A child myopia risk assessment method based on traditional Chinese medicine constitution analysis

Abstract

The present application belongs to the technical field of children myopia risk assessment, and discloses a children myopia risk assessment method based on traditional Chinese medicine constitution analysis. The method synchronously collects traditional Chinese medicine constitution objectification data, eye physiological data, basic information and environmental influence data in a non-contact and child-friendly manner, completes constitution typing after preprocessing combined with children physiological characteristics, realizes multi-dimensional feature fusion through attention mechanism, inputs the feature set into an interpretable artificial intelligence model verified by multi-center large sample, and outputs myopia risk level, factor contribution weight, myopia progression, high myopia, true and false myopia identification and other early warning information. The method dynamically retests at a fixed cycle, establishes a constitution-myopia risk time sequence archive, generates a personalized traditional Chinese medicine intervention scheme according to different constitutions, realizes constitution objective quantitative collection and multi-dimensional data deep fusion, and assesses the myopia risk in an interpretable and quantifiable manner, which can be deployed on mobile terminals and wearable devices.

Description

Technical Field

[0001] This invention belongs to the field of children's myopia risk assessment technology, specifically a method for children's myopia risk assessment based on traditional Chinese medicine constitution analysis. Background Technology

[0002] The occurrence and development of myopia in children is usually influenced by multiple factors, including axial length, refractive status, genetic factors, eye use behavior, outdoor activities, and sleep patterns. Current methods for assessing the risk of myopia in children mainly rely on ocular physiological test data and lifestyle data, and determine the risk of myopia through single collection or periodic screening. However, children are in a continuous growth and development stage, and their eye accommodation ability, the rate of depletion of farsighted reserve, and lifestyle all have significant variability. Relying solely on ocular indicators and external behavioral data is insufficient to fully reflect the impact of a child's internal state on the formation of myopia risk.

[0003] In the context of TCM constitution identification and children's visual health management, constitution imbalance is believed to affect children's eye nourishment, accommodative stability, and fatigue recovery ability. However, existing constitution identification methods mostly rely on questionnaires, parental descriptions, or physician experience judgment. The collected results are easily affected by subjective understanding, cooperation level, and differences in evaluation standards, making it difficult to form calculable and reproducible constitution representation data. Even when constitution information is used for myopia risk assessment, it is usually only used as an additional reference item. There is a lack of integration mechanism based on the collection quality, time changes, and related response relationships between constitution and other data such as refractive error, axial length, hyperopia reserve, and screen time. This makes the role path of constitution factors in risk formation unclear, and the assessment results difficult to interpret.

[0004] Furthermore, existing risk assessment methods mostly rely on single-time risk level outputs, failing to establish continuous records based on changes in children's physical condition, ocular physiology, and environmental behavior. They also lack mechanisms to verify assessment results using changes in refractive error, axial length, and hyperopic reserve within a retesting period. When a child's physical condition changes but ocular physiology indicators are not yet significantly abnormal, the system struggles to identify risk trends in a timely manner. When ocular physiology indicators change, it is difficult to trace their correlation with physical condition, eye use behavior, and environmental factors. Summary of the Invention

[0005] The purpose of this invention is to provide a method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis. Through objective constitution data collection, improved algorithm model, multi-dimensional feature fusion and dynamic tracking intervention, it achieves accurate assessment, interpretable early warning and personalized prevention and control of myopia risk in children. At the same time, it is adaptable to lightweight deployment, improving clinical applicability and scalability.

[0007] To achieve the above-mentioned objectives, this invention provides a method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis, comprising the following steps: Step S1: Multimodal objectification data acquisition: Using non-invasive, child-friendly data collection methods, the following four types of data were collected simultaneously to ensure the objectivity, comprehensiveness, and quantifiability of the data: Objective data on TCM constitution: Videos of children's faces are captured using a remote photoplethysmography (RPPG) acquisition device. Heart rate variability (HRV) signals are extracted, and time-domain features (SDNN, RMSSD, pNN50) and frequency-domain features (LF, HF, LF / HF ratio) are calculated to quantify the state of autonomic nervous system function and assist in determining constitution types such as Qi deficiency, Yang deficiency, and Yin deficiency. At the same time, AI image recognition algorithms are used to identify and quantify the children's tongue appearance (tongue body, tongue coating color and shape), complexion (skin color, gloss), and periorbital features (periorbital skin color, eye bag shape). Combined with age correction factors, the children's constitution type is accurately determined.

[0008] Ocular physiological data: Using equipment such as fully automated computer optometry, optical biometry, and accommodation sensitivity testing, five core indicators of children's refractive power, axial length, corneal curvature, accommodation sensitivity, and hyperopia reserve are collected simultaneously to comprehensively reflect the physiological state of the eyes.

[0009] Basic information: Collect the child's age, gender, height, weight, and the parents' myopia genetic information (whether the parents are myopic, the degree of myopia, and the age of onset of myopia) to establish a basic identity and genetic risk profile.

[0010] Environmental impact data: Through smart wearable devices (children's wristbands, smart eye patches) and parents' mobile apps, we continuously collected data on daily screen time (near-range and far-range screen time), outdoor activity time, total daily sleep time and sleep quality for 7-30 days. At the same time, we collected the frequency and amount of intake of key nutrients such as lutein, DHA and vitamin A in children's daily diet.

[0011] Step S2: Data Preprocessing The collected multimodal data undergoes standardization processing to eliminate dimensional differences, remove outliers, and fill in missing data. Specifically, this includes: For RPPG signals: wavelet transform algorithm is used to remove baseline drift, power frequency interference and motion artifacts, and extract effective HRV feature parameters; for tongue and facial color image data: histogram equalization and Gaussian filtering algorithm are used for image enhancement to improve the feature extraction accuracy of AI recognition model and reduce the influence of environmental factors such as lighting and angle.

[0012] For numerical data (ocular physiological data, HRV characteristics, dietary data, etc.): Z-score normalization is used for normalization, with the following formula:

[0013] in This is the original data. This represents the average data for children of the same age group. The standard deviation is used to eliminate dimensional differences between different indicators.

[0014] For outliers and missing values: outliers deviating from the normal range are removed using the 3σ principle; for a small number of missing data, the K-nearest neighbor (KNN) algorithm is used to complete the data, where the K value is between 5 and 10 to ensure data integrity.

[0015] Regarding children's personal information: data anonymization processing is adopted (such as encrypting and anonymizing names and ID numbers).

[0016] Step S3: Physique typing and multi-dimensional feature fusion: Improved Pediatric Constitution Classification: Based on preprocessed HRV characteristics, quantitative scores of tongue appearance / facial color, and age correction factors, a weighted scoring model is constructed, with the following formula:

[0017] in For the overall physical fitness score, The weighting coefficients are set to values ​​of 0.4, 0.35, and 0.25 respectively. The normalized values ​​of HRV features, facial features, and age correction factors were used to classify children's constitutions into five types: balanced constitution, qi deficiency constitution, yin deficiency constitution, yang deficiency constitution, and phlegm-dampness constitution. Compared with traditional classification methods, the classification accuracy is improved by more than 15%.

[0018] Multi-dimensional feature fusion: By combining the attention mechanism with the theory of viscera and qi in traditional Chinese medicine, a multi-dimensional feature fusion model is constructed. The model focuses on strengthening the correlation weight between physical characteristics and ocular physiological characteristics (with a weight ratio of no less than 30%). At the same time, it integrates basic information and environmental impact data to generate a myopia risk assessment feature set containing 5 primary features and 22 secondary features. This comprehensively covers the four dimensions of children's physical constitution, physiology, genetics, and environment, providing comprehensive feature support for subsequent assessments.

[0019] Step S4: Explainable Myopia Risk Assessment The preprocessed feature set is input into an interpretable artificial intelligence (XAI) evaluation model that has been validated by a large number of samples from multiple centers. This model is built based on an optimized random forest algorithm, and the specific steps are as follows: Model Training: Collect multi-dimensional data and myopia incidence information of more than 10,000 children aged 6-12 years to construct a training dataset, and use a grid search algorithm to optimize the number of decision trees in the random forest. Values ​​range from 100 to 300, maximum depth ( Hyperparameters such as 10-20 can be used to improve the model's fitting accuracy and generalization ability.

[0020] Risk level assessment: The model outputs the risk level of myopia in children, divided into three levels: low risk (risk value < 0.3), medium risk (0.3 ≤ risk value < 0.7), and high risk (risk value ≥ 0.7). It also outputs the risk contribution weight of each feature, among which the contribution weight of physical characteristics is not less than 30%, clarifying the specific impact of each factor on the risk of myopia.

[0021] Precise early warning function: The model can simultaneously output the predicted value of the rate of myopia progression in children and the early warning result of high myopia risk (a high risk is determined when the early warning probability is ≥80%), and distinguish between pseudomyopia and true myopia, providing clear and quantifiable basis for clinical intervention.

[0022] Step S5: Dynamic Tracking and Closed-Loop Intervention Dynamic tracking management: Every 3 months is an assessment cycle. Steps S1-S4 are repeated to establish a time series profile of children’s “physical condition-myopia risk”. The time series data is analyzed using a long short-term memory network (LSTM) model to dynamically update the myopia risk level and the contribution weight of each feature, and accurately capture the risk change trend in the process of children’s growth and development.

[0023] Personalized TCM intervention plan generation: Based on the body constitution classification results and myopia risk level, a targeted, gentle, and safe personalized intervention plan is generated, specifically including: Qi deficiency constitution: Spleen-strengthening and Qi-tonifying intervention program, including acupoint massage (Zusanli, Zhongwan, and Pishu acupoints, massage for 5-10 minutes each time, twice a day), dietary regulation (increase intake of spleen-strengthening foods such as yam, millet, and red dates, and reduce intake of raw and cold foods), and outdoor activity guidance (no less than 2 hours of outdoor activity per day, mainly light exercise such as jogging and rope skipping). Yin deficiency constitution: Liver and kidney tonifying intervention plan, including eye hot compress (apply a 40-42℃ hot towel to the eyes for 10-15 minutes each time, 1-2 times a day), dietary regulation (increase intake of liver and kidney tonifying foods such as wolfberry, mulberry, and black sesame, and reduce intake of spicy and fried foods), and work and rest guidance (ensure at least 9 hours of sleep every day and avoid staying up late). Yang deficiency constitution: Warming Yang and unblocking the meridians intervention plan, including ear acupressure (selecting liver, kidney and eye acupoints, pressing for 3-5 minutes each time, twice a day), dietary conditioning (consuming more warming Yang foods such as mutton, longan and ginger, and reducing the intake of cold foods), and exercise guidance (engaging in 15-20 minutes of outdoor activities under sunlight every day, combined with gentle exercises such as Tai Chi and Baduanjin). For those with a balanced constitution: the basic protective intervention plan includes maintaining a balanced diet and regular sleep schedule, strengthening outdoor activities and eye relaxation training to prevent myopia.

[0024] Multi-terminal deployment and interconnection: It supports the deployment of assessment models and intervention plans on terminal devices such as mobile apps, WeChat mini programs, children's smart bracelets, and smart eye patches to achieve data synchronization and sharing; at the same time, it can be connected with hospital electronic medical record systems and optometry center management systems to achieve collaborative management between the medical end and the home end.

[0025] The beneficial effects of this invention are as follows: 1. This invention unifies the collection and preprocessing of facial videos, tongue images, facial images, periorbital images, ocular physiological test data, basic information, and environmental behavior data of children to be assessed. It extracts remote photoplethysmography (PPG) signals from facial videos to obtain heart rate variability characteristics, and simultaneously extracts constitution image features from tongue, facial, and periorbital images. This achieves the effect of transforming traditional Chinese medicine constitution information, which relies on questionnaires, parental descriptions, or manual experience judgment, into calculable and repeatable constitution representation data for children. Since the constitution representation results no longer rely solely on subjective judgment but are formed jointly by heart rate variability characteristics and image features, it can improve the stability and usability of constitution data in children's myopia risk assessment, providing a clearer data foundation for subsequent risk assessment.

[0026] 2. This invention introduces age-appropriate and developmental-stage-specific correction parameters into the process of physical characterization of children. Based on the correlation strength between the physical characterization results and ocular physiological test data, it performs gated fusion of physical characterization results, ocular physiological test data, basic information, and environmental behavioral data. This achieves the effect of allowing health data from different sources to participate in risk assessment according to data quality, collection time, and the degree of correlation of changes. Compared to simply inputting physical category, eye indicators, and lifestyle behavioral data side-by-side, this invention reduces the interference of low-quality, outdated, or weakly correlated data on the assessment results. It makes the myopia risk assessment feature vector more accurately reflect the true relationship between the child's current physical condition, ocular physiological state, and environmental behavioral state, thereby improving the completeness and consistency of the risk assessment basis.

[0027] 3. This invention inputs the myopia risk assessment feature vector into a risk assessment model trained with labeled samples, and uses the changes in refractive error, axial length, and hyperopia reserve within a predetermined follow-up period as supervision labels to achieve the effect of making the model assessment results correspond to objective ocular physiological changes. Since the risk assessment model retains the feature path used to calculate the contribution while outputting the myopia risk value and risk level, it can give the contribution of various input features to the risk value. This allows the assessment results to not only indicate the level of risk, but also reflect the main data sources of risk formation, making it easier for families or medical institutions to review and understand the risk results, and reducing the problem of insufficient interpretation caused by simple black-box risk output.

[0028] 4. This invention, by repeatedly performing data collection, preprocessing, physical characterization, gating fusion, and risk assessment according to a predetermined retesting cycle, forms a time-series archive of the physical characterization results, ocular physiological test data, environmental behavior data, and myopia risk values ​​of the child to be assessed. This achieves the effect of continuously tracking and cross-validating the process of myopia risk changes in children. When the direction of risk change is consistent with the direction of ocular physiological change within the continuous retesting cycle, the credibility of the risk change indication can be enhanced. When there are inconsistent changes among the physical characterization results, environmental behavior data, or ocular physiological test data, it can also prompt for review and management, thereby avoiding static judgments based solely on a single test result, and making the assessment results traceable, retestable, and verifiable.

[0029] 5. This invention generates health management prompts based on physical constitution characteristics, risk levels, and environmental behavior data, achieving the effect of linking risk assessment results with subsequent management information. These health management prompts are used to advise on management matters such as outdoor activities, near-vision control, sleep adjustment, nutritional intake, and periorbital care, and are not directly used as disease diagnosis or treatment conclusions. Therefore, while maintaining the attributes of health information processing, it can achieve closed-loop processing from data collection, feature fusion, risk assessment, contribution interpretation to retesting management, thereby improving the continuity and applicability of long-term management of myopia risk in children. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the process for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis, as described in this invention. Figure 2 This is a flowchart of the data acquisition and preprocessing process of the present invention; Figure 3 This is a flowchart of the closed-loop process for constitution integration and risk assessment in this invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figures 1 to 3 As shown, this embodiment of the invention provides a method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis, including the following steps: Step S1: Multimodal Objective Data Acquisition: Using a non-contact, child-friendly method, objective data on children's TCM constitution, ocular physiology, basic information, and environmental impact data are collected simultaneously. The specific collection content and correlation algorithms are as follows: (1) Collection of objective data on TCM constitution: RPPG signal and HRV feature acquisition: Children's facial videos are acquired through mobile devices or wearable devices, and remote photoplethysmography (RPPG) signals are extracted. An adaptive filtering algorithm is used to remove signal noise, and a peak detection algorithm is used to identify the RR interval (the time interval between two adjacent pulse wave peaks) in the RPPG signal. Then, the time domain and frequency domain features of HRV are calculated to quantify the differences in autonomic nerve function among Qi deficiency, Yang deficiency, and Yin deficiency constitutions.

[0033] Formula for calculating HRV time-domain features: SDNN (Standard deviation of normal RR interval): ; in, The duration of the i-th normal RR interval (in ms). The average value of all normal RR intervals is given, and N is the number of normal RR intervals. A decrease in SDNN value corresponds to Qi deficiency and Yang deficiency, reflecting a weakening of autonomic nervous system regulation.

[0034] RMSSD (Root Mean Square of the Difference Between Adjacent RR Periods):

[0035] A lower RMSSD value corresponds to Yin deficiency constitution, reflecting weakened parasympathetic nerve function, and is associated with dry eyes and eye fatigue.

[0036] HRV frequency domain feature calculation: The RR interval sequence is converted into a frequency domain signal using Fast Fourier Transform (FFT), and the LF (low-frequency power, 0.04-0.15Hz), HF (high-frequency power, 0.15-0.4Hz), and LF / HF ratio are calculated using the following formulas: , ,

[0037] in, The power spectral density of the HRV signal is represented by the LF / HF ratio. An increased LF / HF ratio corresponds to a Yang deficiency constitution, reflecting sympathetic nerve excitation and poor blood circulation. A decreased ratio corresponds to a Yin deficiency constitution, reflecting parasympathetic nerve dominance.

[0038] Tongue appearance, facial color and periorbital features acquisition: Images of children's tongue and face are acquired through a mobile camera. Image enhancement algorithms (such as Retinex algorithm) are used to optimize image quality. Then, a CNN (convolutional neural network) model is used to extract features and identify features such as tongue color, tongue coating, teeth marks, cracks, facial color, periorbital bruising, and eye bags. A quantitative score (0-10 points) for each feature is output for physical constitution assessment.

[0039] The simplified structure of the CNN feature extraction model is as follows: The input image is processed by 3 convolutional layers (Conv2d), 2 pooling layers (MaxPool2d), and 2 fully connected layers (Linear), and the output feature vector is as follows: Convolutional layer calculation:

[0040] Pooling layer calculation:

[0041] Fully connected layer computation:

[0042] in, This is the weight matrix. For bias terms, The activation function is ReLU. For convolution operation, k is the pooling kernel size, s is the stride, and y is the feature quantization score vector.

[0043] (2) Collection of ocular physiological data, basic information and environmental impact data: ocular physiological data were collected through optometry equipment, including refractive power (D), axial length (AL, mm), corneal curvature (K, D), accommodative sensitivity (CS, cpm) and hyperopic reserve (HR, D); basic information included age (A), gender (G, male=1, female=0) and parental myopia inheritance information (H, both parents are myopic=2, one parent is myopic=1, neither is myopic=0); environmental impact data were obtained through user reporting and equipment monitoring, including daily screen time (T, h), outdoor activity time (O, h), sleep duration (S, h) and nutritional intake score (N, 0-10 points).

[0044] Step S2 Data Preprocessing: All multimodal data collected in Step S1 are standardized, denoised, outlier removed, and missing value imputed. Simultaneously, data containing children's personal information is anonymized (using a hash algorithm) to ensure data compliance and usability. (1) Standardization: The Z-score standardization algorithm is used to convert all quantitative data into standardized data with a mean of 0 and a standard deviation of 1, eliminating the influence of units. The formula is as follows:

[0045] Where x is the original data, This is the mean of this type of data. The standard deviation of this type of data. This is the standardized data.

[0046] (2) Noise reduction: For RPPG signals, an illumination anti-interference algorithm (adaptive brightness adjustment) and a motion artifact removal algorithm (Kalman filtering) are used, as shown in the following formulas: Kalman filter prediction equation:

[0047] Kalman filter update equation: , ,

[0048] in, Let A be the predicted value at time k, A be the state transition matrix, and B be the control matrix. This is the control variable at time k-1. For Kalman gain, Let H be the observation value at time k (RPPG signal), H be the observation matrix, R be the observation noise covariance, P be the covariance matrix, and I be the identity matrix.

[0049] (3) Outlier removal: The 3σ principle is used to remove outliers. Data range; missing value completion: The K-nearest neighbor (KNN) algorithm is used to complete missing values ​​based on the corresponding data of similar samples, with K set to 5.

[0050] Step S3: Physique typing and multi-dimensional feature fusion: (1) Improved TCM constitution classification algorithm: Combining the characteristics of children's "immature Yin and Yang" and their easily changeable constitution, age and growth and development stage correction factors are introduced. Based on the preprocessed TCM constitution objective data (HRV features, tongue appearance and facial color feature quantitative scores), a weighted summation algorithm is used to complete the constitution classification. The formula is as follows: Physical fitness score calculation:

[0051] in, This is a comprehensive score for a specific constitution (Qi deficiency / Yin deficiency / Yang deficiency / balanced constitution), where m is the number of constitution-related characteristics. The weight of the i-th feature is determined by the analytic hierarchy process (AHP): HRV feature weight 0.4, tongue appearance feature weight 0.4, and facial color and periorbital feature weight 0.2. Let i be the standardized value of the i-th feature. The age correction factor is α=0.1 for 3-6 years old and α=0.05 for 7-12 years old, where A is the child's age. The constitution with the highest comprehensive score is taken as the child's final constitution classification result (balanced constitution, qi deficiency constitution, yin deficiency constitution, yang deficiency constitution).

[0052] (2) Multi-dimensional feature fusion: Using the attention mechanism and combining the theory of viscera and qi and blood in traditional Chinese medicine, the intrinsic relationship between physical characteristics and ocular physiological characteristics is strengthened, and a multi-dimensional myopia risk assessment feature set is constructed. The formula is as follows: Attention weight calculation:

[0053] Fusion feature vectors:

[0054] in, The attention score for the i-th feature (calculated by the fully connected layer) is denoted by n, which represents the total number of features (4 physical features, 5 ocular physiological features, 3 basic information features, and 4 environmental impact data features). Let j be the attention weight for the j-th feature. Let F be the standardized vector of the j-th feature, and F be the fused multi-dimensional evaluation feature set.

[0055] Step S4: Explainable Myopia Risk Assessment: An Explainable Artificial Intelligence (XAI) assessment model, optimized based on the random forest algorithm, is used to process the fused feature set, outputting the myopia risk level and the risk contribution weight of each influencing factor. Simultaneously, a myopia-related risk warning is provided. The specific algorithm and formula are as follows: (1) Random Forest Model Construction: It consists of M decision trees. Each decision tree obtains training samples through bootstrap sampling. The Gini coefficient is used as the feature splitting criterion. The formula for calculating the Gini coefficient is:

[0056] Where C represents the number of categories (high / medium / low risk of myopia). The percentage of class c samples in the current node; the output of each decision tree is the risk category of the sample, and the final output of the model is determined by voting (the category with the most votes is the final risk level).

[0057] (2) Risk contribution weight calculation: The contribution weight of each feature to the myopia risk assessment is calculated using the feature importance score (Gini importance), as shown in the following formula: ,

[0058] in, The importance score for the j-th feature, where T is the number of decision trees. Let Gini coefficient be the amount by which the j-th feature in the t-th decision tree is split. The risk contribution weight for the j-th feature requires that the total contribution of physical characteristics be no less than 30%.

[0059] (3) Early warning of myopia-related risks: Based on the fusion feature set, the rate of myopia progression is predicted by a linear regression model, as shown in the following formula:

[0060] Where V represents the rate of myopia progression (D / year). For the intercept term, Let be the regression coefficient of the j-th feature. The random error term is used. Combining axial length and refractive error data, a high myopia risk threshold is set (axial length ≥26mm is considered high risk). A logistic regression model is then used to distinguish between pseudomyopia and true myopia, as shown in the following formula:

[0061] Where P is the probability of true myopia. For the intercept term, is the regression coefficient of the j-th feature; when P≥0.6, it is judged as true myopia, and when P<0.6, it is judged as pseudomyopia.

[0062] The XAI evaluation model was trained on multi-center, large-sample children's data (sample size ≥ 1000 cases) and externally validated, with a validation set accuracy of ≥ 85%, meeting the compliance requirements for medical AI.

[0063] Step S5: Dynamic Tracking and Closed-Loop Intervention Repeat steps S1 to S4 regularly (every 3 months) to establish a time-series profile of children's physical condition and myopia risk. Use a time-series predictive model (LSTM) to dynamically update the myopia risk assessment results. The core formula of the LSTM model is as follows: Forgotten Gate:

[0064] Input Gate: ,

[0065] Cell state:

[0066] Output gate: ,

[0067] in, , , The outputs of the forget gate, input gate, and output gate are respectively... The cell state at time t. The output of the hidden layer at time t. Input features at time t, This is the weight matrix. For bias terms, This is element-wise multiplication. is the sigmoid activation function, and tanh is the hyperbolic tangent activation function.

[0068] Based on children's physical constitution classification and myopia risk level, targeted personalized TCM intervention suggestions are generated, forming a complete closed loop of "collection, assessment, tracking and intervention". The method can be deployed on mobile terminals (mobile APP, mini program) or wearable devices (children's bracelet, smart eye patch), and supports interconnection with children's vision records and TCM physical constitution records to achieve real-time collection, rapid assessment and risk warning.

[0069] Data authorization and multi-source health data collection Before conducting the assessment, the system first establishes an anonymous assessment number for the child to be assessed. The child's name, ID number, contact information, home address, and guardian information are not directly involved in model training and risk assessment. When it is necessary to connect with medical records, the identity mapping is completed by the authorized interface. During the risk assessment process, the anonymized number, age group, gender, and parents' myopia information are used.

[0070] The system acquires multi-source health data of children to be evaluated that have been authorized for collection. The multi-source health data includes facial videos, tongue images, facial images, periorbital images, eye physiological test data, basic information, and environmental behavior data.

[0071] Facial video is used to extract remote photoplethysmography (PPG) signals. During acquisition, the camera is pointed towards the child's face, and the acquisition area includes at least the skin area around the forehead, cheeks, and nose. The system prompts the child to maintain a natural sitting posture and natural expression on the acquisition interface. The acquisition environment only needs to meet basic lighting requirements and does not require a special medical testing environment. During the acquisition process, the system judges in real time whether the face is complete, whether the lighting is too dim, and whether the head is moving too much. If the acquisition conditions are not met, the system prompts to re-acquire or extend the acquisition time.

[0072] The tongue image is used to extract the characteristics of tongue color, tongue coating color, tongue shape, and tongue coating thickness. During the acquisition, the system prompts the child to naturally extend their tongue, and the camera captures a frontal image of the tongue. The system locates the tongue area. If the image is severely blurred, overexposed, occluded by the tongue, or the tongue area is insufficient, the image is marked as a low-quality image and will not be included in the subsequent calculation of physical constitution image features.

[0073] Facial images are used to extract facial color features, while periorbital images are used to extract periorbital skin color features and periorbital contour features. Facial and periorbital images can be extracted from facial videos or obtained through independent photography. When using video extraction, the system selects images from frames with high signal quality and stable facial poses. When using photography, the system assesses facial integrity, periorbital clarity, and illumination uniformity.

[0074] The ocular physiological test data includes refractive error, axial length, corneal curvature, accommodative sensitivity, and hyperopic reserve. Refractive error can be obtained from computerized refractometers, optometric screening instruments, or medical institution examination records; axial length can be obtained from optical biometers; corneal curvature, accommodative sensitivity, and hyperopic reserve can be obtained from optometric examination equipment or authorized access to medical records. Data from different sources are accompanied by the collection time and source identification to facilitate subsequent assessment of data timeliness and reliability.

[0075] Basic information includes age, gender, and parental myopia information. Parental myopia information can include whether the father and mother are myopic and the range of their myopia levels. Environmental and behavioral data includes duration of close-range eye use, outdoor activity time, sleep duration, and nutritional intake records. Duration of close-range eye use can be obtained from smart eye patches, mobile device usage records, or parental input. Outdoor activity time can be obtained from light sensor data and exercise data from wearable devices, or parental records. Sleep duration can be obtained from children's wristbands or parental records. Nutritional intake records can be based on daily intake frequency or intake categories, and are not required to be accurate to a medical nutrition prescription.

[0076] Through the above-described data collection method, this embodiment does not use a single eye physiological test result as the basis for evaluation, nor does it rely solely on questionnaires to obtain the constitution type. Instead, it unifies facial video physiological signals, tongue images, facial images, periorbital images, eye physiological test data, basic information, and environmental behavior data into a computable data foundation.

[0077] Facial video preprocessing and heart rate variability feature extraction The system performs facial region tracking on facial videos. Specifically, the system identifies facial regions frame by frame and determines multiple regions of interest around the forehead, left cheek, right cheek, and nose. For each frame, the system removes areas that are obviously obscured by hair, glasses, masks, or shadows, and retains only areas with stable skin color that are continuously trackable.

[0078] The system performs temporal separation of skin color channels in the preserved regions of interest. Specifically, the system extracts the sequence of color channel changes over time from each region of interest and performs detrending processing and bandpass filtering on the color change sequence of each region to obtain a remote photoplethysmography (PPG) signal that reflects subtle changes in blood volume. The signals obtained from different skin regions can be weighted and synthesized according to signal stability. Regions with higher signal stability have larger weights, while regions that are more affected by occlusion, strong reflection, or local motion have smaller weights.

[0079] The system segments valid heartbeat segments based on signal quality evaluation results. The signal quality evaluation results are jointly determined by face tracking stability, illumination fluctuation amplitude, head displacement amplitude, and heartbeat cycle continuity. Face tracking stability reflects the degree of continuous trackability of the face region in the video; illumination fluctuation amplitude reflects the impact of changes in ambient brightness on the signal during acquisition; head displacement amplitude reflects the impact of children's movements on the signal; and heartbeat cycle continuity reflects whether the extracted peaks and troughs satisfy the continuity of physiological rhythm.

[0080] When a video segment has low face tracking stability, excessive lighting fluctuations, excessive head displacement, or discontinuous heartbeat cycles, the system marks the video segment as invalid and does not use it for heart rate variability feature extraction. For valid heartbeat segments, the system identifies adjacent heartbeat peaks, obtains the heartbeat interval sequence, and calculates heart rate variability features based on the heartbeat interval sequence.

[0081] Heart rate variability features include multiple features such as mean heart rate interval, standard deviation of heart rate interval, root mean square of the difference between adjacent heart rate intervals, proportion of the difference between adjacent heart rate intervals exceeding a preset time difference, low-frequency power, high-frequency power, and low-frequency to high-frequency ratio. These features are used to characterize the state of children's autonomic nervous activity. The system does not directly determine the constitution type based on a single heart rate variability feature, but rather uses it as an input feature of the constitution representation vector, which participates in the calculation together with subsequent constitution image features.

[0082] Through this processing, this embodiment transforms the originally difficult-to-quantify constitution-related states into repeatable physiological signal characteristics, avoiding the determination of constitution tendencies based solely on subjective questionnaires or human experience.

[0083] Tongue image, face image and periorbital image processing The system assesses the acquisition quality of tongue, face, and periorbital images. The acquisition quality assessment includes sharpness assessment, exposure assessment, target area integrity assessment, and pose deviation assessment. Images with sharpness below a preset sharpness threshold, abnormal exposure, missing target areas, or excessive pose deviation are not included in the feature extraction process. Instead, the system prompts for re-acquisition or marks them as low-reliability data.

[0084] For a qualified tongue image, the system performs tongue body region segmentation to obtain the tongue body region and the tongue coating region. The tongue body region is used to extract tongue body color features and tongue body morphology features; the tongue coating region is used to extract tongue coating color features and tongue coating thickness features. Tongue body color features can include color distribution, color uniformity, and local color difference; tongue body morphology features can include tongue body width, edge shape, and tongue surface texture; tongue coating color features can include tongue coating brightness, hue, and coverage area; tongue coating thickness features can be determined by comprehensively considering the texture density, brightness variation, and coverage ratio of the tongue coating region.

[0085] For facial images, the system first locates the facial region, then excludes areas such as glasses, hair, strong reflections, and shadows, and extracts facial color features, including facial hue, brightness, saturation, gloss, and local color uniformity.

[0086] For images around the eyes, the system locates the upper eyelid, lower eyelid, under-eye bag area, and periorbital area, and extracts skin color features and contour features around the eyes. Skin color features around the eyes include brightness, hue, and color difference with other areas of the face; contour features around the eyes include the lower eyelid contour, under-eye bag area morphology, and texture changes around the eyes.

[0087] The system converts tongue, face, and periorbital images into corresponding image feature vectors and retains the acquisition quality score for each type of image. The acquisition quality score is used for subsequent gating fusion and is not directly used as a risk judgment conclusion. Therefore, even if some images are of low quality, the entire evaluation process will not be simply discarded. Instead, the impact on the final evaluation result is reduced by ensuring data reliability.

[0088] Processing of eye physiological test data, basic information and environmental behavior data The system performs source consistency and time consistency processing on eye physiological test data. For the same indicator with multiple source records, the data with the most recent collection time and higher source reliability is given priority. If the differences between multiple source data exceed the preset allowable range, the indicator is marked as pending review and will not be directly used as high-weight data in the fusion.

[0089] Refractive error data can be converted into equivalent spherical values. Axial length, corneal curvature, accommodative sensitivity, and hyperopic reserve are normalized according to the data range of children of the same age. Normalization does not change the original record. The system saves the original value, normalized value, acquisition time, and source identifier at the same time.

[0090] The age in the basic information is used to determine the age group correction parameters, the gender is used to determine the reference range for the same age group, and the parents' myopia information is used to form genetic risk characteristics. The parents' myopia information is only used for risk assessment and is not the sole basis for judging the risk of myopia in children.

[0091] Environmental behavior data is compiled according to a predetermined statistical period. The duration of close-range eye use, outdoor activity time, sleep duration and nutritional intake can be recorded daily and then statistical characteristics can be obtained by week or month. The statistical characteristics include average value, degree of fluctuation and the proportion of days that meet the standards. For data entered by parents, the system records the entry time and the number of consecutive recording days. When the number of consecutive recording days is insufficient or the proportion of missing days is high, the reliability of this type of data is reduced accordingly.

[0092] The system standardizes the units of numerical data, marks outliers, and completes missing items. Outlier marking can be determined based on the reference range for the same age group, the normal output range of the device, and the individual's historical variation range. For missing items, the system prioritizes using the same child's historical records for completion. When the same child's historical records are insufficient, statistical values ​​of reference data for the same age group and gender can be used for completion. The completed data is accompanied by a completion label. During subsequent gating and fusion, the completed data is assigned a lower reliability to avoid the completed data having an excessive impact on the risk results.

[0093] Generation of children's physical fitness characteristics The system generates a child's physical constitution representation vector based on heart rate variability characteristics and physical constitution image characteristics. The child's physical constitution representation vector includes a heart rate variability sub-vector, a tongue image sub-vector, a facial image sub-vector, and a periorbital image sub-vector. Each sub-vector does not form a final physical constitution conclusion on its own, but is incorporated into the physical constitution representation model for comprehensive calculation.

[0094] The system adjusts the effective weights of each feature in the physical fitness representation vector according to the child's age group and developmental stage. The age group can be divided according to the preschool, lower grades, middle and upper grades, and pre- and post-pubertal stages, or according to the age grouping rules for children adopted by medical institutions or screening institutions. The developmental stage can be determined based on age, height and weight change trends, parent records, and developmental status records in medical files. The age group correction parameter and the developmental stage correction parameter are used to address the problem of unstable feature expression caused by changes in children's physical fitness status with growth and development.

[0095] In this embodiment, the results of children's physical constitution characterization include physical constitution tendency values ​​for balanced tendency, qi deficiency tendency, yin deficiency tendency, yang deficiency tendency, and phlegm-dampness tendency. Each physical constitution tendency value can be represented by a normalized score or a tendency level. The physical constitution tendency value is jointly determined by heart rate variability features, physical constitution image features, age group correction parameters, and developmental stage correction parameters. The system can use a rule-based scoring model, an ensemble tree model, or a lightweight neural network model to form the physical constitution tendency value. Regardless of the model used, the model input includes heart rate variability features and physical constitution image features after quality control, and the output is always the physical constitution tendency value, rather than a simple physical constitution category from a manual questionnaire.

[0096] In one implementation, the system first calculates the support of heart rate variability features, tongue image features, facial image features, and periorbital image features for each constitution tendency, respectively. Then, it adjusts the effective weight of each support based on age group correction parameters and developmental stage correction parameters to obtain the final constitution tendency value. If the quality of a certain type of data is low, the effective weight of that type of data in the constitution tendency value is reduced, rather than directly deleting the corresponding constitution representation result.

[0097] With this setting, this embodiment differs from simple tongue image recognition or simple constitution questionnaire recognition. Its constitution representation results are derived from the joint calculation of facial video physiological signals, tongue images, facial images, periorbital images, and child development corrections, which can form a repeatable and comparable data-driven constitution representation.

[0098] Gated fusion of myopia risk assessment feature vectors The system performs gating fusion of children's physical fitness results, eye physiological test data, basic information and environmental behavior data based on data availability, collection time interval and correlation response strength between children's physical fitness results and eye physiological test data to obtain a myopia risk assessment feature vector.

[0099] Data availability is used to characterize whether a certain type of data is complete, meets the acquisition quality requirements, and has completion markers. For facial video data, data availability is determined by the proportion of effective heartbeat segments, signal quality evaluation results, and heartbeat cycle continuity. For image data, data availability is determined by sharpness, exposure, target area completeness, and pose deviation. For eye physiological detection data, data availability is determined by source reliability, acquisition time, and consistency of the same indicator across multiple sources. For environmental behavior data, data availability is determined by record continuity, missing proportion, and data entry source.

[0100] The data collection time interval is used to characterize the freshness of the data. The closer the data is to the current assessment time, the higher its timeliness weight; the farther the data is from the current assessment time, the lower its timeliness weight. For indicators such as axial length, refractive error and hyperopic reserve, which have clinical significance over time, the system prioritizes the most recent valid data while retaining historical trends.

[0101] The correlation response strength is used to characterize whether there is a synchronous change relationship between the results of children's physical fitness assessment and ocular physiological test data in the historical retesting period. Specifically, in at least two retesting periods, the system calculates the degree of consistency between the direction of change of physical fitness tendency value and the direction of change of axial length, refractive error and hyperopia reserve. When a change of physical fitness tendency value and ocular physiological change have a high degree of consistency in multiple retesting periods, the fusion weight of the physical fitness tendency value in the myopia risk assessment feature vector is increased; when the consistency is low or the historical retesting data is insufficient, the physical fitness tendency value is retained as a basic physical fitness characteristic, but is not given an excessively high weight.

[0102] Gated fusion is not simply a matter of splicing together various features. The system first calculates the reliability of physical constitution characteristics, eye physiological test data, basic information and environmental behavior data separately. Then, it determines the fusion weight of various data based on timeliness and relevant response intensity. The final myopia risk assessment feature vector includes physical tendency features, eye physiological state features, genetic basis features, environmental behavior features, data reliability features and historical change features.

[0103] By using gated fusion, the impact of low-quality, outdated, or weakly correlated data with the current state of children on the assessment results is reduced; the role of high-quality, timely data that responds to changes in ocular physiology is enhanced. This technology can be distinguished from conventional schemes that only input ocular physiology data, genetic information, and eye-use behavior data into the prediction model, as well as schemes that only output physical constitution identification results without participating in the time-series assessment of myopia risk.

[0104] Risk assessment model training and risk output The system inputs the myopia risk assessment feature vector into the risk assessment model trained with labeled samples, and outputs the myopia risk value, risk level, and the contribution of various input features to the myopia risk value.

[0105] The risk assessment model can adopt an interpretable ensemble tree model. In this embodiment, the interpretable ensemble tree model can be a random forest model or other ensemble tree models with tree path output capabilities. The reason for adopting an ensemble tree model is that it can retain the split nodes, split directions and split contributions of each feature in multiple tree paths, which facilitates the subsequent calculation of contribution, rather than just outputting the unexplainable risk level.

[0106] The training samples are derived from authorized historical children's health data and follow-up data. Each training sample includes a myopia risk assessment feature vector at a certain assessment time point, as well as the results of the child's ocular physiological changes during the subsequent predetermined follow-up period. The supervision label is jointly determined by the changes in refractive error, axial length, and hyperopia reserve during the predetermined follow-up period. The changes in refractive error, axial length, and hyperopia reserve are all derived from objective test records, and the doctor's subjective impression is not used as the sole label.

[0107] Supervisory labels can be constructed based on risk values, or based on low-risk, medium-risk, and high-risk levels. In practice, the label division rules can be determined based on the distribution of follow-up data of children of the same age, the review rules adopted by optometry screening institutions, or the objective change thresholds in the historical data of medical institutions.

[0108] During model training, the system divides the training samples into training and validation sets and records the usage of various features in the tree path during training. After the model training is completed, the system retains the feature path used to calculate the contribution. The feature path includes the splitting node of a certain input feature in the tree model, the splitting direction, the change in risk score before and after the split, and the feature category to which the splitting node belongs.

[0109] The model outputs a myopia risk value, which can be represented by a continuous value between zero and one. The higher the value, the higher the probability of an increase in the risk of myopia or an increase in the risk of myopia progression within the predetermined follow-up period. The system determines the risk level based on the comparison between the myopia risk value and the low-risk threshold and the high-risk threshold. The low-risk threshold and the high-risk threshold can be determined based on the validation set distribution, the screening rules of medical institutions, or historical follow-up data of the same age group, and can be adjusted as the model version is updated.

[0110] The model simultaneously outputs the contribution of various input features to the myopia risk value. The contribution can be summarized according to physical characteristics, ocular physiology, basic information, environmental behavior, and historical changes. Specifically, the system counts the feature split nodes that the child sample to be evaluated passes through in multiple tree paths, calculates the contribution of each split node to the change in risk value, and then summarizes them according to feature category to obtain the contribution of various input features. The contribution is used to explain the main source of the risk result and is not used as an independent risk judgment conclusion.

[0111] For example, when a child has a high risk of myopia, the system can show that the increased risk is mainly due to the combined effect of changes in axial length, decreased hyperopic reserve, excessive near-vision time, and an increased predisposition to a certain physical condition. In this case, the system does not directly determine that the child has developed a specific disease, but rather suggests that the child needs further optometric review or health management.

[0112] Through the above training and output methods, the model in this embodiment does not simply use existing ocular physiological data to predict myopia risk, nor does it use the results of TCM constitution identification as a simple additional label. Instead, it constructs a verifiable training and evaluation chain by combining constitution characteristics, ocular physiology, environmental behavior, and follow-up labels.

[0113] Time-series archive creation and retesting verification The system repeatedly performs multi-source health data collection, data preprocessing, children's physical fitness characterization, gating fusion, and risk assessment according to the predetermined retesting cycle, forming a time-series archive of the physical fitness characterization results, eye physiological test data, environmental behavior data, and myopia risk values ​​of the children to be assessed. The predetermined retesting cycle can be set by medical institutions, screening institutions, or parents, and can be executed monthly, quarterly, or by semester.

[0114] The time-series archive includes at least the collection time, physical predisposition value, eye physiological test data, environmental behavior statistics, data reliability, myopia risk value, risk level and contribution of each assessment. The system retains the original data source identifier and desensitized feature data for each assessment, which facilitates subsequent tracing of the reasons for risk changes.

[0115] The system generates a myopia risk change prompt based on the consistency between the direction of risk change and the direction of eye physiological change within a continuous retesting period. The direction of risk change can be an increase in risk value, a decrease in risk value, or a basically stable risk value. The direction of eye physiological change can be determined by whether the refractive error changes towards myopia, whether the axial length of the eye increases, and whether the hyperopic reserve decreases.

[0116] When the risk value increases during a continuous retesting period, and at the same time the axial length increases, the refractive power changes towards myopia, or the hyperopic reserve decreases, the system generates a high-confidence warning of increased risk. When the risk value increases but the ocular physiological test data does not show the same change, the system generates a review prompt, suggesting checking the collection quality, the completeness of environmental and behavioral records, or recent eye use status. When the risk value decreases and the changes in ocular physiological test data tend to stabilize, the system generates a prompt to continue observation.

[0117] By using time-series archives and retest verification, this embodiment avoids outputting static risk levels based solely on single test results. Its technical effect is that changes in physical condition, eye physiology, and environmental behavior can be recorded and compared over a continuous time dimension, thereby improving the traceability and verifiability of risk alerts.

[0118] Health management prompt generation The system generates health management tips based on children's physical characteristics, risk level, contribution, and environmental behavior data. These tips include tips for outdoor activities, near-vision control, sleep adjustment, nutritional intake, and eye care.

[0119] When the risk contribution mainly comes from excessive close-range eye use, the system generates prompts to reduce continuous close-range eye use and increase rest frequency. When the risk contribution mainly comes from insufficient outdoor activity time, the system generates prompts to increase outdoor activity time. When the risk contribution mainly comes from insufficient sleep duration or large fluctuations in work and rest, the system generates sleep adjustment prompts. When the proportion of physical constitution characteristics in the risk contribution is high, the system can combine physical constitution tendency values ​​to generate corresponding prompts for eye care, work and rest and diet management.

[0120] The health management tips are general health management suggestions and do not replace the results of medical institution examinations, nor should they be used as disease diagnosis conclusions or treatment plans. For children with high risk levels, increasing risk with repeated tests, or significant changes in eye physiological test data, the system will prioritize generating a review prompt, suggesting that they go to a professional institution for an optometry examination.

[0121] Through the above settings, the system can form a complete closed loop from data collection, feature extraction, physical constitution characterization, gating fusion, risk assessment, contribution interpretation, time series retesting to health management prompts, so that each prompt can be traced back to the corresponding data source and model output results.

[0122] System Implementation This invention can also be implemented through a children's myopia risk assessment system based on traditional Chinese medicine constitution analysis. This system includes a data acquisition module, a preprocessing module, a constitution characterization module, a gating fusion module, a risk assessment module, and a time-series management module.

[0123] The data acquisition module is used to acquire multi-source health data of children to be evaluated that have been authorized for collection. The data acquisition module can connect to cameras, smart bracelets, smart eye patches, vision testing equipment, parent-side applications, and medical record interfaces, and record the collection time, source identifier, and authorization status for each type of data.

[0124] The preprocessing module is used to extract remote photoplethysmography (PPG) signals from facial videos and obtain heart rate variability features. It is also used to obtain body image features from tongue images, facial images, and periorbital images. The preprocessing module is also used to complete the dimensional unification, outlier marking, and missing item completion of eye physiological test data, basic information, and environmental behavior data.

[0125] The constitution characterization module is used to generate children's constitution characterization results based on heart rate variability characteristics, constitution image characteristics, age group correction parameters, and developmental stage correction parameters. The children's constitution characterization results include constitution tendency values ​​for balanced tendency, qi deficiency tendency, yin deficiency tendency, yang deficiency tendency, and phlegm-dampness tendency.

[0126] The gated fusion module is used to generate a myopia risk assessment feature vector based on data availability, acquisition time interval, and physical condition and ocular physiological response intensity. The gated fusion module sets fusion weights for data from different sources, so that data with higher quality, stronger timeliness, and correlation with ocular physiological changes can play a greater role in the assessment.

[0127] The risk assessment module is used to output the myopia risk value, risk level, and contribution of various input features to the myopia risk value based on the myopia risk assessment feature vector. The risk assessment module adopts an interpretable ensemble tree model with tree path output capability, which can retain the feature path used to calculate the contribution.

[0128] The time series management module is used to generate time series archives of physical fitness results, eye physiological test data, environmental behavior data and myopia risk values ​​according to a predetermined retest cycle, and to generate myopia risk change prompts and health management prompts. The time series management module is also used to save the time, data reliability, risk level and contribution of each assessment, so that the family and medical ends can view the risk change trend.

[0129] Technical distinction from potential comparison documents The technical solution of this embodiment differs from solutions that predict myopia risk based solely on refractive error, axial length, parental myopia information, and eye-use environment data. Such solutions typically input ocular physiological data and environmental data into the prediction model, focusing on outputting myopia risk results. In contrast, this embodiment further converts heart rate variability features, tongue surface image features, facial image features, and periocular image features extracted from non-contact facial videos into children's physical characteristics, and modifies these characteristics based on age and developmental stage, enabling physical factors to participate in myopia risk assessment in a calculable and reproducible form.

[0130] The technical solution in this embodiment also differs from solutions that perform TCM constitution identification or tongue image identification alone. Such solutions typically output constitution category or health status prompts. However, this embodiment does not stop at the constitution identification result itself, but rather performs gating fusion of children's constitution characterization results with eye physiological test data, basic information and environmental behavior data. Based on the changes in refractive error, axial length and hyperopia reserve during the follow-up period, a risk assessment model is trained to establish a verifiable correspondence between constitution characterization and myopia risk.

[0131] The technical solution of this embodiment also differs from the risk assessment method of simple multi-feature splicing. The simple splicing method does not consider data quality, collection time, physical condition, and eye physiological response intensity, which can easily cause low-quality data, expired data, or data with weak correlation to interfere with the assessment results. This embodiment determines the fusion weight by data availability, collection time interval, and related response intensity, which can make the myopia risk assessment feature vector more consistent with the child's current state.

[0132] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0133] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis, characterized in that, The method is executed by a health information processing terminal or server, and includes the following steps: S1. Obtain authorized multi-source health data of the child to be evaluated, including facial video, tongue image, facial image, periorbital image, eye physiological test data, basic information and environmental behavior data; S2. Perform face region tracking and skin color channel temporal separation on the facial video, extract remote photoplethysmography (PPG) signals, and segment effective heartbeat segments based on signal quality evaluation results to obtain heart rate variability features from the effective heartbeat segments; perform acquisition quality judgment, image correction, and target region segmentation on the tongue image, facial image, and periorbital image to obtain physical fitness image features; perform dimensional unification, outlier marking, and missing item completion on the eye physiological detection data, basic information, and environmental behavior data. S3. Generate a child's physical fitness representation vector based on the heart rate variability features and physical fitness image features, and adjust the effective weights of each feature in the child's physical fitness representation vector according to the child's age group and developmental stage to obtain the child's physical fitness representation result. S4. Based on data availability, collection time interval, and the correlation response strength between children's physical fitness assessment results and eye physiological test data, gating fusion is performed on children's physical fitness assessment results, eye physiological test data, basic information, and environmental behavior data to obtain a myopia risk assessment feature vector. S5. Input the myopia risk assessment feature vector into the risk assessment model trained with labeled samples, and output the myopia risk value, risk level, and contribution of various input features to the myopia risk value; wherein, the training of the risk assessment model uses the change in refractive power, the change in axial length, and the change in hyperopia reserve within a predetermined follow-up period as supervision labels, and the risk assessment model retains the feature path used to calculate the contribution when outputting the myopia risk value; S6. Repeat steps S1 to S5 according to the predetermined retesting cycle to form a time-series archive of the physical characteristics, eye physiological test data, environmental behavior data and myopia risk value of the child to be evaluated. Based on the consistency between the direction of risk change and the direction of eye physiological change within the continuous retesting cycle, generate myopia risk change prompts and health management prompts.

2. The method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 1, characterized in that, In step S2, the signal quality evaluation result is determined based on the face tracking stability, illumination fluctuation amplitude, head displacement amplitude, and heart rate cycle continuity in the facial video; when the signal quality evaluation result of any video segment is lower than the preset quality threshold, the video segment will not be used for heart rate variability feature extraction.

3. The method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 2, characterized in that, In step S1, the heart rate variability features include multiple items from the following: mean heart interval, standard deviation of heart interval, root mean square of the difference between adjacent heart intervals, proportion of the difference between adjacent heart intervals exceeding a preset time difference, low-frequency power, high-frequency power, and low-frequency to high-frequency ratio.

4. The method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 3, characterized in that, In step S2, the physical constitution image features include tongue color features, tongue coating color features, tongue shape features, tongue coating thickness features, facial color features, skin color around the eyes features, and eye contour features; wherein, after segmenting the target region, the tongue image, facial image, and eye periorbital image generate corresponding image feature vectors, which are then input into the physical constitution representation model to form the child's physical constitution representation vector.

5. The method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 4, characterized in that, In step S3, the eye physiological test data includes refractive error, axial length, corneal curvature, accommodative sensitivity, and hyperopic reserve; the basic information includes age, gender, and parental myopia information; and the environmental behavior data includes near-vision time, outdoor activity time, sleep duration, and nutritional intake records.

6. The method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 5, characterized in that, In step S3, the results of the child's physical constitution characterization include physical constitution tendency values ​​for balanced tendency, qi deficiency tendency, yin deficiency tendency, yang deficiency tendency, and phlegm-dampness tendency; the physical constitution tendency values ​​are jointly determined by heart rate variability characteristics, physical constitution image characteristics, age group correction parameters, and developmental stage correction parameters.

7. The method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 6, characterized in that, In step S4, the gating fusion includes: The reliability of each data category is determined based on its missing percentage, collection quality, and the interval between the collection time and the current evaluation time. The intensity of the physical fitness-related eye physiological response was determined based on the degree of synchronous change between the results of children's physical fitness assessment and the data of eye physiological testing during the historical retesting period. Based on the reliability of the data and the intensity of the physical and ocular physiological related responses, the fusion weights of children's physical characteristics, ocular physiological test data, basic information and environmental behavior data in the myopia risk assessment feature vector are adjusted.

8. The method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 7, characterized in that, In step S5, the risk assessment model is an interpretable ensemble tree model, and the contribution is obtained by statistically analyzing the split contribution of various input features to the myopia risk value in multiple tree paths; the risk level is determined based on the comparison results between the myopia risk value and the low-risk threshold and the high-risk threshold.

9. A method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis according to claim 8, characterized in that, In step S6, the health management prompts include outdoor activity prompts, near-field eye use control prompts, sleep adjustment prompts, nutritional intake prompts, and periorbital care prompts. The health management prompts are generated based on the child's physical characteristics, risk level, and environmental behavior data, and are used to prompt families or medical institutions to conduct review and management, but are not used as disease diagnosis or treatment conclusions.

10. A child myopia risk assessment system based on traditional Chinese medicine constitution analysis, characterized in that: The system, applicable to the method for assessing the risk of myopia in children based on traditional Chinese medicine constitution analysis as described in any one of claims 1 to 9, comprises: The system includes a data acquisition module, a preprocessing module, a physical fitness characterization module, a gating fusion module, a risk assessment module, and a time series management module. The data acquisition module is used to acquire multi-source health data of children to be evaluated that have been authorized for collection; The preprocessing module is used to extract remote photoplethysmography (PPG) signals from facial videos and obtain heart rate variability features, and is also used to obtain body image features from tongue images, facial images and periorbital images. The physical fitness characterization module is used to generate children's physical fitness characterization results based on heart rate variability characteristics, physical fitness image characteristics, age group correction parameters, and developmental stage correction parameters. The gated fusion module is used to generate a myopia risk assessment feature vector based on data availability, acquisition time interval, and physical eye physiological response intensity. The risk assessment module is used to output the myopia risk value, risk level, and contribution of various input features to the myopia risk value based on the myopia risk assessment feature vector. The time-series management module is used to generate time-series archives of physical fitness results, eye physiological test data, environmental behavior data, and myopia risk values ​​according to a predetermined retest cycle, and to generate myopia risk change prompts and health management prompts.

Images