed to be effective and pushed to all scoring terminals; otherwise, it is rolled back to the last stable version and the model iteration log is recorded.

[0028] Preferably, after completing the update of the myopia risk scoring model and pushing it to the terminal node, the myopia risk scoring model enters the continuous evolution monitoring stage; this stage takes a periodic evaluation mechanism as the core, and every other complete intervention feedback cycle, a new round of physiological parameters and behavior records are collected from the user end to evaluate the difference between the myopia risk scoring model prediction value and the actual feedback indicators in real time; if a certain myopia risk scoring model version shows an upward trend in scoring error rate in two consecutive cycles, the myopia risk scoring model evolution process is triggered; the evolution process includes three steps of model structure reconstruction, parameter initialization optimization and behavior factor re-screening; in the structure reconstruction link, based on the high-frequency patterns in the current user behavior space and their physiological response trends, a new causal topology substructure is automatically generated to replace the original branch with poor fitting effect; in the parameter initialization optimization process, the latest high-weight feedback samples in the last cycle are introduced as pre-training data to ensure that the initial state of the new myopia risk scoring model is more suitable for the current behavior scenario; in the behavior factor re-screening step, multi-round factor importance ranking and information gain evaluation methods are used to simplify and expand the original risk factor set, retaining the top N main factors with the largest contribution to the scoring result, while removing factors with insufficient correlation or redundancy; finally, a new round of scoring model candidate version is formed and enters the verification queue for deployment.

[0029] Preferably, while the myopia risk scoring model continues to evolve, the intervention strategy is also periodically optimized and upgraded. The intervention strategy template includes four modules: a behavioral suggestion library, a push timing algorithm, an interaction mode configuration, and a compliance prompt mechanism. Strategy optimization is based on the behavioral myopia risk index output by the myopia risk scoring model, matching different intervention intensities and suggestion types according to different risk levels. By retrospectively analyzing the user's historical intervention effects, efficient strategy combinations and suggestion expressions with high user acceptance are extracted and recorded in the strategy optimization database. After each optimization cycle, the priority suggestion sequence in the intervention strategy template is adjusted, and its corresponding execution probability vector is updated to improve strategy reach efficiency and user response rate. In terms of push timing, a user's daily routine model is introduced to dynamically calculate the optimal intervention reminder window, ensuring that suggestions are pushed during the time period when users have the highest executability. In the interaction mode configuration, a reinforcement learning mechanism is introduced to automatically select the most suitable combination of multimodal interaction methods such as text prompts, voice broadcasts, and image animations based on the user's past feedback preferences. The compliance prompt mechanism generates adaptive feedback text based on the user's completion rate and lag time for suggestions, improving user motivation and reducing behavior churn through personalized expression.

[0030] Preferably, after model evolution and strategy optimization are completed, a closed-loop control process is initiated to achieve adaptive scheduling of personalized intervention strategies. The closed-loop control process constructs a private intervention feedback loop for each user, including four stages: intervention suggestion generation, execution detection, feedback evaluation, and strategy update. Based on the current myopia risk scoring model output and the user's most recent feedback indicators, a new round of intervention suggestions is generated, and the suggestion intensity level is dynamically selected in combination with historical feedback trends. Subsequently, the implementation of suggestions and changes in physiological responses are detected in real time through smart devices. The collected data is processed in the cloud and then backfilled into the causal graph and scoring model as the basis for the next round of evaluation. During the feedback evaluation process, a time-weighted backtracking algorithm is used to evaluate the actual changes in the user's behavioral path and physiological trends in this round of intervention, and a fine-tuned suggestion template is generated based on the user's behavioral execution deviation. Finally, an intervention strategy scheduling table is formed, which makes a predetermined arrangement of intervention tasks in the future cycle, including intervention time points, suggestion content, execution form, priority, and dynamic adjustment threshold. If the user's behavior pattern or physiological state changes abruptly within a certain window period, the strategy adjustment mechanism is automatically triggered, temporarily suspending the current intervention plan and generating alternative suggestions based on the latest scoring results.

[0031] (III) Beneficial Effects

[0032] This invention provides a method for assessing the risk of myopia occurrence based on multi-source data. It has the following beneficial effects:

[0033] 1. The application realizes precise dynamic assessment of myopia risk by collecting user physiological and behavioral data through multi-sensor fusion; combines pre-set multi-dimensional behavior labels and physiological labels to construct a detailed causal relationship model, effectively revealing the internal relationship between behavior and physiological changes; the combination of sliding window algorithm and risk scoring model can capture short-term behavior improvement and risk rebound phenomenon in real time, ensuring the timeliness and accuracy of risk warning; refines the scoring change process, accurately identifies the delay window period and false risk decline stage, and avoids intervention failure caused by misjudgment; through multiple screening mechanisms and regression correction, the robustness and warning sensitivity of the risk peak marker are improved.

[0034] 2. The application introduces edge computing and asynchronous parallel processing mechanism in the overall design, ensuring the real-time and stability of data acquisition, processing and transmission; the pre-set multi-layer disaster recovery and load balancing strategy greatly enhances the reliability and high-concurrency response capability of the method, meeting the long-term monitoring needs of large-scale users; through detailed division of user group behavior patterns and risk paths, hierarchical and refined management of intervention suggestions is realized; the user's cognition and feedback willingness for their own data are improved, promoting the effective implementation of intervention measures; periodic follow-up combined with individualized effect report improves the intervention closed loop, supporting long-term tracking and model self-optimization; the combination of overall architecture and algorithm realizes the all-round, dynamic and efficient management of myopia risk monitoring, greatly improving the level of public health management. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the application will be described below in a clear and complete manner. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the application.

[0036] The embodiment of the present application provides a myopia occurrence risk assessment method based on multi-source data. First, a standardized self-checking process is performed on all sensor devices before 7:00 every morning. The process is triggered by a background control module immediately after the device is powered on. First, a built-in calibration module is used to start zero-point calibration of an ocular axis measuring instrument for measuring ocular axis length. Then, it is detected whether the data channel response meets the set sensitivity requirement. Finally, the stability of the output signal during the collection process is tested to determine whether the device is within the repeatability error standard range. The automatic optometry instrument and the corneal topography instrument also perform response detection and channel testing simultaneously to ensure the accuracy and reliability of the measured diopter and corneal curvature data. In the formal data collection stage, students queue into the detection channel. After the identity is recognized by the camera, the corresponding physiological data is automatically recorded, including the pupil distance information recorded by the pupil distance ruler and the unaided visual acuity level determined by using the standard visual acuity chart. The above collected information is labeled with a unified label and a time stamp. Before collecting dynamic behavior data, the light sensor, posture camera and distance sensor used to record dynamic behavior are detected for real-time connectivity. The device response check, whether the data sampling rate meets the set standard frequency, and the signal-to-noise ratio analysis are performed in sequence to evaluate the collection quality and exclude abnormal collection. The amsTSL2591 type sensor is used to monitor outdoor light, the VishayVEML7700 type sensor is used to monitor indoor reading light, and the ROHM BH175 type sensor is used to monitor light frequency. The ams AS7265x type multispectral sensor monitors the proportion of 400-450nm high-energy blue light. The distance sensor is an STVL53L5CX type sensor, which simultaneously detects the reading distance and the visual angle offset of 30-40c. The posture camera is an Intel RealSenseD455 type camera, which quantifies the head tilt angle. If the tilt angle is greater than 15°, an alarm is given. After all the original data is collected, the data preprocessing process is entered. The history data model is used to screen out and eliminate possible mutation values in the behavior flow. The linear interpolation algorithm is used to fill in the missing sections of dynamic behavior data caused by network delay, and the time sequence continuity is restored.

[0037] Then, the labeling link is entered. According to the students' eye use time in the classroom in the morning and afternoon and its continuity, behavior labels such as high-intensity close-range eye use are generated. If frequent light changes or long-time low-light exposure are identified, high-frequency variable light exposure or insufficient light labels are supplemented. If the ocular axis growth rate exceeds the threshold, physiological labels such as rapid ocular axis growth are given. All label data are bound with time stamps, data confidence values and sensor numbers for tracing.

[0038] The built-in causal graph framework is called to load the student's historical behavior records and physiological variation curves, confirm the possible paths between behavior nodes such as long-time reading or lack of outdoor activities and eye axis growth, and update the path credibility according to external databases, including updating the path frequency, sample consistency score, etc. Through interpolation, the low-frequency sample behavior change curve is restored and repeated data is compressed to realize data alignment to a unified time axis. On the time axis, a multi-dimensional vector sequence containing behavior labels, physiological labels, original numerical values and myopia risk index is generated and stored in the database. A sliding window algorithm is started to detect whether the student's short-term behavior has improved every hour with a 24-hour window. If high-frequency close-to-eye behavior is reduced and light is improved in three windows, it is considered that the behavior trend is improved, and it is marked as an improved section. At this time, it is analyzed whether the behavior change drives the synchronous improvement of physiological parameters. If the eye axis growth curve tends to be flat, it is considered that the trend is consistent and the myopia risk index scoring mechanism is triggered. The scoring value is generated by fusing behavior and physiological labels to form a trend change curve. If the score is found to be down but the physiological parameters have not reached the improvement threshold, it is marked as a delay window period and continued to be observed whether it rebounds. If the score rises again, it is confirmed that there is a rebound event, and the score peak and time node are recorded for subsequent intervention.

[0039] When it is confirmed that the student's behavior is stable and meets the preset standard, such as high-frequency outdoor activities for three consecutive days and classroom eye behavior meeting the standard, the physiological parameter verification process is triggered. From the initial stage, continuous eye axis and other physiological data records are collected before each class. In the delay stage, the recent three score changes are compared. If the eye axis change slope is still below the set threshold, the observation period is extended. Otherwise, it enters the maintenance stage to continue to verify stability. If the score improves but the eye axis does not improve synchronously or the score rebounds, the closed-loop intervention process is started. A standard intervention scheme including recommended outdoor time, reading distance and light environment is generated and pushed to the student terminal. The intervention content, level and associated risk node are recorded. After the user performs the recommended behavior, the behavior feedback flow is collected through the sensor and combined with the subjective feeling log to complete the closed-loop intervention feedback.

[0040] After the intervention is performed, it is confirmed whether there is consistency between the intervention record and the collected data. If the intervention exists but the score does not improve or the physiological parameters worsen, it is recorded as abnormal compliance and the user's behavior is re-labeled. The differences in behavior and physiology before and after the intervention are analyzed, the intervention effect score is calculated and used for subsequent model correction. If the score trend and physiological data change conflict are detected in the feedback, the score fuse process is triggered to suspend the current risk assessment and perform device calibration. Finally, the data in each stage is summarized to perform spatial clustering analysis to identify high-risk group behavior and physiological evolution path, generate intervention suggestion list and causal path correction for the school student group. All updates are fed back to the scoring model core node to trigger weight synchronization, parameter fine-tuning and version update process to form an evolving myopia risk index closed loop.

[0041] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A method for assessing a risk of myopia occurrence based on multi-source data, characterized in that, The method comprises the following steps: S1, collecting physiological parameter data and dynamic behavior data of a user through a multi-sensor, generating behavior labels and physiological labels after preprocessing, wherein the physiological parameter data comprises axial length, diopter, corneal curvature, interpupillary distance and visual acuity grade, and the dynamic behavior data comprises illumination parameters, reading distance and head tilt angle; S2, mapping the behavior labels and the physiological labels to establish a causal relationship and constructing a causal graph; aligning the physiological parameter data and the dynamic behavior data with different collection frequencies on a time axis through an interpolation and redundancy compression method to construct a time series data sequence; in the time series data sequence, a multi-dimensional vector composed of the behavior labels and the physiological labels is generated for each time node and stored in a time series database; S3, applying a sliding window algorithm to the time series data to detect a short-term improvement trend of user behavior, constructing a historical baseline based on a weighted mean of physiological parameters and dynamic behavior data of the user in a fixed period, and then constructing a myopia risk score model in combination with the historical baseline to generate a myopia risk index, generating a myopia risk index change process according to user historical and current behavior data and physiological parameter change trend, wherein the myopia risk index change process comprises a score rising interval, a score falling interval and a score stable interval, when a time period in which the myopia risk index decreases in the score falling interval but the change amplitude of the physiological parameter data does not reach a set physiological improvement threshold is identified, a delay window period is marked; in the delay window period, if the myopia risk index is detected to rise again and exceed the previous baseline value, a myopia risk index rebound event is marked, and the score falling interval is marked as a false risk falling interval; a cumulative rebound risk peak value is marked at a myopia risk index rebound peak point; when there are multiple cumulative rebound risk peak values, the peak point with the highest myopia risk index and the fastest rising rate is selected and recorded as a final rebound risk warning point; S4, when it is detected that the user behavior continuously meets a preset standard condition, a physiological parameter verification process is started, if no positive change of the physiological parameter data is observed or a rebound event is triggered, a closed-loop intervention suggestion is generated based on the final rebound risk warning point, and intervention feedback information is generated after the closed-loop intervention suggestion is executed; S5, when the myopia risk index change process is opposite to the change direction of the actually measured physiological parameter data, or the difference between the dynamic behavior data collected by different sensors exceeds a tolerance range, a conflict event is triggered; after the conflict event is triggered, the current myopia risk index calculation process is suspended, a data processing window is locked and a fuse log report is generated, and a device calibration reminder is sent, wherein the device calibration reminder is generated based on device usage frequency and calibration history. ​ ​ ​ ​ S6, for the user showing behavior deviation characteristics in the intervention execution process, extract the prevention and control compliance abnormal user data, update the multi-dimensional vector and the causal graph, generate the comprehensive feedback index based on the closed-loop intervention suggestion execution record, the accumulated rebound risk peak value and the equipment calibration information; adjust the causal relationship mapping weight of the behavior label to the physiological label according to the comprehensive feedback index; regenerate the myopia risk index using the adjusted myopia risk scoring model, and repeat the steps S3-S6 until the warning is removed if the condition of the final rebound risk warning point is still met; synchronize the myopia risk scoring model to all running instances when the threshold is reached; and regenerate the closed-loop intervention suggestion based on the final rebound risk warning point again. 2.The myopia risk assessment method based on multi-source data according to claim 1, characterized in that: In the S2: The causal graph includes behavior nodes, physiological nodes and intermediate variable nodes, wherein the physiological nodes include rebound risk peak nodes, and a dedicated intermediate variable node is configured for identifying a delay window period, for mapping the indirect relationship between behavior changes and physiological changes; The causal relationship mapping of the behavior label and the physiological label includes: based on the public medical database or the database, the weight of the causal path in the causal graph is incrementally updated, the incremental update includes identifying the newly added causal path and combining the confidence level of the path source for fusion evaluation, for optimizing the causal path weight related to the cumulative rebound risk peak value in the false risk reduction interval; And using the user historical data to optimize and adjust the core causal path weight between the behavior nodes and the physiological nodes; the cumulative rebound risk peak node is the physiological parameter data at the time point corresponding to the cumulative rebound risk peak value; The alignment time axis includes aligning the key turning points of the myopia risk index change curve by using interpolation reconstruction and redundancy compression strategies on different sampling frequency data sources, and retaining the behavior change burst nodes and the high correlation fragments of the corresponding physiological indicators; the behavior change burst node is identified by gradient analysis of dynamic behavior data, which is an event point whose change amplitude in unit time exceeds 2 times of the historical baseline standard deviation; the myopia risk index change curve is a dynamic visualization curve generated based on the myopia risk index change process; The time series data construction includes generating a multi-dimensional vector containing behavior characteristics and physiological characteristics at each time node, the multi-dimensional vector contains original data values, myopia risk index gradient characteristics, state markers of the delay window period and time window weight indicators; the behavior characteristics and physiological characteristics include original data values plus behavior labels and physiological labels; the original data values are the original values of the physiological parameter data and the dynamic behavior data. 3.The myopia risk assessment method based on multi-source data according to claim 2, characterized in that: In the S3: at the end of the preset delay window, the rebound change of the myopia risk index is captured based on the fitting model of the behavior label and the physiological label, and the cumulative rebound risk peak value is generated to correct the myopia risk index result; the previous baseline value refers to the dynamic weighted average value of the myopia risk index within 24 hours before the start of the delay window period; when the local maximum value of the myopia risk index rebound curve appears and meets the peak value marking condition, the point is recorded as the cumulative rebound risk peak value; The marking condition of the false risk reduction interval includes any one of the following conditions: (a) the myopia risk index rises after the end of the delay window period and exceeds the score value before the delay window period; (b) the myopia risk index rises by more than 20% of the lowest score value during the delay window period after the end of the delay window period; (c) the myopia risk index rises at a rate of more than 0.5 points / hour; (d) the score reduction interval lasts less than 2 hours; (e) the behavior improvement trend in the score reduction interval is inconsistent with the physiological parameter improvement trend; (f) the physiological parameter changes by more than a preset threshold in the score reduction interval; The cumulative rebound risk peak is marked at the myopia risk index rebound peak point if any one of the following conditions is met: (a) the myopia risk index reaches a historical highest value after rebounding; (b) the myopia risk index value enters a high risk level interval after rebounding and exceeds the high risk level threshold; (c) the myopia risk index exceeds 90 points after rebounding; (d) the myopia risk index is in a medium-high risk level and rises at a rate of more than 0.5 points / hour after rebounding. 4.The myopia risk assessment method based on multi-source data according to claim 1, characterized in that: In the S4, the judgment standard of the physiological parameter data verification process is that the change trend of the physiological parameter data continuously meets a preset target improvement direction, and the change slope or amplitude exceeds a set minimum physiological improvement threshold in N consecutive preset observation periods; The closed-loop intervention suggestions include increasing outdoor time, optimizing reading distance, and adjusting lighting environment; the intervention feedback information includes closed-loop intervention suggestion execution record, dynamic behavior data change, physiological parameter data detection result, and subjective feeling log; The determination condition of the prevention and control compliance abnormal user includes any one of the following conditions: (a) the execution record of the closed-loop intervention suggestion is missing in the intervention feedback information; (b) after executing the closed-loop intervention suggestion, the myopia risk index changes by less than 5 points of the set myopia risk index change threshold within a specified period; (c) after executing the intervention suggestion, the change direction of the physiological parameter data is opposite to the expected improvement direction, or the change amplitude is less than the set physiological response threshold; (d) after executing the intervention suggestion, the myopia risk index or physiological parameter data changes in a direction opposite to the preset improvement direction, and the change amplitude is more than 5%; (e) after executing the intervention suggestion, the dynamic behavior data presents a change feature including abnormal fluctuation of eye use behavior or abnormal rearrangement of daily routine.

5. The myopia risk assessment method based on multi-source data according to claim 1, characterized in that: In the S6, the generation process of the comprehensive feedback indicator includes combining: the intervention behavior execution record of the user, the myopia risk index change trend, the prevention and control compliance abnormal user data, and the device calibration log data, and comparing and analyzing similar mode samples in the historical risk event library to form a feedback vector for weight update; The nonlinear updating method of the causal relationship mapping weight comprises extracting a key behavior factor from the comprehensive feedback index and constructing a weight updating vector; the key behavior factor is a deviation amplitude of an outdoor activity duration, an average deviation value of an eye distance, and a deviation mean of an illumination level extracted from the comprehensive feedback index; The weight updating vector comprises multiple dimensional components representing behavior execution deviation, including outdoor activity duration deviation, eye distance deviation, and environmental illumination deviation, each of which represents a deviation degree between a current behavior state of a user and a target behavior baseline; The myopia risk scoring model adjusts a weight of a causal path in a causal graph and a sensitivity coefficient in the scoring model according to a structure and a direction of the weight updating vector, so as to improve a judgment accuracy of intervention effectiveness of the model and enhance an adaptive capacity of the model to different user behavior reaction differences; The synchronous trigger condition of the myopia risk scoring model comprises that a feature vector of a current behavior mode of a user is out of a boundary range covered by the current myopia risk scoring model, a false risk reduction interval in the last three iterations has a misjudgment rate exceeding a threshold value, and a rebound event trigger frequency in a unit time is higher than a set upper limit of the frequency; After the pre-warning is released, verification is needed, and after the verification is passed, a new myopia risk scoring model version is automatically released and synchronized to all running instances, and a periodic verification and dynamic anomaly detection double mechanism is adopted, when a user behavior mode deviates from a coverage range of an existing myopia risk scoring model, a false risk reduction misjudgment rate rises, or a rebound event frequency is abnormal, a new myopia risk scoring model training and deployment process is automatically triggered.