Corneal molding lens fitting ensemble learning method based on corneal feature attention

Abstract

This invention relates to the field of data processing technology, and more particularly to an integrated learning method for orthokeratology lens fitting based on corneal feature attention. The method includes: S100, acquiring raw input data; S200, a feature subset module dividing the raw input data into multiple training subsets using a medical prior grouping strategy or a dynamic importance grouping strategy, and using these multiple training subsets to train a base learner, which then outputs its predicted values; S300, the raw input data entering a feature attention module to obtain a final attention fusion result; and S400, training a meta-learner based on the base learner's predicted values ​​and the final attention fusion result, which then outputs its final prediction result. This invention improves the accuracy of orthokeratology lens fitting.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an integrated learning method for orthokeratology lens fitting based on corneal feature attention. Background Technology

[0002] Existing patent CN116564540A discloses a method and system for predicting orthokeratology lens parameters based on an ensemble learning algorithm, used to improve the accuracy of orthokeratology lens parameter prediction. The method includes: segmenting a historical optometry dataset into multiple feature subsets; predicting orthokeratology lens parameters for each feature subset to obtain candidate prediction data; performing error analysis between the real data and the candidate prediction data to obtain error data; calculating the weights of each base learner to obtain weight data for each base learner; performing sample weight analysis on the candidate prediction data to generate a sample weight set; weighting multiple candidate prediction data to obtain a feature dataset to be analyzed; weighting the feature dataset to be analyzed using the weight data corresponding to each base learner to obtain a target feature dataset; and inputting the target feature dataset into a meta-learner to predict orthokeratology lens parameters to obtain target prediction data.

[0003] Existing technologies often struggle to effectively integrate prior medical knowledge into data-driven models, leading to predictions that deviate from clinical logic. High-dimensional biometric features exhibit dynamic redundancy and interaction effects, making it difficult for traditional feature selection methods to capture their nonlinear relationships. Furthermore, medical features possess strong biological correlations, and traditional random or single-feature removal grouping strategies disrupt the intrinsic connections between features, contradicting real-world scenarios in clinical practice where "entire feature sets are missing" (e.g., the loss of an entire set of corneal examination data). Additionally, the feature attention update mechanism is inadequate, failing to fully integrate sparse attention weights. Moreover, the lack of system architecture support in feature alignment, dimensionality adaptation (equal dimension preservation, missing value imputation and dimensionality increase, priority-based dimensionality reduction), and alignment layer health monitoring results in poor feature representation and fusion, ultimately impacting prediction accuracy and efficiency. Summary of the Invention

[0004] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides an integrated learning method for orthokeratology lens fitting based on corneal feature attention, thereby improving the fitting accuracy of orthokeratology lenses.

[0005] This invention provides an integrated learning method for orthokeratology lens fitting based on corneal feature attention, comprising the following steps:

[0006] S100: Obtain the raw data and perform feature standardization to form the raw input data;

[0007] S200, the original input data enters the feature subset module. The feature subset module divides the original input data into multiple training subsets through a medical prior grouping strategy or a dynamic importance grouping strategy. The base learner is trained using the multiple training subsets, and the base learner outputs the base learner prediction value.

[0008] S300, the raw input data enters the feature attention module;

[0009] The feature attention module performs corneal feature enhancement, feature fusion, and feature importance calibration operations on the corneal features in the original input data, thereby updating the initial feature weights corresponding to the original features of the original input data to the calibrated feature weights;

[0010] The feature alignment layer of the feature attention module updates the input feature dimension of the original input data after calibrating the feature weights.

[0011] The feature attention module obtains the sparse attention weights of the original input data after adjusting the dimension of the input features by differentiable sparse attention;

[0012] A feature update mechanism is set in the feature attention module. The feature update mechanism uses the sparse attention weights to process the original features of the original input data to obtain the final attention fusion result.

[0013] S400 trains a meta-learner based on the base learner's predictions and the final attention fusion result, and the meta-learner outputs the final prediction result.

[0014] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention is that step S200 includes the following steps:

[0015] S201, the feature subset module presets groups, including corneal information group, biometrics group, refractive data group, and demographic group. The medical prior grouping strategy excludes one group at a time to generate a training subset.

[0016] S202, when the feature subset module does not have a preset group, dynamic feature sampling is enabled, and the number of features in the original input data is greater than the maximum number of base learners, the dynamic importance grouping strategy is enabled;

[0017] When n > 100

[0018] Where n represents the number of samples in the original input data. This represents the feature importance score of the j-th original feature. This represents the Pearson correlation coefficient. The vector representing the j-th original feature. This is a real label;

[0019] When n≤100

[0020] in, The LassoCV coefficient represents the j-th original feature in the feature subset module;

[0021] Normalize the feature importance score for each original feature, based on... Choose the most important one. One primary feature is designated as the key feature, and the rest are secondary features.

[0022]

[0023] in, This represents the maximum number of base learners. This represents the total number of features in the original input data;

[0024] right For each key feature, generate several training subsets that exclude each key feature; for secondary features, generate a training subset that includes all secondary features; and generate a training subset for all original features.

[0025] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention is that S300 includes the following steps:

[0026] S301, the feature attention module processes corneal-related features in the original features of the original input data through a dedicated embedding layer, and uses a gating mechanism to dynamically enhance corneal-related features into enhanced corneal features;

[0027]

[0028] in, Indicates original features, Indicates corneal-related characteristics, Embedded representations of corneal-related features Indicates cross-enhancement features, Indicates the first gating weight. Indicates enhanced corneal features, This represents the first learnable weight matrix. This represents the second learnable weight matrix. This represents the third learnable weight matrix. This represents the fourth learnable weight matrix. This represents the fifth learnable weight matrix. This represents the first learnable bias vector. This represents the second learnable bias vector. This represents the third learnable bias vector. This represents the fourth learnable bias vector. This represents the fifth learnable bias vector. This represents the sigmoid function. Presentation layer normalization operation;

[0029] S302, the feature attention module processes the original input data through a dedicated embedding layer to obtain a conventional feature embedding. This will enhance corneal features. Compared with conventional feature embedding Add, The characteristics after fusion are obtained. ;

[0030] S303, for Take the average. corneal influence factor was obtained. ,in, This indicates the operation of taking the average value;

[0031] According to corneal impact factors Obtain the calibration feature weights ;

[0032]

[0033] in, This represents the sixth learnable weight matrix. This represents the seventh learnable weight matrix. This represents the sixth learnable bias vector. This represents the seventh learnable bias vector. Indicates the initial feature weights. Indicates the calibration feature weights, represents the learnable scaling factor, and Q represents the query vector.

[0034] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention lies in that the feature alignment layer of the feature attention module updates the input feature dimension of the original input data after calibrating the feature weights, specifically including the following steps:

[0035] S311, the input feature dimension of the original input data after calibrating the feature weights. The embedding layer target dimension of the feature alignment layer of the feature attention module. Make comparisons;

[0036] like If the scene is determined to be of equal dimensions, the feature alignment layer directly matches the dimension.

[0037] like If it is determined to be a dimensional upgrade scenario, then ,

[0038] in, This represents the output feature vector after feature alignment. This represents the alignment transformation function. Indicates zero-filling operation;

[0039] like If it is determined to be a dimensionality reduction scenario, then ,in, Represents the dynamic fusion coefficient. Indicates the preceding Dimensional original features, , This represents the L2 norm of the original eigenvector.

[0040] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention is that S300 further includes the following step:

[0041] Set up a feature alignment layer protection mechanism. ,in, This represents the information loss rate. When the information loss rate exceeds a set threshold of 0.25, the feature alignment layer is frozen.

[0042] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention is that it further includes the following steps:

[0043] A course learning strategy is set in the feature attention module. When the input feature dimension matches the target dimension of the embedding layer, the feature attention module executes the course learning strategy to control the feature retention ratio of the feature attention module. The course learning strategy is as follows:

[0044]

[0045] in, Indicates the current training round. Indicates the total number of training rounds. Indicates the minimum feature retention ratio. This indicates the percentage of the maximum feature retained.

[0046] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention is that the processing of differentiable sparse attention includes the following steps:

[0047] S321, Adjust the global interaction importance of the feature attention module;

[0048]

[0049] in, This represents batch matrix multiplication, where Q represents the query vector. Represents the key vector. This indicates the number of features in the original input data. Represents the global interaction matrix. This represents the element in the i-th row and j-th column of the global interaction matrix. Indicates the importance of global interactions;

[0050] S322, Calculate the final score of the calibration feature weights of the original input data:

[0051]

[0052] in, This represents the global weighted score. This represents the global weighted score of the i-th calibration feature weight. Indicates the final score;

[0053] S323, based on the final score, calculates the differentiable mask, sparse weights, and second gating value, and then calculates the sparse attention weights.

[0054] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention is that S323 includes the following steps:

[0055] When calculating the differentiable mask, the dynamic K value is predicted based on the final score, and then the differentiable mask is calculated.

[0056] The dynamic K-value represents the number or proportion of key features that should ultimately be retained for each sample in the original input data. The formula for calculating the dynamic K-value is:

[0057]

[0058] in, This represents the eighth learnable weight matrix. This represents the ninth learnable weight matrix. This represents the eighth learnable bias vector. This represents the ninth learnable bias vector. This represents the feature retention ratio coefficient obtained after two layers of linear transformation and activation, ranging from (0,1), and is used for subsequent interpolation to obtain the final retention count. This indicates the number or proportion of key features that should ultimately be retained for each sample;

[0059] Add Gamper noise to the final score.

[0060]

[0061] in, This represents the noisy feature score after adding Gamper noise. Let represent a random variable that follows a uniform distribution on the interval (0,1). This represents a learnable temperature parameter;

[0062] TopK selection

[0063]

[0064] in, This represents a binary mask, indicating whether the j-th feature of the i-th sample is selected, with 1 indicating selection and 0 indicating non-selection. This represents the score of the i-th noisy feature; Represents the i-th sample ; This represents the TopK selection operation, which selects the top K original features with the highest scores from the input noisy feature scores, where K is determined by... Dynamic adjustment;

[0065] Calculate the differentiable mask.

[0066]

[0067] in, Indicates a soft mask. express mean This represents the differentiable mask ultimately used for end-to-end training. This represents the binary mask selected by TopK. This represents the gradient separation operation;

[0068] When calculating sparse weights, the scores are sorted according to the final scores, and then the support set and projection threshold are calculated to obtain the sparse weights.

[0069] Sort the final scores:

[0070] in, This represents the final score of the first-ranked item. This represents the final score of the second-ranked item. This represents the final score of the Nth rank. , This represents the final sequence of scores after sorting.

[0071] Compute the support set,

[0072]

[0073] in, Indicates the size of the candidate's support set. This represents the final score of the k-th sorted number. This represents the sum of the final scores of the top k largest values ​​after sorting.

[0074] Calculate the projection threshold :

[0075] Output sparse weights,

[0076] in, Indicates sparse weights;

[0077] Second gating value calculate:

[0078]

[0079] in, This represents the tenth learnable weight matrix. This represents the eleventh learnable weight matrix;

[0080] Performing gating fusion operations yields sparse attention weights:

[0081]

[0082] in, This represents sparse attention weights.

[0083] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention lies in setting a feature update mechanism in the feature attention module. The feature update mechanism uses the sparse attention weights to process the original features of the original input data to obtain the final attention fusion result, including the following steps:

[0084] Weight the query vector.

[0085] in, This represents the query vector after adjusting the sparse attention weights. Represents the sparse attention weights after dimension alignment;

[0086] Obtaining the low-rank projection ,

[0087]

[0088] in, Represents the low-rank projection , This represents the low-rank projection matrix of the query vector. This represents the key vector after low-rank projection. Represents the original key vector. The low-rank projection matrix representing the key vector. and The dimension is , , This represents the feature dimension of a single head in a multi-head attention mechanism;

[0089] Obtain attention score,

[0090]

[0091] in, This represents a Boolean mask used to label features with sparse attention weights greater than 0.1; This represents the attention score. This represents the Softmax function, used to normalize attention scores to a probability distribution that sums to 1. This represents an element in a Boolean mask, used to mark the first... Whether the original features of the location are preserved. Indicates reservation;

[0092] Perform vector aggregation operation: ,in, Represents the context vector. Represents a value vector;

[0093] Perform feature copy and update operations:

[0094]

[0095] in, The projection result representing the original features. Indicates the number of copies. Indicates the characteristics after replication. This represents the mask after expanding the dimensions. Indicates the updated features;

[0096] Perform a residual gating weight fusion operation to obtain the fused features.

[0097]

[0098] in, Indicates the residual gating weight. Indicates the characteristics after fusion. This represents the twelfth learnable weight matrix. This represents the twelfth learnable bias vector;

[0099] Calculate the final output of the attention-weighted calculation.

[0100] in, This represents the final output of the attention-weighted algorithm. This represents the thirteenth learnable weight matrix. This represents the thirteenth learnable bias vector;

[0101] The final attention fusion result is obtained by fusing the attention-weighted final outputs. .

[0102] A further improvement of the corneal feature attention-based integrated learning method for orthokeratology lens fitting in this invention is that step S400 includes the following steps:

[0103] S401, Input the base learner predictions into the meta-attention module, and calculate the sample-level weight matrix through the meta-attention module:

[0104] in, This represents the prediction matrix of the base learner. This represents the sample-level weight matrix, where F represents the number of base learners. This indicates the self-attention mechanism. This represents the fourteenth learnable weight matrix. This represents the fifteenth learnable weight matrix;

[0105] The base learner predictions are weighted using the sample-level weight matrix to obtain weighted base learner predictions. ;

[0106] in, This represents the sample-level weight matrix corresponding to the i-th base learner. This represents the predicted value of the i-th base learner, and the meta-learner is trained using a linear regression model.

[0107] S402, weighted base learner predictions Input the meta-learner, and the meta-learner will output the final prediction result;

[0108] S403, during step S402, the meta-attention module and the feature attention module perform collaborative optimization through a joint loss function, which is used to consider the error between the final prediction result output by the meta-learner and the true label, as well as the prediction contribution of the feature attention module.

[0109] The joint loss function is:

[0110]

[0111] in, Indicates the loss value. This represents the first learnable fusion parameter. This represents the second learnable fusion parameter. This indicates the actual label.

[0112] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0113] By acquiring raw data and standardizing its features, differences arising from varying data sources and scales can be eliminated, laying a solid foundation for subsequent accurate analysis. The feature subset module employs a medical prior grouping strategy or a dynamic importance grouping strategy to divide the training subset. This approach fully considers the differences in professional knowledge and data importance within the medical field, enabling the base learner to be trained on more reasonable and targeted data, thereby outputting more accurate base learner predictions. Simultaneously, the established minimum protection threshold ensures the stability and reliability of the model.

[0114] By performing corneal feature enhancement, feature fusion, and feature importance calibration on the raw input data, key features can be effectively highlighted, medical prior knowledge can be integrated into the machine learning model, and the model's ability to capture important information can be improved. Adjusting the target dimension of the embedding layer of the feature alignment layer ensures dimensionality matching during data processing and avoids information loss or errors caused by dimensional inconsistencies.

[0115] The learning strategy can gradually adjust the learning content according to the learning progress and difficulty, thereby improving learning efficiency. By obtaining sparse attention weights through differentiable sparse attention, it is possible to more accurately focus on features that have a significant impact on the prediction results. The feature update mechanism further optimizes the attention weights, resulting in the final attention fusion result, enabling the model to better utilize key information in the data.

[0116] The meta-learner integrates the predictions from the base learner and the final attention fusion result, treating the attention network as an independent predictor and fusing it in parallel with the ensemble learning model. This fully leverages the advantages of ensemble learning, comprehensively considering multiple aspects of information. The attention network incorporates medical knowledge as a regularization term into the loss function, guiding the model to think according to medical priors, thereby obtaining more accurate and reliable final prediction results. This provides strong technical support for the precise fitting of orthokeratology lenses.

[0117] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0118] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0119] Figure 1 This is a block diagram illustrating an integrated learning method for orthokeratology lens fitting based on corneal feature attention provided by the present invention.

[0120] Figure 2 This is a diagram illustrating the sparse attention execution in the corneal feature attention-based integrated learning method for orthokeratology lens fitting provided by this invention.

[0121] Figure 3 This is a block diagram illustrating the execution of a single head of the multi-head attention mechanism in the corneal feature attention-based integrated learning method for orthokeratology lens fitting provided by this invention. Detailed Implementation

[0122] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but should not be used to limit the scope of this invention.

[0123] The following is combined Figure 1 The present invention describes an integrated learning method for orthokeratology lens fitting based on corneal feature attention, comprising the following steps:

[0124] S100: Obtain the raw data and perform feature standardization to form the raw input data;

[0125] S200, the original input data enters the feature subset module. The feature subset module divides the original input data into multiple training subsets through a medical prior grouping strategy or a dynamic importance grouping strategy. The base learner is trained using the multiple training subsets, and the base learner outputs the base learner prediction value.

[0126] S300, the raw input data enters the feature attention module;

[0127] The feature attention module performs corneal feature enhancement, feature fusion, and feature importance calibration operations on the corneal features in the original input data, thereby updating the initial feature weights corresponding to the original features of the original input data to the calibrated feature weights;

[0128] The feature alignment layer of the feature attention module updates the input feature dimension of the original input data after calibrating the feature weights.

[0129] The feature attention module obtains the sparse attention weights of the original input data after adjusting the dimension of the input features by differentiable sparse attention;

[0130] A feature update mechanism is set in the feature attention module. The feature update mechanism uses the sparse attention weights to process the original features of the original input data to obtain the final attention fusion result.

[0131] S400 trains a meta-learner based on the base learner's predictions and the final attention fusion result, and the meta-learner outputs the final prediction result.

[0132] Ideally, in stage S100, acquiring raw data and standardizing its features eliminates differences caused by varying data sources and scales, laying a solid foundation for subsequent accurate analysis. In stage S200, the feature subset module uses either a medical prior grouping strategy or a dynamic importance grouping strategy to divide the training subset. This approach fully considers the differences in professional knowledge and data importance within the medical field, enabling the base learner to be trained on more reasonable and targeted data, thus outputting more accurate base learner predictions. The determined minimum protection threshold also ensures the model's stability and reliability. Step S300 performs corneal feature enhancement, feature fusion, and feature importance calibration on the raw input data, effectively highlighting key features, integrating medical prior knowledge into the machine learning model, and improving the model's ability to capture important information. The feature alignment layer updates the input feature dimensions of the raw input data after calibrating the feature weights, ensuring dimensionality matching during data processing and avoiding information loss or errors due to inconsistent dimensions. By obtaining sparse attention weights through differentiable sparse attention, it is possible to more accurately focus on features that significantly impact the prediction results. The feature update mechanism further optimizes the attention weights, resulting in the final attention fusion result, enabling the model to better utilize key information in the data. Finally, in the S400 stage, the meta-learner integrates the predictions from the base learner and the final attention fusion result, fully leveraging the advantages of ensemble learning and comprehensively considering multiple aspects of information to obtain a more accurate and reliable final prediction result, providing strong technical support for the precise fitting of orthokeratology lenses.

[0133] Specifically, the raw data is loaded from a CSV file, specifying feature names (feature_names) and corneal features (corneal_features). The dataset is then split into training and testing sets (80% training, 20% testing). Model initialization involves setting parameters such as the base learner (random forest), meta-learner (linear regression), feature names, corneal feature indices, and attention training epochs. The feature attention module and meta-attention module are initialized, and corneal-related feature indices are identified, i.e., anatomical features that are highly correlated with the prediction results are segmented based on prior medical knowledge. Where C represents the corneal feature index set and f represents the feature. In the orthokeratology lens fitting dataset: C={Flat_K,Steep_K,Eccentricity,Corneal_Diameter}.

[0134] Furthermore, S200 includes the following steps:

[0135] S201, the feature subset module presets groups, including corneal information group, biometrics group, refractive data group, and demographic group. The medical prior grouping strategy excludes one group at a time to generate a training subset.

[0136] S202, when the feature subset module does not have a preset group, dynamic feature sampling is enabled, and the number of features in the original input data is greater than the maximum number of base learners, the dynamic importance grouping strategy is enabled;

[0137] Calculate the initial importance of each original feature in the original input data;

[0138] When n > 100

[0139] Where n represents the number of samples in the original input data. This represents the feature importance score of the j-th original feature. This represents the Pearson correlation coefficient. The vector representing the j-th original feature. This is a real label;

[0140] When n≤100

[0141] in, The LassoCV coefficient represents the j-th original feature in the feature subset module;

[0142] Normalize the feature importance score for each original feature, based on... Choose the most important one. One primary feature is designated as the key feature, and the rest are secondary features.

[0143]

[0144] in, This represents the maximum number of base learners. This represents the total number of features in the original input data;

[0145] right For each key feature, generate several training subsets that exclude each key feature; for secondary features, generate a training subset that includes all secondary features; and generate a training subset for all original features.

[0146] Preferably, S201 involves systematically excluding specific groups to delve deeper into the impact of different group characteristics on orthokeratology lens fitting results, thereby providing a richer and more targeted dataset for subsequent model training. The dynamic importance grouping strategy in S202 dynamically calculates the importance of each original feature based on the number of samples and features in the original input data, and accordingly classifies key and secondary features. This strategy not only effectively filters out the features with the greatest impact on the model when there are numerous data features, but also accurately assesses the importance of features through model coefficients when the data sample size is small, ensuring the accuracy and reliability of feature selection.

[0147] Further, step S300 includes the following steps:

[0148] S301, the feature attention module processes corneal-related features in the original features of the original input data through a dedicated embedding layer, and uses a gating mechanism to dynamically enhance corneal-related features into enhanced corneal features;

[0149]

[0150] in, Indicates original features, Indicates corneal-related characteristics, Embedded representations of corneal-related features Indicates cross-enhancement features, Indicates the first gating weight. Indicates enhanced corneal features, This represents the first learnable weight matrix. This represents the second learnable weight matrix. This represents the third learnable weight matrix. This represents the fourth learnable weight matrix. This represents the fifth learnable weight matrix. This represents the first learnable bias vector. This represents the second learnable bias vector. This represents the third learnable bias vector. This represents the fourth learnable bias vector. This represents the fifth learnable bias vector. This represents the sigmoid function. Presentation layer normalization operation;

[0151] S302, the feature attention module processes the original input data through a dedicated embedding layer to obtain a conventional feature embedding. This will enhance corneal features. Compared with conventional feature embedding Add, The characteristics after fusion are obtained. ;

[0152] S303, for Take the average. corneal influence factor was obtained. ,in, This indicates the operation of taking the average value;

[0153] According to corneal impact factors Obtain the calibration feature weights ;

[0154]

[0155] in, This represents the sixth learnable weight matrix. This represents the seventh learnable weight matrix. This represents the sixth learnable bias vector. This represents the seventh learnable bias vector. Indicates the initial feature weights. Indicates the calibration feature weights, represents the learnable scaling factor, and Q represents the query vector.

[0156] Preferably, S301 is a corneal feature enhancement step that precisely captures corneal morphological parameters through a dedicated embedding layer. Combined with feature cross-enhancement technology, it effectively mines deeper features such as corneal surface curvature and astigmatism distribution, and utilizes a dynamic gating mechanism to achieve personalized feature enhancement. This processing method significantly improves the accuracy of identifying corneal abnormalities, especially demonstrating high sensitivity for feature extraction in complex cases such as keratoconus and irregular astigmatism. Simultaneously, layer normalization ensures the stability of data across different batches, providing a reliable feature foundation for subsequent analysis.

[0157] S302 is the feature fusion step, which spatially aligns and fuses enhanced corneal features with conventional biological parameters to form a multi-dimensional feature vector that includes morphological, biomechanical, and optical properties. This fusion strategy preserves the corneal-specific geometric information, making the feature representation more clinically interpretable.

[0158] S303 is a feature importance calibration step. The corneal influence factor constructed through averaging can quantitatively assess the contribution weight of corneal features in the overall fitting decision. This indicator can dynamically adjust the learning rate of different feature dimensions, automatically suppress noise feature interference during training, and strengthen the learning signal of key parameters (such as flat corneal curvature (Flat_K) and steep corneal curvature (Steep_K)).

[0159] Furthermore, the feature alignment layer of the feature attention module updates the input feature dimension of the original input data after calibrating the feature weights, specifically including the following steps:

[0160] S311, the input feature dimension of the original input data after calibrating the feature weights. The embedding layer target dimension of the feature alignment layer of the feature attention module. Make comparisons;

[0161] like If the scene is determined to be of equal dimensions, the feature alignment layer directly matches the dimension.

[0162] like If it is determined to be a dimensional upgrade scenario, then ,

[0163] in, This represents the output feature vector after feature alignment. This represents the alignment transformation function. Indicates zero-filling operation;

[0164] like If it is determined to be a dimensionality reduction scenario, then ,in, Represents the dynamic fusion coefficient. Indicates the preceding Dimensional original features, , This represents the L2 norm of the original eigenvector.

[0165] Ideally, the feature dimension comparison and corresponding processing in step S311 can more accurately capture the unique features of the cornea. Direct dimension matching in the same-dimensional scenario ensures the complete transmission of feature information and avoids information loss; the combination of zero-filling operation and alignment transformation function in the up-dimensional scenario effectively improves the expressive power of features, enabling the model to handle more complex corneal morphologies; while the dynamic fusion strategy in the down-dimensional scenario achieves effective feature compression and optimization by retaining key features and adjusting their weights. These processing steps together form the basis of the ensemble learning method based on corneal feature attention, providing strong support for personalized fitting of orthokeratology lenses.

[0166] Preferably, step S311 enables the input feature dimension of the original input data after calibrating the feature weights to match the target dimension expected by the feature alignment layer, allowing the feature attention module to accurately process the original input data after calibrating the feature weights, avoiding information loss during information processing, and improving the accuracy of information processing by the feature attention module.

[0167] Furthermore, the S300 also includes the following steps:

[0168] Set up a feature alignment layer protection mechanism. ,in, This represents the information loss rate. When the information loss rate exceeds a set threshold of 0.25, the feature alignment layer is frozen.

[0169] Preferably, the feature alignment layer protection mechanism freezes the feature alignment layer in a timely manner when the information loss rate approaches or exceeds a set threshold, ensuring the stability and reliability of feature information and providing solid technical support for the accurate fitting of orthokeratology lenses.

[0170] Furthermore, it also includes the following steps:

[0171] A course learning strategy is set in the feature attention module. When the input feature dimension matches the target dimension of the embedding layer, the feature attention module executes the course learning strategy to control the feature retention ratio of the feature attention module. The course learning strategy is as follows:

[0172]

[0173] in, Indicates the current training round. Indicates the total number of training rounds. Indicates the minimum feature retention ratio. This indicates the percentage of the maximum feature retained.

[0174] Preferably, the learning strategy dynamically adjusts the feature retention ratio. In the early stages of training, more features are retained to fully learn the diverse morphologies of the cornea. As training iterations increase, the feature retention ratio is gradually reduced, thus focusing on the most discriminative corneal features. This strategy not only helps prevent overfitting but also improves the model's adaptability to complex corneal morphologies.

[0175] Furthermore, the processing of differentiable sparse attention includes the following steps:

[0176] S321, Adjust the global interaction importance of the feature attention module;

[0177]

[0178] in, This represents batch matrix multiplication, where Q represents the query vector. Represents the key vector. This indicates the number of features in the original input data. Represents the global interaction matrix. This represents the element in the i-th row and j-th column of the global interaction matrix. Indicates the importance of global interactions;

[0179] S322, Calculate the final score of the calibration feature weights of the original input data:

[0180]

[0181] in, This represents the global weighted score. This represents the global weighted score of the i-th calibration feature weight. Indicates the final score;

[0182] S323, based on the final score, calculates the differentiable mask, sparse weights, and second gating value, and then calculates the sparse attention weights.

[0183] Ideally, stage S321, through global interaction matrix calculation, delves into the complex relationships between corneal features, providing crucial information for subsequent feature selection. Stage S322 quantifies the importance of each feature by calculating its final score, enabling the model to more accurately capture features critical to orthokeratology lens fitting. In stage S323, based on the final scores of the original features, the model further calculates a differentiable mask, sparse weights, and a second gating value. These parameters collectively influence the calculation of sparse attention weights, achieving refined selection and attention allocation of corneal features. This series of calculations not only improves the accuracy of orthokeratology lens fitting but also provides strong data support for subsequent ensemble learning methods.

[0184] Furthermore, such as Figure 2 As shown, S323 includes the following steps:

[0185] When calculating the differentiable mask, the dynamic K value is predicted based on the final score, and then the differentiable mask is calculated.

[0186] The dynamic K-value represents the number or proportion of key features that should ultimately be retained for each sample in the original input data. The formula for calculating the dynamic K-value is:

[0187]

[0188] in, This represents the eighth learnable weight matrix. This represents the ninth learnable weight matrix. This represents the eighth learnable bias vector. This represents the ninth learnable bias vector. This represents the feature retention ratio coefficient obtained after two layers of linear transformation and activation, ranging from (0,1), and is used for subsequent interpolation to obtain the final retention count. This indicates the number or proportion of key features that should ultimately be retained for each sample;

[0189] Add Gamper noise to the final score.

[0190]

[0191] in, This represents the noisy feature score after adding Gamper noise. Let represent a random variable that follows a uniform distribution on the interval (0,1). This represents a learnable temperature parameter;

[0192] TopK selection

[0193]

[0194] in, This represents a binary mask, indicating whether the j-th feature of the i-th sample is selected, with 1 indicating selection and 0 indicating non-selection. This represents the score of the i-th noisy feature; Represents the i-th sample ; This represents the TopK selection operation, which selects the top K original features with the highest scores from the input noisy feature scores, where K is determined by... Dynamic adjustment;

[0195] Calculate the differentiable mask.

[0196]

[0197] in, Indicates a soft mask. express mean This represents the differentiable mask ultimately used for end-to-end training. This represents the binary mask selected by TopK. This represents the gradient separation operation;

[0198] When calculating sparse weights, the scores are sorted according to the final scores, and then the support set and projection threshold are calculated to obtain the sparse weights.

[0199] Sort the final scores:

[0200] in, This represents the final score of the first-ranked item. This represents the final score of the second-ranked item. This represents the final score of the Nth rank. , This represents the final sequence of scores after sorting.

[0201] Compute the support set,

[0202]

[0203] in, Indicates the size of the candidate's support set. This represents the final score of the k-th sorted number. This represents the sum of the final scores of the top k largest values ​​after sorting.

[0204] Calculate the projection threshold :

[0205] Output sparse weights,

[0206] in, Indicates sparse weights;

[0207] Second gating value calculate:

[0208]

[0209] in, This represents the tenth learnable weight matrix. This represents the eleventh learnable weight matrix;

[0210] Performing gating fusion operations yields sparse attention weights:

[0211]

[0212] in, This represents sparse attention weights.

[0213] Ideally, differentiable masking computation, through operations such as dynamic K-value prediction, adding Gumble noise, and TopK selection, accurately filters out key information from the original features. This key information reflects the corneal characteristics across different dimensions and is crucial for accurately understanding corneal morphology. Sparse weight computation then sorts the original features based on their final scores to determine the support set and projection threshold, ultimately yielding the sparse weights. This process highlights the weights of features that significantly impact orthokeratology lens fitting while mitigating the interference of secondary features. Sparse attention weights can focus more intently on corneal features that are decisive for orthokeratology lens fitting, enabling the entire ensemble learning method to more accurately adapt to the corneal characteristics of different individuals, thus improving the accuracy and effectiveness of orthokeratology lens fitting.

[0214] Furthermore, such as Figure 3 As shown, a feature update mechanism is set in the feature attention module. The feature update mechanism uses the sparse attention weights to process the original features of the original input data to obtain the final attention fusion result, including the following steps:

[0215] Weight the query vector.

[0216] in, This represents the query vector after adjusting the sparse attention weights. Represents the sparse attention weights after dimension alignment;

[0217] Perform low-rank projection.

[0218]

[0219] in, Represents the low-rank projection , This represents the low-rank projection matrix of the query vector. This represents the key vector after low-rank projection. Represents the original key vector. The low-rank projection matrix representing the key vector. and The dimension is , , This represents the feature dimension of a single head in a multi-head attention mechanism;

[0220] Generate a dynamic mask.

[0221]

[0222] in, This represents a Boolean mask used to label features with sparse attention weights greater than 0.1; This represents the attention score. This represents the Softmax function, used to normalize attention scores to a probability distribution that sums to 1. This represents an element in a Boolean mask, used to mark the first... Whether the original features of the location are preserved. Indicates reservation;

[0223] Vector aggregation: ,in, Represents the context vector. Represents a value vector;

[0224] Feature replication and updating:

[0225]

[0226] in, The projection result representing the original features. Indicates the number of copies. Indicates the characteristics after replication. This represents the mask after expanding the dimensions. Indicates the updated features;

[0227] Residual gating weight fusion

[0228]

[0229] in, Indicates the residual gating weight. Indicates the characteristics after fusion. This represents the twelfth learnable weight matrix. This represents the twelfth learnable bias vector;

[0230] Calculate the final output of the attention-weighted calculation.

[0231] in, This represents the final output of the attention-weighted algorithm. This represents the thirteenth learnable weight matrix. This represents the thirteenth learnable bias vector;

[0232] The final attention fusion result is obtained by fusing the attention-weighted final outputs. .

[0233] Preferably, the base learner has a minimum protection threshold, which defaults to 0.25, to ensure that corneal features maintain appropriate importance during model training and prediction, consistent with prior knowledge in the medical field. For corneal features, if their sparse_weights are less than the minimum protection threshold, they are assigned the minimum protection threshold. This design ensures that the importance of corneal features in the model is not too low, while maintaining training stability and flexibility.

[0234] Ideally, the query vector is first weighted and then combined with sparse attention weights after dimension alignment to more accurately reflect feature importance. Next, a low-rank projection operation is performed to project the query and key vectors into a low-dimensional space, reducing computational complexity while preserving key feature information. During dynamic mask generation, features with sparse attention weights exceeding a certain threshold are flagged, ensuring accurate capture of important features. The vector aggregation step combines value vectors with context vectors to generate a more comprehensive feature representation. Feature copying and updating operations further enhance feature expressiveness by copying and expanding dimensions, allowing features to better adapt to different task requirements. The residual-gated weight fusion step introduces residual connections and gating mechanisms to dynamically adjust and fuse features. Finally, the final attention-weighted output is calculated, and multiple attention-weighted results are fused to obtain the final attention fusion result, providing a more accurate and reliable basis for orthokeratology lens fitting.

[0235] Furthermore, S400 includes the following steps:

[0236] S401, Input the base learner predictions into the meta-attention module, and calculate the sample-level weight matrix through the meta-attention module:

[0237] in, This represents the prediction matrix of the base learner. This represents the sample-level weight matrix, where F represents the number of base learners. This indicates the self-attention mechanism. This represents the fourteenth learnable weight matrix. This represents the fifteenth learnable weight matrix;

[0238] The base learner predictions are weighted using the sample-level weight matrix to obtain weighted base learner predictions. ;

[0239] in, This represents the sample-level weight matrix corresponding to the i-th base learner. This represents the predicted value of the i-th base learner, and the meta-learner is trained using a linear regression model.

[0240] S402, weighted base learner predictions Input the meta-learner, and the meta-learner will output the final prediction result;

[0241] S403, during step S402, the meta-attention module and the feature attention module perform collaborative optimization through a joint loss function, which is used to consider the error between the final prediction result output by the meta-learner and the true label, as well as the prediction contribution of the feature attention module.

[0242] The joint loss function is:

[0243]

[0244] in, Indicates the loss value. This represents the first learnable fusion parameter. This represents the second learnable fusion parameter. This indicates the actual label.

[0245] Ideally, the meta-attention module processes and highlights features more critical to orthokeratology lens fitting, thereby improving the model's ability to capture important features. Based on the difference between the model's predictions and the true labels, an appropriate loss function is used to calculate the loss value, which measures the accuracy of the model's predictions and provides direction for subsequent model optimization. The input is the weighted predictions of the base learners, which are trained using a linear regression model by default. During training, the meta-learner learns how to fuse and adjust the predictions of the base learners. The meta-attention module incorporates medical knowledge as a regularization term into the joint loss function, guiding the model to think according to medical priors. By continuously optimizing the model parameters, the output becomes the model's final prediction, which can more accurately guide the fitting of orthokeratology lenses.

[0246] It should be noted that the learnable weight matrix, learnable bias vector, and learnable fusion parameter mentioned above are parameters inherent to the model, and this invention does not impose specific limitations on them.

[0247] The superiority of the learning method of this invention was verified through comparative experiments. The comparative experiments were set up as follows: the learning method of this invention (Experiment 1), the base learner of this invention (Experiment 2), XGBoost (Experiment 3), LightGBM (Experiment 4), GradientBoosting (Experiment 5), SVR (Experiment 6), and Stacking (Experiment 7) were compared to the core parameters AC (adaptation arc) and Reduction (reduction) of orthokeratology lenses, and the R values ​​after the comparative experiments were statistically analyzed.2 The accuracy of the MAE data is shown in Tables 1 and 2. According to the comparative experiments, the learning method of the present invention has higher accuracy in the output results of AC (adaptation arc) and Reduction (reduction amplitude) compared with the existing algorithms.

[0248] Table 1 Comparison of AC (Adaptive Arc) Experiments

[0249]

[0250] Table 2 Comparison of Reduction (Amount of Decrease) Experiments

[0251]

[0252] Further analysis of these data reveals that, in the comparative experiments of AC (adaptive arc), the R of Experiment 1 (the learning method of this invention) is higher. 2 The R value reached 0.9103, higher than other experiments, and the MAE value was 0.2680, lower than most experiments. This indicates that the learning method of this invention can fit the data more accurately when predicting AC (adaptation arc), with a smaller error between the predicted and actual values. In the comparison experiment of Reduction, Experiment 1's R... 2 The value was 0.9594, the highest among all experiments, while the MAE value was 0.2001, also at a relatively low level, indicating that this method has a significant advantage in predicting reduction. Compared with existing algorithms, the learning method of this invention demonstrates higher accuracy in predicting both core parameters. This improvement in accuracy has important clinical significance for the fitting of orthokeratology lenses. Accurate AC (adaptation arc) and Reduction parameters can improve the fit between the orthokeratology lens and the patient's cornea, reduce discomfort during wear, and improve the corrective effect. At the same time, it also helps to reduce the risk of complications caused by inaccurate parameters, such as corneal damage and infection.

[0253] The advancement of the learning method of this invention is verified through ablation experiments. Experiments 11 to 19 below illustrate the ablation experiment setup method:

[0254] Experiment 11: The complete integrated learning method for orthokeratology lens fitting according to this invention;

[0255] Experiment 12: Disable the corneal feature enhancement operation (step S301) in the complete corneal reshaping lens fitting integrated learning method of the present invention, and treat the corneal features as ordinary features. The purpose is to evaluate the importance of special treatment of corneal features.

[0256] Experiment 13: Disabling sparse attention (steps S321 to S323) in the complete corneal reshaping lens fitting integrated learning method of this invention, and replacing sparse attention with a simple MLP (multilayer perceptron), with attention weights replaced by average weights, the purpose of which is to evaluate the attention mechanism;

[0257] Experiment 14: The meta-learner in the complete corneal reshaping lens fitting integrated learning method of this invention is disabled, and the base learner predictions are fused using fixed weights. The purpose is to evaluate the fusion effect of the meta-learner on the base learner predictions.

[0258] Experiment 15: Disable the training subset in the complete corneal reshaping lens fitting integrated learning method of the present invention (step S200), replace the feature subset with the complete feature set, and select a value of 8-12 for the maximum depth in the random forest parameters of the base learner;

[0259] Experiment 16: Disable the dynamic scaling adjustment (step S323) in the complete corneal reshaping lens fitting integrated learning method of the present invention, and use a fixed K value of 0.3 to evaluate the effect of dynamic K value adjustment;

[0260] Experiment 17: Disable the feature update mechanism in the complete corneal reshaping lens fitting ensemble learning method of this invention, and directly use the attention output;

[0261] Experiment 18: Disable feature importance calibration (S302 and S303) in the complete corneal reshaping lens fitting integrated learning method of this invention.

[0262] Experiment 19: The feature alignment layer protection mechanism in the complete corneal reshaping lens fitting integrated learning method of this invention was disabled in order to evaluate the necessity of the feature alignment layer protection mechanism;

[0263] The results of the ablation experiments are shown in Tables 3 and 4:

[0264] Table 3 AC ablation experimental data

[0265]

[0266] Table 4 Reduction ablation experimental data

[0267]

[0268] Ideally, as can be seen from the data in Tables 3 and 4, the complete model (Experiment 11) performs well in R... 2 It performed best in both the R and MAE metrics. Specifically, in the AC ablation experiment, the complete model achieved the best R... 2The R-value reached 0.9103, and the MAE value was 0.2680. Compared with other models lacking different components, it showed higher goodness of fit and lower prediction error. In the Reduction ablation experiment, the R-value of the complete model was... 2 Its value is as high as 0.9594, while its MAE value is only 0.2001, making it the best performing in all experiments.

[0269] Further analysis revealed that the absence of sparse attention (Experiment 13) and the lack of a feature update mechanism (Experiment 17) significantly impacted model performance. In the AC experiments, the R-value of the model without sparse attention was [missing data]. 2 The value decreased to 0.8965, and the MAE value increased to 0.2861; in the Reduction experiment, the R-value of the model without a feature update mechanism decreased. 2 The value dropped significantly to 0.9295, while the MAE value rose significantly to 0.2636. This indicates that sparse attention and feature update mechanisms play a crucial role in improving model performance.

[0270] Furthermore, models without corneal feature enhancement (Experiment 12), without meta-learners (Experiment 14), without training subset partitioning (Experiment 15), without dynamic scaling (Experiment 16), without feature importance calibration (Experiment 18), and without alignment protection (Experiment 19) also experienced a performance decrease, but the impact was relatively small. This indicates that while these components contribute to model performance, they are not the decisive factors.

[0271] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An ensemble learning method for orthokeratology lens fitting based on corneal feature attention, characterized in that, Includes the following steps: S100: Obtain the raw data and perform feature standardization to form the raw input data; S200, the original input data enters the feature subset module. The feature subset module divides the original input data into multiple training subsets through a medical prior grouping strategy or a dynamic importance grouping strategy. The base learner is trained using the multiple training subsets, and the base learner outputs the base learner prediction value. S300, the raw input data enters the feature attention module; The feature attention module performs corneal feature enhancement, feature fusion, and feature importance calibration operations on the corneal features in the original input data, thereby updating the initial feature weights corresponding to the original features of the original input data to the calibrated feature weights; The feature alignment layer of the feature attention module updates the input feature dimension of the original input data after calibrating the feature weights. The feature attention module obtains the sparse attention weights of the original input data after updating the input feature dimensions by using differentiable sparse attention; A feature update mechanism is set in the feature attention module. The feature update mechanism uses the sparse attention weights to process the original features of the original input data to obtain the final attention fusion result. S301, the feature attention module processes corneal-related features in the original features of the original input data through a dedicated embedding layer, and uses a gating mechanism to dynamically enhance corneal-related features into enhanced corneal features; in, Indicates original features, Indicates corneal-related characteristics, Embedded representations of corneal-related features Indicates cross-enhancement features, Indicates the first gating weight. Indicates enhanced corneal features, This represents the first learnable weight matrix. This represents the second learnable weight matrix. This represents the third learnable weight matrix. This represents the fourth learnable weight matrix. This represents the fifth learnable weight matrix. This represents the first learnable bias vector. This represents the second learnable bias vector. This represents the third learnable bias vector. This represents the fourth learnable bias vector. This represents the fifth learnable bias vector. This represents the sigmoid function. Presentation layer normalization operation; S302, the feature attention module processes the original input data through a dedicated embedding layer to obtain a conventional feature embedding. This will enhance corneal features. Compared with conventional feature embedding Add, The characteristics after fusion are obtained. ; S303, for Take the average. corneal influence factor was obtained. ,in, This indicates the operation of taking the average value; According to corneal impact factors Obtain the calibration feature weights ; in, This represents the sixth learnable weight matrix. This represents the seventh learnable weight matrix. This represents the sixth learnable bias vector. This represents the seventh learnable bias vector. Indicates the initial feature weights. Indicates the calibration feature weights, This represents the learnable scaling factor, and Q represents the query vector; S400 trains a meta-learner based on the base learner's predictions and the final attention fusion result, and the meta-learner outputs the final prediction result.

2. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 1, characterized in that, Step S200 includes the following steps: S201, the feature subset module presets groups, including corneal information group, biometrics group, refractive data group, and demographic group. The medical prior grouping strategy excludes one group at a time to generate a training subset. S202, when the feature subset module does not have a preset group, dynamic feature sampling is enabled, and the number of features in the original input data is greater than the maximum number of base learners, the dynamic importance grouping strategy is enabled; When n > 100 Where n represents the number of samples in the original input data. This represents the feature importance score of the j-th original feature. This represents the Pearson correlation coefficient. The vector representing the j-th original feature. This is a real label; When n≤100 in, The LassoCV coefficient represents the j-th original feature in the feature subset module; Normalize the feature importance score for each original feature, based on... Choose the most important one. One primary feature is designated as the key feature, and the rest are secondary features. in, This represents the maximum number of base learners. This represents the total number of features in the original input data; right For each key feature, generate several training subsets that exclude each key feature; for secondary features, generate a training subset that includes all secondary features; and generate a training subset for all original features.

3. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 2, characterized in that, The feature alignment layer of the feature attention module updates the input feature dimension of the original input data after calibrating the feature weights, specifically including the following steps: S311, the input feature dimension of the original input data after calibrating the feature weights. The embedding layer target dimension of the feature alignment layer of the feature attention module. Make comparisons; like If the scene is determined to be of equal dimensions, the feature alignment layer directly matches the dimension. like If it is determined to be a dimensional upgrade scenario, then , in, This represents the output feature vector after feature alignment. This represents the alignment transformation function. Indicates zero-filling operation; like If it is determined to be a dimensionality reduction scenario, then ,in, Represents the dynamic fusion coefficient. Indicates the preceding Dimensional original features, , This represents the L2 norm of the original eigenvector.

4. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 3, characterized in that, The S300 also includes the following steps: Set up a feature alignment layer protection mechanism. ,in, This represents the information loss rate. When the information loss rate exceeds a set threshold of 0.25, the feature alignment layer is frozen.

5. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 4, characterized in that, It also includes the following steps: A course learning strategy is set in the feature attention module. When the input feature dimension matches the target dimension of the embedding layer, the feature attention module executes the course learning strategy to control the feature retention ratio of the feature attention module. The course learning strategy is as follows: in, Indicates the current training round. Indicates the total number of training rounds. Indicates the minimum feature retention ratio. This indicates the percentage of the maximum feature retained.

6. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 5, characterized in that, Differentiable sparse attention processing includes the following steps: S321, Adjust the global interaction importance of the feature attention module; in, This represents batch matrix multiplication, where Q represents the query vector. Represents the key vector. This indicates the number of features in the original input data. Represents the global interaction matrix. This represents the element in the i-th row and j-th column of the global interaction matrix. Indicates the importance of global interactions; S322, Calculate the final score of the calibration feature weights of the original input data: in, This represents the global weighted score. This represents the global weighted score of the i-th calibration feature weight. Indicates the final score; S323, based on the final score, calculates the differentiable mask, sparse weights, and second gating value, and then calculates the sparse attention weights.

7. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 6, characterized in that, S323 includes the following steps: When calculating the differentiable mask, the dynamic K value is predicted based on the final score, and then the differentiable mask is calculated. The dynamic K-value represents the number or proportion of key features that should ultimately be retained for each sample in the original input data. The formula for calculating the dynamic K-value is: in, This represents the eighth learnable weight matrix. This represents the ninth learnable weight matrix. This represents the eighth learnable bias vector. This represents the ninth learnable bias vector. This represents the feature retention ratio coefficient obtained after two layers of linear transformation and activation, ranging from (0,1), and is used for subsequent interpolation to obtain the final retention count. This indicates the number or proportion of key features that should ultimately be retained for each sample; Add Gamper noise to the final score. in, This represents the noisy feature score after adding Gamper noise. Let represent a random variable that follows a uniform distribution on the interval (0,1). This represents a learnable temperature parameter; TopK selection in, This represents a binary mask, indicating whether the j-th feature of the i-th sample is selected, with 1 indicating selection and 0 indicating non-selection. This represents the score of the i-th noisy feature; Represents the i-th sample ; This represents the TopK selection operation, which selects the top K original features with the highest scores from the input noisy feature scores, where K is determined by... Dynamic adjustment; Calculate the differentiable mask. in, Indicates a soft mask. express mean This represents the differentiable mask ultimately used for end-to-end training. This represents the binary mask selected by TopK. This represents the gradient separation operation; When calculating sparse weights, the scores are sorted according to the final scores, and then the support set and projection threshold are calculated to obtain the sparse weights. Sort the final scores: in, This represents the final score of the first-ranked item. This represents the final score of the second-ranked item. This represents the final score of the Nth rank. , This represents the final sequence of scores after sorting. Compute the support set, in, Indicates the size of the candidate's support set. This represents the final score of the k-th sorted number. This represents the sum of the final scores of the top k largest values ​​after sorting. Calculate the projection threshold : Output sparse weights, in, Indicates sparse weights; Second gating value calculate: in, This represents the tenth learnable weight matrix. This represents the eleventh learnable weight matrix; Performing gating fusion operations yields sparse attention weights: in, This represents sparse attention weights.

8. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 7, characterized in that, A feature update mechanism is set up in the feature attention module. The feature update mechanism uses the sparse attention weights to process the original features of the original input data to obtain the final attention fusion result, including the following steps: Weight the query vector. in, This represents the query vector after adjusting the sparse attention weights. Represents the sparse attention weights after dimension alignment; Obtaining the low-rank projection , in, Represents the low-rank projection , This represents the low-rank projection matrix of the query vector. This represents the key vector after low-rank projection. Represents the original key vector. The low-rank projection matrix representing the key vector. and The dimension is , , This represents the feature dimension of a single head in a multi-head attention mechanism; Obtain attention score, in, This represents a Boolean mask used to label features with sparse attention weights greater than 0.1; This represents the attention score. This represents the Softmax function, used to normalize attention scores to a probability distribution that sums to 1. This represents an element in a Boolean mask, used to mark the first... Whether the original features of the location are preserved. Indicates reservation; Perform vector aggregation operation: ,in, Represents the context vector. Represents a value vector; Perform feature copy and update operations: in, The projection result representing the original features. Indicates the number of copies. Indicates the characteristics after replication. This represents the mask after expanding the dimensions. Indicates the updated features; Perform a residual gating weight fusion operation to obtain the fused features. in, Indicates the residual gating weight. Indicates the characteristics after fusion. This represents the twelfth learnable weight matrix. This represents the twelfth learnable bias vector; Calculate the final output of the attention-weighted calculation. in, This represents the final output of the attention-weighted algorithm. This represents the thirteenth learnable weight matrix. This represents the thirteenth learnable bias vector; The final attention fusion result is obtained by fusing the attention-weighted final outputs. .

9. The integrated learning method for orthokeratology lens fitting based on corneal feature attention according to claim 8, characterized in that, S400 includes the following steps: S401, Input the base learner predictions into the meta-attention module, and calculate the sample-level weight matrix through the meta-attention module: in, This represents the prediction matrix of the base learner. This represents the sample-level weight matrix, where F represents the number of base learners. This indicates the self-attention mechanism. This represents the fourteenth learnable weight matrix. This represents the fifteenth learnable weight matrix; The base learner predictions are weighted using the sample-level weight matrix to obtain weighted base learner predictions. ; in, This represents the sample-level weight matrix corresponding to the i-th base learner. This represents the predicted value of the i-th base learner, and the meta-learner is trained using a linear regression model. S402, weighted base learner predictions Input the meta-learner, and the meta-learner will output the final prediction result; S403, during step S402, the meta-attention module and the feature attention module perform collaborative optimization through a joint loss function, which is used to consider the error between the final prediction result output by the meta-learner and the true label, as well as the prediction contribution of the feature attention module. The joint loss function is: in, Indicates the loss value. This represents the first learnable fusion parameter. This represents the second learnable fusion parameter. This indicates the actual label.

Images