Method and electronic device for comprehensive assessment of eye development

Abstract

The application provides an eye development comprehensive evaluation method and an electronic device. A contrast encoder is used to encode an N-dimensional weighted eye development comprehensive index EDI module length of an evaluated person into a low-dimensional representation. The EDI module length is constructed by N-dimensional standardized data and weights. The weights of the N-dimensional standardized data are optimized based on a self-adaptive genetic algorithm. A fitness function adopted includes a representation item representing normal and abnormal separability. Finally, the eye development comprehensive evaluation of the evaluated person is performed according to the low-dimensional representation obtained by encoding. In this way, an eye development comprehensive index with a dimensionally expandable eye development comprehensive index is used to dynamically evaluate the eye development trajectory deviation. The multi-dimensional data affecting the eye development is considered, so that the evaluation result and the intervention measure are more accurate and efficient.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and electronic device for comprehensive assessment of eye development. Background Technology

[0002] Once true myopia occurs, it is irreversible. According to statistics, the population suffering from myopia is showing characteristics of younger age and an increasing proportion of high myopia. Myopia has seriously affected the visual health, learning and life, and employment of Chinese teenagers and children.

[0003] Current myopia assessment methods have the following main shortcomings:

[0004] (1) A single indicator cannot detect hidden risks. Approximately 15% to 20% of children with premyopia exhibit an inconsistent pattern of "normal refractive error (SE) but elongated axial length (AL)". Most assessment methods use a single indicator (SE or AL) to determine whether it is normal, but the refractive state of the eyeball (e.g., SE) and the anatomical structure (e.g., AL) may not match.

[0005] (2) Existing parameter-by-parameter comparisons lack a comprehensive index. For example, patient ocular parameters are compared independently with population norms, but there is a lack of methods to combine multiple parameters into a comprehensive index.

[0006] (3) There is a lack of an easily understandable indicator similar to the Body Mass Index (BMI). BMI reflects body shape through the ratio of height to weight. The field of optometry lacks a similar comprehensive indicator.

[0007] (4) The assessment dimensions are too narrow. The current assessment only focuses on SE and AL, ignoring the influence of corneal morphology (thickness, curvature), intraocular structure (anterior chamber depth), extraocular muscle load (heterotaxis / convergence function) and systemic metabolic status on eye development.

[0008] (5) Normative data cannot be updated adaptively. The existing norms are static data and cannot be automatically calibrated as new data accumulates.

[0009] (6) Lack of dynamic assessment of "developmental trajectory deviation". Existing methods are mostly single cross-sectional assessments, such as "whether the AL is above the standard in this examination", ignoring that eye development is a dynamic process. For the same 8-year-old child with AL=24.0mm, if their AL has been consistently above P90 since age 6 (stable trajectory), it has a completely different clinical significance than if their AL rapidly increases from P50 to P90 (trajectory deviation) - the latter is the high-risk signal that truly requires urgent intervention. Existing methods cannot distinguish between these two situations, leading to over-warning or under-warning. Summary of the Invention

[0010] This application proposes a comprehensive assessment method and electronic device for eye development, which can solve one of the problems existing in the background art.

[0011] To achieve the above objectives, this application adopts the following technical solution:

[0012] Firstly, a comprehensive assessment method for eyeball development is provided, including:

[0013] To obtain the N-dimensional raw data required for a comprehensive assessment of the subject’s eye development, the N-dimensional raw data includes: refractive error SE and axial length AL;

[0014] The original N-dimensional data is standardized to obtain standardized N-dimensional data. ,in, The dimension where the standardized data is located;

[0015] By comparing the encoder, the N-dimensional standardized data will be... and weight The modulus of the constructed N-dimensional weighted eye development index (EDI) is mapped to a low-dimensional representation space to obtain a low-dimensional representation. To maximize the separation of the normal group and the abnormal group, and based on the weight optimization mechanism of the adaptive genetic algorithm, the multidimensional standardized data is optimized to obtain the desired results. weight The fitness function of the adaptive genetic algorithm includes a separability term that characterizes the separability of the search target directly pointing to the weights after contrastive learning. , This is the iteration step size;

[0016] Based on the low-dimensional representation A comprehensive assessment of the eye development of the subjects was conducted.

[0017] In one possible design of the first aspect, the modulus of the N-dimensional weighted ocular development index is:

[0018]

[0019] In one possible design approach of the first aspect, the adaptive genetic algorithm specifically includes:

[0020] Initialize the population with equally weighted prior individuals As a seed individual;

[0021] Weight Substitute the modulus of the N-dimensional weighted eye development index EDI, calculate the fitness on the validation set with follow-up annotations according to the fitness function, and generate a new population through selection, crossover and mutation. This process is repeated until convergence.

[0022] In one possible design approach for the first aspect, P chromosomes are randomly generated in a real-number encoding manner during population initialization; during selection, tournament selection is used to construct a mating pool and retain elites; crossover is performed using simulated binary crossover; and mutation is performed using adaptive polynomial mutation.

[0023] The adaptive genetic algorithm further includes constraint repair, which involves truncating out-of-bounds genes to [0,1] and renormalizing them.

[0024] In one possible design approach of the first aspect, chromosomes are encoded as age function parameter pairs. , No. Dimension weights with age Continuous change is represented as: This transforms cross-age group weight fluctuations from external penalty terms into structural priors.

[0025] In one possible design approach of the first aspect, clinical hard constraints are applied during constraint repair, and dimensions are grouped according to anatomical relevance, with defined within-group weight ratios, wherein the clinical hard constraints include: ,in, Weighted by refractive power, Weighted by axial length.

[0026] In one possible design approach of the first aspect, several angular boundary parameters of different bias types are co-evolved into chromosome and weighted evolution. The EDI angle corresponding to the angular boundary parameter is an N-dimensional direction vector defined by a principal angle and auxiliary angles, wherein the principal angle is derived from diopter-normalized data. Standardized data with axial length The auxiliary angle is defined as being derived from the k-th dimension standardized data. and all standardized data definitions.

[0027] In one possible design of the first aspect, the contrast encoder is a two-layer perceptron, and the loss function of the contrast encoder is a supervised contrast loss;

[0028] During training, the contrast encoder uses EDI trajectory segments from the same child follow-up to form positive sample pairs; it embeds supervised contrastive loss with deviation type as group label and learns a type prototype for each type; and it assigns higher loss weights to samples with EDI modulus in the 1.0±0.2 boundary range.

[0029] In one possible design of the first aspect, the N-dimensional raw data further includes: mean corneal curvature K, axial length-corneal matching degree AL / CR, corneal thickness CCT, anterior chamber depth ACD, convergence load index CFI, and panphysiological index, wherein CR is the corneal curvature radius; the standardization is Z-score processing.

[0030] In a second aspect, an electronic device is provided, comprising: a processor, and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory such that the electronic device performs the method as described in any possible implementation of the first aspect.

[0031] Beneficial effects:

[0032] Based on the above technical solution, a contrastive encoder is used to encode the N-dimensional weighted eye development comprehensive index EDI modulus of the subjects into a low-dimensional representation. The EDI modulus is constructed from N-dimensional standardized data and its weights, and the weights of the N-dimensional standardized data are optimized based on an adaptive genetic algorithm. The fitness function used includes a representation term that represents the separability between normal and abnormal data. Finally, the comprehensive assessment of the subjects' eye development is carried out based on the encoded low-dimensional representation. In this way, an eye development comprehensive index with expandable dimensions can be used to dynamically assess deviations from the eye development trajectory, taking into account multi-dimensional data that affect eye development, making the assessment results and intervention measures more accurate and efficient. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart of a method for processing data on the assessment of children's eye development based on the Integrated Eye Development Index (EDI) provided in an embodiment of this application. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0036] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0038] This embodiment relates to the interdisciplinary field of optometry and statistics, specifically to a method for processing data on the assessment of children's eye development based on the Integrated Eye Development Index (EDI).

[0039] like Figure 1 As shown, the specific solution of this embodiment is as follows.

[0040] 5.1 Data Sources and Preprocessing

[0041] The input data in this embodiment comes from the following collection channels:

[0042] Unified definition table of parameter symbols:

[0043] The parameters involved in the following content are uniformly explained as follows.

[0044] (1) Parameters related to the adaptive genetic algorithm: P is the population size (preferably P=50); T is the maximum number of generations (preferably T=100); η is the mutation probability decay coefficient (default η=2.0); δ is the termination threshold, which terminates early when the optimal fitness improves less than δ for 10 consecutive generations; σ² is the population fitness variance; ε is the variance threshold, used to adaptively adjust the mutation probability; α, β, and γ are the weight coefficients of the Accuracy, Stability, and Complexity terms in the fitness function, respectively; α1 and α2 are the weight coefficients of the Accuracy and Sep terms in the extended fitness function; P c Crossover probability (P) c =0.9); η c The SBX distribution index (η) c =20); P m0 Let P be the initial mutation probability. m(t) Let a be the mutation probability in generation t; i b iThe parameter pairs for age-functionalized weighted encoding; w * This is the optimal weight vector after AGA convergence.

[0045] (2) Comparative learning of relevant parameters: z i z j z k Let be the representation vectors of sample i and its positive sample j and negative sample k within the batch, respectively, after being mapped by the encoder; N is the dimension of the input EDI vector; τ is the temperature parameter (τ=0.07); Sep is the inter-class distance D between the normal group and the abnormal group in the representation space. between Distance within class D within The ratio; ΔEDI is the change in EDI modulus between two consecutive follow-up visits.

[0046] In the following supervised comparison loss, sim(z) i ,z j ) represents sample z i With z j The similarity between them is calculated using cosine similarity; sim w (z i ,z j ) indicates the optimal weight w * Modulated weighted cosine similarity is used to highlight the differences in high-weight dimensions and suppress the influence of low-weight dimensions.

[0047] Data preprocessing:

[0048] (a) Outlier detection: SE < -15D or > +10D, AL < 18mm or > 30mm are marked as outliers and are not included in the calculation.

[0049] (b) Norm matching: Based on the subject's age (accurate to the month) and gender, the corresponding parameters are obtained from the norm database.

[0050] Detailed algorithm for norm matching:

[0051] Step B1: Age stratification and positioning;

[0052] Ages are grouped by whole year: 3 years old, 4 years old, ..., 18 years old, for a total of 16 age groups;

[0053] Enter age Unit: years, accurate to two decimal places, e.g., 8.25 represents 8 years and 3 months.

[0054] Determine the interval to which it belongs: , ;

[0055] In the formula for determining the interval mentioned above, A represents the subject's continuous age (in years, accurate to two decimal places); A floorThe largest integer age not greater than A is called the lower bound age group; A ceil The smallest integer age not less than A is called the upper bound age group; the interpolation weight w = AA floor Its value ranges from 0 to 1. For any parameter, the normative mean μ(A) and standard deviation σ(A) corresponding to age A are obtained by applying A... floor A ceil The corresponding statistics of two adjacent age groups are obtained by linear interpolation.

[0056] Step B2: Gender matching;

[0057] Select the corresponding gender norm sub-bank based on the subject's gender identifier (male / female);

[0058] If the gender identifier is missing, the mixed gender norm (male:female = 1:1 weighted) will be used by default.

[0059] Step B3: Linear interpolation calculation;

[0060] like Age is an integer (e.g.) ), directly read the parameters for that age group;

[0061] like Non-integer age (e.g.) Perform bilinear interpolation:

[0062] Interpolation weights ;

[0063] Step B4: Percentile value interpolation;

[0064] Linear interpolation is also performed for percentiles P5, P25, P50, P75, and P95;

[0065] Output the complete norm entries: {μ(A),σ(A),P5(A),P25(A),P50(A),P75(A),P95(A)};

[0066] 5.2 Step 1: Construction of Age-Gender Stratified Norms

[0067] Normative data are stratified according to the following dimensions:

[0068] (a) Age stratification: 3-18 years old, one stratum per year (can be refined to half a year).

[0069] (b) Gender stratification: Norms were established for males and females respectively.

[0070] (c) Each stratified unit contains statistics for each parameter: mean μ, standard deviation σ, and P5 / P25 / P50 / P75 / P95 percentiles.

[0071] Normative data sources: (1) publicly published national survey data (such as the 2022 expert consensus: AL baseline value of 6 years old is 22.46 mm, SE reference value is +1.38 D); (2) measured data accumulated in this project (120,000+ control cases).

[0072] Verification of the reliability of norm data: The norms of this project were cross-validated with the reference values ​​published in the 2022 National Expert Consensus on Vision Health of Children and Adolescents. The deviation of the AL reference value for 6-year-olds was <0.15mm, and the deviation of the SE reference value was <0.20D, showing good consistency. For some age groups (such as AL for 14-year-old males), the deviation was found to be due to the higher proportion of myopia in this age group in the sample. This has been corrected by increasing the sample size to n>800 in the norms.

[0073] Norm storage structure: Stored in JSON format, each parameter-age-gender combination stores one norm entry:

[0074] {"param":"SE","age":8,"sex":"male","mu":-0.75,"sigma":0.85,"p5":-2.10,"p95":0.50};

[0075] 5.3 Step 2: Z-score standardization calculation

[0076] For each parameter, calculate its Z-score (standardized score), which represents the degree to which that parameter deviates from the mean of the same-age and same-sex group. Take SE or AL, etc.:

[0077] The meaning of Z-score: Z=0 means that it is exactly at the mean level for the same age and gender; Z>0 means that it is above the mean (AL is high or SE is farsighted); Z<0 means that it is below the mean (AL is short or SE is nearsighted).

[0078] Example of parameter calculation (8-year-old boy, SE=-2.00D, AL=24.50mm):

[0079] The above formula indicates that SE is 1.25D more nearsighted than the average for the same age, which is about 1.47 standard deviations.

[0080] The above formula indicates that AL is 0.90 mm longer than the age-average value, approximately 2 standard deviations.

[0081] 5.4 Step 3: EDI Calculation (Core Innovation)

[0082] 5.4.1 Basic Two-Dimensional EDI (SE+AL)

[0083] Will and Perform a polar coordinate transformation to generate two components of the EDI:

[0084] The above formula represents the degree of deviation from normal development.

[0085] The above formula indicates the direction of deviation (which type of parameter deviates more).

[0086] Parameter description:

[0087] EDI modulus: Similar to BMI, the larger the value, the further it deviates from the normal range. A modulus <1.0 is within the normal range, 1.0~2.0 indicates a slight deviation, and >2.0 indicates a significant deviation.

[0088] EDI perspective: The negative sign is used because the more negative the SE (the higher the myopia), the higher the risk. The larger the value, the more severe the myopia.

[0089] Example calculation (continued from the previous example): EDI module length = √((-1.47)² + (2.00)²) = √(2.16 + 4.00) = √6.16 = 2.48 EDI angle = arctan2(2.00, 1.47) = 53.7°

[0090] Interpretation: EDI modulus = 2.48 (significantly deviates from normal), angle 53.7° corresponds to the type of "SE myopia protrusion + AL protrusion combined deviation".

[0091] 5.4.2 Automatic Determination of Seven Deviation Types

[0092] Based on the range of the EDI angle, the deviation type is automatically determined:

[0093] Geometric derivation of angular interval boundary values / parameters:

[0094] The boundary values ​​are determined by the angular division of a regular octagon: the 360° circumference is divided into 8 sectors at 45° intervals, where:

[0095] 25.7° = arctan(0.5), which corresponds to the dividing line where the weight ratio of SE to AL is 2:1;

[0096] 64.3° = arctan(2.0), which corresponds to the boundary line where the weight ratio of SE to AL is 1:2;

[0097] 115.7° = 180° - 64.3°, which is the symmetrical boundary of the second quadrant;

[0098] 154.3° = 180° - 25.7°, which is the symmetrical boundary of the second quadrant;

[0099] This division ensures that each sector corresponds to an independent type of deviation, and that the boundaries have clear clinical significance.

[0100] 5.4.3 N-dimensional progressively extended EDI

[0101] When more dimensions of data become available, EDI automatically expands:

[0102] After the weights of each dimension are optimized by the Adaptive Genetic Algorithm (AGA), the N-dimensional EDI modulus is expressed in a weighted form:

[0103] in For the first The weight coefficients for each dimension, with a default value of [value]. (Equal weights), the optimal weights are selected after AGA optimization. .

[0104] In the basic two-dimensional extension (N=2), each dimension is equally weighted (i.e., Since the Z-score has been standardized to the same scale (mean is 0, standard deviation is 1), it can directly participate in polar coordinate transformation without additional weighting.

[0105] If differentiated weights are required (such as adjusting the relative weights of SE and AL according to clinical importance), the AGA optimization scheme can be referenced, and the optimal weight combination can be adaptively determined by the genetic algorithm.

[0106] Extended path:

[0107] Definition of N-dimensional angle system:

[0108] For N-dimensional space, we define two levels: principal angle and auxiliary angle.

[0109] Main perspective (First principal plane angle): by and Angles of the reference plane formed:

[0110] The meaning is the same as in the two-dimensional case.

[0111] Auxiliary angle (k-th dimension elevation angle), representing the degree of deviation of the k-th dimension relative to the hyperplane formed by the first k-1 dimensions:

[0112] N-dimensional direction vector:

[0113] The aforementioned N-dimensional direction vector is composed of the principal angle θ1 and auxiliary angles θ2,...,θ N-1 A defined unit direction vector, whose components represent the direction cosine combination of EDI in the 1st to Nth dimensions, is used to locate the deviation direction in the N-dimensional normalized space.

[0114] By using auxiliary perspectives, higher-dimensional hidden risk patterns such as "SE normal + AL normal but K abnormal" can be identified.

[0115] Missing dimension handling: When data for a certain extended dimension is unavailable, the missing dimension is handled accordingly. (i.e., without deviation), it automatically reduces the dimensionality to the existing number of dimensions in the EDI. This ensures that evaluation is still possible even when the data is incomplete.

[0116] 5.4.4 AGA Weight Optimization

[0117] When differentiated weights are required for each dimension, an Adaptive Genetic Algorithm (AGA) is introduced to optimize the weight vector. The input is an N-dimensional Z-score vector. (in , (The rest are extended dimensions), and the output is the optimal weight vector w. * and weighted EDI module length The execution process is as follows:

[0118] Step G1 (Population Initialization): Randomly generate P chromosomes (preferably P=50) using real number encoding, each gene... And it satisfies the normalization constraint. And inject equal-weighted prior individuals As seed individuals to accelerate convergence;

[0119] Step G2 (fitness evaluation): Substitute the weights into the weighted EDI modulus and calculate the fitness on the validation set with follow-up annotations according to Fitness = α·Accuracy + β·Stability − γ·Complexity;

[0120] Step G3 (Selection): The tournament selects to build a mating pool (size 3 preferred), and retains the top 5% of elites;

[0121] Step G4 (Cross): SBX simulates binary cross ( Distribution index );

[0122] Step G5 (Mutation): Adaptive polynomial mutation, mutation probability It decays with each generation and adaptively adjusts according to the population fitness variance σ²;

[0123] Step G6 (Constraint Repair): Outbound genes are truncated to [0,1] and renormalized;

[0124] Step G7 (New Population Generation): Return to step G2 iteration;

[0125] Step G8 (Termination and Output): Terminate when the maximum number of generations T is reached (preferably T=100) or the optimal fitness is improved by less than the threshold δ for 10 consecutive generations, and output the optimal weight vector w. * .

[0126] To deeply couple weight optimization with eye development assessment, three measures are introduced:

[0127] Measure 1 (Age-based weighted encoding): Encode chromosomes as age function parameter pairs. The weight of the i-th dimension changes continuously with age. This transforms cross-age group weight fluctuations from an external penalty term into a structural prior; thus, the problem of cross-age group weight fluctuations is transformed from an external penalty term into a structural prior—the property of weights evolving smoothly with age is directly encoded into the genome, eliminating weight jumps with age from a mechanistic perspective, realizing age-adaptive weights, and significantly reducing the parameter tuning burden of variance penalty terms in fitness assessment.

[0128] Measure 2 (Domain Knowledge Constraint Repair): In constraint repair step G6, apply clinical hard constraints such as refractive core dimension weights and ≥0.5, and group the dimensions according to anatomical relevance, limiting the weight ratio within each group; the sum of the refractive core dimension weights should not be lower than a preset threshold (e.g., w). SE +w AL ≥0.5); and the dimensions are divided into refractive group, corneal group and eye position group according to anatomical relevance, with the weight ratio within each group limited to a clinically given range; chromosomes that violate the constraints are repaired according to preset rules rather than being directly eliminated, so that the evolutionary search is always carried out within a clinically reasonable subspace, avoiding the evolution of statistically effective but clinically uninterpretable weight combinations.

[0129] Measure 3 (Co-evolution of Bias Angle Boundaries): The angle boundary parameters of the seven bias types are incorporated into the chromosome and co-evolved with the weights. The angle boundary parameters of the seven bias types (such as the tangent ratios corresponding to 25.7° and 64.3°) are incorporated into the chromosome and co-evolved with the dimension weights: the weights determine the stretching degree of each dimension of EDI, and the angle boundaries determine the dividing line of the bias type. The co-optimization of the two makes the boundary of the bias type determination adaptively adjusted according to the distribution of real data, rather than using a fixed geometric division, so that the coordinate transformation itself also has data-driven adaptive capability.

[0130] Validated by 120,000 clinical cases, the AUC of the AGA-optimized weighted EDI in identifying latent myopia risk increased from 0.82 to 0.91, which is significantly better than the fixed weight scheme.

[0131] 5.4.5 Contrastive Learning Representations and Refined Risk Stratification

[0132] Based on the aforementioned weighted EDI, further contrastive learning is used to map the N-dimensional weighted EDI vector to a low-dimensional representation space, maximizing the separation between "normal development" and "abnormal development" patterns in this space. This mechanism is consistent with innovation point six. The encoder employs a two-layer MLP: Input(N) → Linear(64) → ReLU → Dropout(0.3) → Linear(32) → L2 normalization, outputting a 32-dimensional representation vector z; the loss function uses supervised contrastive loss SupCon. ,in For temperature parameters, samples were labeled "normal group" and "abnormal group" according to follow-up outcomes.

[0133] Three measures were introduced simultaneously during the training:

[0134] Measure 1 (Time-series comparison of developmental trajectories): positive sample pairs are formed by using EDI trajectory segments from two consecutive follow-ups of the same child to capture the dynamic evolution pattern of developmental trajectories rather than static snapshots.

[0135] Measure 2 (Fine-grained comparison of deviation type perception): The SupCon loss is embedded with seven deviation types as group labels and a type prototype is learned for each type.

[0136] Measure 3 (Mining of Difficult-to-Negative Samples for Determining Boundary): Assign higher loss weights to samples whose EDI modulus is in the boundary range of 1.0±0.2.

[0137] During the assessment, the weighted EDI vector of the subject is input into the trained encoder to obtain the representation z. Based on the distance between z and the prototype center of the normal group and the abnormal group, as well as the prototype of the deviation type, refined risk stratification and intervention suggestions are output.

[0138] On the independent test set (n=2000), the sensitivity for identifying latent myopia (normal SE but longer AL) was further improved from 0.91 for AGA to 0.94, and the specificity was improved from 0.85 to 0.90.

[0139] The aforementioned weight optimization and contrastive learning representation form a closed-loop collaborative structure, rather than a unidirectional chain, which is achieved through the following three connection points:

[0140] Association point A (contrastive representation of separability feeds back into the fitness function): The fitness function is expanded to Fitness = α1·Accuracy + α2·Sep + β·Stability − γ·Complexity, where Sep = D between / D within This is the ratio of the inter-class distance to the intra-class distance between the normal and abnormal groups in the representation space after the weights are mapped by the contrastive encoder. This ensures that the search objective of AGA directly targets the final separability of the weights after contrastive learning, aligning the weight optimization objective with the representation learning objective.

[0141] In the fitness function of the adaptive genetic algorithm described above, Accuracy represents the classification accuracy, which is the consistency ratio obtained by comparing the validation set samples with the follow-up labels after determining whether the weighted EDI magnitude falls within the normal range and the deviation type range of the EDI angle. Stability represents the change range of the optimal weight between adjacent generations, which is used to suppress weight oscillation. Complexity represents the sparsity or entropy of the weight vector, which is used to avoid overfitting. The three terms together measure the overall merits of the weight vector in terms of classification performance, continuous stability and model complexity.

[0142] Relevance point B (weight-guided positive sample construction and similarity modulation): When constructing positive sample pairs, a strong perturbation is applied to the high-weight dimensions of the AGA output to enhance them, while the low-weight dimensions remain approximately unchanged. This forces the encoder to remain sensitive to clinically important dimensions and invariant to secondary dimensions. The cosine similarity in the contrast loss uses weighted similarity modulation. w(zi,zj) With optimal weight w * The contributions of each representation dimension are weighted to make the distance structure of the representation space consistent with the clinical priors learned by AGA.

[0143] Related point C (alternating iterative closed loop): Step S a : Fix the current weights w(t), train the contrastive encoder, and calculate the representation separability Sep(t); Step S b : Run AGA evolution with Sep as the fitness term to obtain new weights w(t+1); Step Sc: If |Sep(t+1)−Sep(t)|<ε sepStop if the maximum number of alternations is reached; otherwise, return to step S. a After convergence, the joint optimal weight vector w is output. * With encoder parameters.

[0144] In the above alternating iterative closed loop, ε sep To characterize the convergence threshold of separability Sep, when the change in Sep obtained from two adjacent rounds of alternating optimization is less than ε sep (preferably ε) sep The iteration stops when the value is 1e-4.

[0145] 5.5 Step 4: Comprehensive Judgment

[0146] Based on the EDI modulus and deviation type, a comprehensive judgment result is generated:

[0147] 5.6 Step 5: Adaptive Update of Norm

[0148] Normative parameters are automatically calibrated as new data accumulates, and updated using Bayesian posterior regression.

[0149] in:

[0150] The original norm sample size refers to the total number of historical data samples used to construct the current norm.

[0151] Newly collected sample size, i.e., the number of new samples added in the most recent data accumulation period;

[0152] , : The mean and standard deviation of the original norm;

[0153] , The sample mean and sample standard deviation of the newly collected data;

[0154] Triggering condition: When new data accumulates to 10% of the original norm (i.e. This triggers an update.

[0155] Alternative solution: Sliding window update – retain only data from the most recent few years (e.g., 5 years) to build a norm, and automatically discard outdated data.

[0156] 5.7 Step 6: Calculation of CFI Convergence Load Index

[0157] CFI quantifies the degree of extraocular muscle fatigue during near work:

[0158] in:

[0159] Near-range esophoria (△), negative values ​​indicate exophoria. Absolute values ​​are used because both esophoria and exophoria increase the load on the eye muscles.

[0160] The base-outward melting point (△) reflects the recovery capacity of the melting phase reserve;

[0161] The fusion fracture point (Δ) with the substrate facing outward reflects the limit of the fusion phase reserve;

[0162] Recovery / rupture ratio: The closer to 1, the stronger the recovery ability.

[0163] Extended CFI (incorporating muscle biomechanics calculations):

[0164] Extending CFI in Basic CFI (CFI) base Based on this, the extraocular muscle strength coefficient F is further incorporated. eom With ciliary muscle tension index T cil To quantify the contributions of extraocular muscles and ciliary muscles to near eye load. Among them, F eom Based on comprehensive near-field eye position, fusion reserve, and other detection indicators, T is estimated. cil The calculation is based on comprehensive adjustment response, adjustment amplitude and other test indicators.

[0165] in This represents the coefficient of extraocular muscle strength. This is the ciliary muscle tension index.

[0166] This formula is applicable to children and adolescents aged 8-18. For individuals under 8 or over 18, the coefficients in the formula (0.05, 0.1, etc.) need to be adjusted according to the age group.

[0167] The coefficients mentioned above (0.05, 0.1, etc.) are based on the correlation analysis between the hemianopsia value and the myopia progression rate in 120,000 cases of data in this project. Specific method: The group with an annual AL increase > 0.2 mm was designated as the "progression group," and the group with an increase ≤ 0.2 mm was designated as the "stable group." The optimal values ​​for each coefficient of the CFI were selected using Logistic regression.

[0168] Cross-validation on different subsamples (divided into 3 age groups) showed that CFI had an AUC of 0.78 for predicting the risk of progression, which was better than using the hysteresis value alone (AUC = 0.65) or AL alone (AUC = 0.72), indicating that the comprehensive quantification of CFI did indeed capture the independent contribution of ocular muscle load to axial length growth.

[0169] 6-8 years old: coefficient × 0.8 (eye muscles are not yet fully developed, and the load tolerance is low);

[0170] 8-18 years old: Use standard coefficients (as shown in the formula above);

[0171] >18 years old: Coefficient × 1.2 (Adults have decreased eye muscle accommodation ability, resulting in higher fatigue under the same load);

[0172] 5.8 Step 7: EDI+CFI Joint Risk Assessment Matrix

[0173] By combining EDI modulus (degree of refractive development deviation) and CFI (degree of eye muscle fatigue), a two-dimensional risk matrix is ​​constructed:

[0174] 5.9 Example

[0175] Example: A 10-year-old girl.

[0176] Input data: SE=-1.25D, AL=24.80mm, K1=42.50D, K2=43.25D, CCT=540μm, phoria near =-4△,break BO =18△, recovery BO =12△.

[0177] Norm matching (10-year-old female): SE norm μ=-0.50D, σ=0.90D; AL norm μ=23.80mm, σ=0.42mm.

[0178] Z-score calculation: Z SE =(-1.25-(-0.50)) / 0.90=-0.83; Z AL =(24.80-23.80) / 0.42=2.38.

[0179] EDI calculation: Modulus length = √(0.83² + 2.38²) = √(0.69 + 5.66) = √6.35 = 2.52. Angle = arctan2(2.38, 0.83) = 70.8°.

[0180] Deviation type: 70.8° in the 64.3°~90° range → "simple AL lengthening" (latent myopia risk! Although SE is only -1.25D which seems not serious, AL is already 2.38 standard deviations longer than the same age).

[0181] CFI calculation: CFI = 4 × (12 / 18) × (1 + 0.05 × |12 - 4|) = 4 × 0.67 × 1.40 = 3.75.

[0182] Joint assessment: EDI=2.52 (significant deviation) + CFI=3.75 (high load) → emergency comprehensive intervention.

[0183] Recommendations: (1) Conduct a detailed examination immediately to rule out pathological myopia; (2) Consider orthokeratology (OK) lenses or defocus lenses to control axial elongation; (3) Improve convergence and divergence function through vision training; (4) Increase outdoor activities to more than 2 hours per day.

[0184] To fully demonstrate the execution process of innovation points five and six under the extended dimension, we will continue to use the aforementioned 10-year-old girl as an example: Let her K... avg Given the constant μ = 42.80D and σ = 1.20D, the constant μ of CCT is 550μm and σ = 30μm, and the constant μ of ACD is 3.20mm and σ = 0.20mm, then Z K =(42.875−42.80) / 1.20≈0.06, Z CCT =(540−550) / 30≈−0.33、Z ACD =(3.10−3.20) / 0.20≈−0.50, and Z SE Z AL This forms a five-dimensional Z-score vector.

[0185] In step 5.4.4 of the AGA weight optimization process: An equal-weighted prior, w=(0.2,0.2,0.2,0.2,0.2), is injected into the initial population as a seed individual, and the remaining individuals are randomly generated. This process involves fitness evaluation, tournament selection, SBX crossover, adaptive mutation, and constraint repair (satisfying w). SE +w AL After multiple iterations (≥0.5), the algorithm converges and outputs the optimal weight w. * Combining with Improvement 1 (age-based weighted encoding), the weights change continuously with age, for example, w AL (age) = a + b·age is taken as 0.35 at age 10, and the importance of axial length is further highlighted as age increases.

[0186] In the contrastive learning representation step 3.5: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] * The weighted five-dimensional EDI vector is input into a two-layer MLP encoder, mapped to a 32-dimensional representation z, and then L2 normalized. This case study demonstrates the EDI... w≈1.48, which is in the deviation range of 1.0~2.0 and does not fall into the boundary region of the difficult negative sample of 1.0±0.2; the z-characteristic is compared with the prototype center distance of the normal group and the abnormal group. Combined with the improvement one (development trajectory time sequence comparison), if the length of the EDI modulus is significantly increased compared with 3 months ago (ΔEDI>0.5), then its "trajectory deviation" attribute is identified and an emergency intervention prompt is output.

[0187] Finally, combining CFI=3.75 (high load) with weighted EDI w The value is approximately 1.48, resulting in a "high" risk level, and intervention recommendations consistent with those mentioned above are provided. This concludes the complete execution steps of this case study, from multidimensional data input, AGA weight optimization and comparative learning representation, to the final risk assessment result.

[0188] The data in the above embodiments are derived from real cases from the collaborating units of this project.

[0189] The patient (female, 10 years old) was re-examined 3 months after the initial assessment, and the AL increased by 0.12 mm (> the age-matched P75 growth rate).

[0190] The risk assessment of "simplistic AL elongation" indicated by the EDI angle of 70.8° was confirmed. After adjusting the intervention program (increasing outdoor activities + visual training), the AL growth rate dropped to 0.06 mm / 3 months after 6 months (returning to the P50 level).

[0191] This case illustrates how the precise location of deviation direction by the EDI angle helps clinicians develop targeted intervention strategies.

[0192] 5.10 Alternative

[0193]

[0194] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0195] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0196] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0197] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0198] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0199] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0200] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0201] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A comprehensive assessment method for eyeball development, characterized in that, include: To obtain the N-dimensional raw data required for a comprehensive assessment of the subject’s eye development, the N-dimensional raw data includes: refractive error SE and axial length AL; The original N-dimensional data is standardized to obtain standardized N-dimensional data. ,in, The dimension where the standardized data is located; By comparing the encoder, the N-dimensional standardized data will be... and weight The modulus of the constructed N-dimensional weighted eye development index (EDI) is mapped to a low-dimensional representation space to obtain a low-dimensional representation. To maximize the separation of the normal group and the abnormal group, and based on the weight optimization mechanism of the adaptive genetic algorithm, the multidimensional standardized data is optimized to obtain the desired results. weight The fitness function of the adaptive genetic algorithm includes a separability term that characterizes the separability of the search target directly pointing to the weights after contrastive learning. , This is the iteration step size; Based on the low-dimensional representation A comprehensive assessment of the eye development of the subjects was conducted.

2. The method as described in claim 1, characterized in that, The modulus of the N-dimensional weighted comprehensive eye development index is:

3. The method as described in claim 1, characterized in that, The adaptive genetic algorithm specifically includes: Initialize the population with equally weighted prior individuals As a seed individual; Weight Substitute the modulus of the N-dimensional weighted eye development index EDI, calculate the fitness on the validation set with follow-up annotations according to the fitness function, and generate a new population through selection, crossover and mutation. This process is repeated until convergence.

4. The method as described in claim 3, characterized in that, In the initialization of the population, P chromosomes are randomly generated using real number encoding; during the selection process, the tournament selection constructs a mating pool and retains the elites; Crossover is performed using simulated binary crossover; mutation is performed using adaptive polynomial mutation. The adaptive genetic algorithm further includes constraint repair, which involves truncating out-of-bounds genes to [0,1] and renormalizing them.

5. The method as described in claim 4, characterized in that, Encode chromosomes as age function parameter pairs , No. Dimension weights with age Continuous change is represented as: This transforms cross-age group weight fluctuations from external penalty terms into structural priors.

6. The method as described in claim 4, characterized in that, In constraint repair, clinical hard constraints are applied, and dimensions are grouped according to anatomical relevance, with within-group weight ratios defined. These clinical hard constraints include: ,in, Weighted by refractive power, Weighted by axial length.

7. The method as described in claim 4, characterized in that, Several types of angular boundary parameters are incorporated into chromosome and weighted co-evolution. The EDI angle corresponding to each angular boundary parameter is an N-dimensional direction vector defined by a principal angle and auxiliary angles. The principal angle is derived from diopter-normalized data. Standardized data with axial length The auxiliary angle is defined as being derived from the k-th dimension standardized data. and all standardized data definitions.

8. The method as described in claim 7, characterized in that, The contrast encoder is a two-layer perceptron, and the loss function of the contrast encoder is supervised contrast loss; During training, the contrast encoder uses EDI trajectory segments from the same child follow-up to form positive sample pairs; it embeds supervised contrastive loss with deviation type as group label and learns a type prototype for each type; and it assigns higher loss weights to samples with EDI modulus in the 1.0±0.2 boundary range.

9. The method as described in claim 1, characterized in that, The N-dimensional raw data also includes: mean corneal curvature K, axial length-corneal matching degree AL / CR, corneal thickness CCT, anterior chamber depth ACD, convergence load index CFI, and panphysiological index, where CR is the corneal curvature radius; the standardization is Z-score processing.

10. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-8.

Images