Child and adolescent myopia development trend prediction method based on multi-modal data fusion

Abstract

This invention relates to the fields of medicine and artificial intelligence, specifically to a method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion, aiming to achieve high-precision prediction. The method includes acquiring raw gene sequencing data files in FASTQ format; sequentially performing sequence alignment and variant detection on the raw gene sequencing data files to generate variant result files; then annotating each gene variant site according to the ACMG genetic variant classification standard for pathogenicity classification; subsequently, counting the number of variant sites corresponding to each type in the pathogenicity classification to obtain ACMG classification statistical features; horizontally concatenating the ACMG classification statistical features with preprocessed traditional Chinese medicine constitution data, refractive parameter data, and fundus structure data to construct a multimodal feature matrix; and inputting the multimodal feature matrix into a pre-trained ensemble learning model to output the probability of myopia occurrence and the predicted degree of myopia.

Description

Technical Field

[0001] This invention relates to the fields of medicine and artificial intelligence, and in particular to a method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion. Background Technology

[0002] Myopia is the most common type of refractive error among children and adolescents, and its incidence continues to rise globally, becoming a major public health problem. Accurately predicting the occurrence and development of myopia is crucial for early intervention and personalized prevention. Currently, myopia prediction methods mainly fall into the following categories: prediction methods based on single indicators (such as refractive error or axial length), which only consider physiological indicators of the eye and ignore comprehensive influencing factors such as genetics and constitution; risk factor analysis methods based on traditional statistical models (such as the Cox proportional hazards model and logistic regression), which assume linear relationships between features and struggle to capture complex nonlinear interactions; and prediction methods based on machine learning (such as random forests and support vector machines) or deep learning (such as convolutional neural networks), which, while making some progress in recent years, mostly use only single-modality information such as structured clinical data or fundus images, failing to effectively integrate multi-source heterogeneous data. Furthermore, traditional Chinese medicine (TCM) constitution identification, as an important component of the TCM theory of "prevention of disease," has shown unique value in the prevention of various diseases, but existing myopia prediction methods have not yet incorporated TCM constitution factors into their prediction models.

[0003] Existing technologies suffer from the following main shortcomings: First, they rely on a single data modality. Most existing methods only utilize one type of data from eye examinations or genetic data, failing to effectively integrate multimodal data such as TCM constitution, refractive parameters, fundus structure, and gene variations, resulting in limited prediction accuracy. Second, they face difficulties in processing heterogeneous data. TCM constitution data (categorical), refractive error data (continuous), and gene variation data (high-dimensional and sparse) have different data structures and feature distributions, and existing methods lack a unified mechanism for fusion of heterogeneous data. Third, they fail to reflect the characteristics of TCM. Existing myopia prediction methods do not incorporate TCM constitution factors into the model, failing to fully utilize the potential value of TCM constitution identification in myopia risk assessment. Fourth, they lack model interpretability. While deep learning methods offer high prediction accuracy, they lack reasonable explanations for the basis of prediction, making it difficult to meet the practical needs of clinicians. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion, aiming to construct a method for predicting the development trend of myopia in children and adolescents with good interpretability and high prediction accuracy.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The present invention provides a method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion, comprising: S1: acquiring raw gene sequencing data files, wherein the raw gene sequencing data files are in FASTQ format; S2: sequentially performing sequence alignment and variant detection on the raw gene sequencing data files to generate variant result files, and then annotating each gene variant site in the variant result files according to the ACMG genetic variant classification standard for pathogenicity classification, and then counting the number of variant sites corresponding to each type in the pathogenicity classification to obtain ACMG classification statistical features. The ACMG genetic variant classification standard classifies gene variant sites into five types: pathogenic, possibly pathogenic, of unclear significance, possibly benign, and benign; S3: horizontally splicing the ACMG classification statistical features with preprocessed TCM constitution data, refractive parameter data, and fundus structure data to construct a multimodal feature matrix; S4: inputting the multimodal feature matrix into a pre-trained ensemble learning model to output the probability of myopia occurrence and the predicted degree of myopia.

[0006] Step S2, after calculating the statistical features of ACMG classification, also includes: extracting high-frequency mutant gene features and human phenotype ontology keyword features, and using the ACMG classification statistical features, high-frequency mutant gene features, and human phenotype ontology keyword features together as input features for gene variation data.

[0007] The method for extracting high-frequency mutant gene features is as follows: all detected genes are sorted in descending order according to the mutation site frequency, and the top three genes are selected as high-frequency mutant gene features; the method for extracting human phenotype ontology keyword features is as follows: the total number of occurrences of human phenotype ontology keywords in each sample is counted.

[0008] The TCM constitution data in step S3 includes nine constitution types: balanced constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution, and special constitution. The preprocessing of the TCM constitution data includes converting the nine constitution types into numerical categorical features using label coding.

[0009] The refractive parameter data in step S3 includes equivalent spherical power, axial length, and corneal curvature; the preprocessing of the refractive parameter data includes filling missing values ​​with the median and performing Z-score standardization on all filled numerical features.

[0010] The fundus structure data in step S3 includes retinal thickness and optic disc area; the preprocessing of the fundus structure data includes filling missing values ​​with the median and performing Z-score standardization on all filled numerical features.

[0011] The horizontal stitching in step S3 is as follows: using the basic information table as the main table, which contains sample number, age, gender and refractive parameter data, ACMG classification statistical features, traditional Chinese medicine constitution data and fundus structure data are connected to the basic information table by left join through sample number to generate a unified multimodal feature matrix. Then, the spherical power and cylindrical power features are removed from the multimodal feature matrix.

[0012] The ensemble learning model in step S4 is the Light Gradient Boosting Machine model. The hyperparameters of the Light Gradient Boosting Machine model are optimized on the validation set using a grid search method. The validation set is obtained by stratifying and sampling the original dataset.

[0013] After outputting the probability of myopia occurrence and the predicted degree of myopia in step S4, the following steps are also included: calculating the importance score of each feature in the multimodal feature matrix using the average impurity reduction method, and identifying key risk factors based on the importance score.

[0014] Step S2 also includes: calculating the ratio of the number of variant sites with unclear significance to the total number of variant sites, obtaining the proportion of variants with unclear significance, and using the proportion of variants with unclear significance as a component of the statistical features of ACMG classification.

[0015] Compared with the prior art, this application has the following advantages: 1. This invention is the first to organically integrate four types of multi-source heterogeneous data: Traditional Chinese Medicine constitution data, refractive parameter data, fundus structure data, and gene variation data. A multimodal feature matrix is ​​constructed through lateral splicing, fully utilizing the complementarity between different modalities. Starting from the original FASTQ sequencing file, an end-to-end processing method is used to extract structured features such as ACMG five-class variant counts, high-frequency mutated genes, and human phenotypic ontology keywords through sequence alignment, variant detection, and ACMG pathogenicity classification annotation. In particular, this invention proposes using the proportion of variants of uncertain significance (VUS) as a predictive feature. Experiments have shown a significant negative correlation between the proportion of VUS and the degree of myopia in children and adolescents, with correlation coefficients ranging from -0.72 to -0.80. This finding provides new clues for the study of the genetic mechanism of myopia, and this feature has never been used in existing myopia prediction models.

[0016] 2. This invention incorporates nine constitution types (balanced, qi-deficient, yang-deficient, yin-deficient, phlegm-dampness, damp-heat, blood stasis, qi stagnation, and special constitution) from the "Classification and Determination of Traditional Chinese Medicine Constitutions" standard into a prediction model after labeling and encoding. Feature importance analysis shows that TCM constitution categories have an independent contribution to the probability of myopia occurrence. This not only verifies the scientific value of TCM's "prevention of disease" and constitution regulation theories in myopia prevention and control, but also provides a quantifiable technical tool for integrated TCM and Western medicine vision health management.

[0017] 3. This invention employs the LightGBM ensemble learning model, optimizing hyperparameters through stratified sampling and grid search, achieving high accuracy and robustness on the test set. Simultaneously, it uses the average impurity reduction method to calculate the importance score of each feature, intuitively demonstrating the contribution of key risk factors such as uncorrected visual acuity, axial length, gene variation characteristics, and TCM constitution category to the prediction results. This solves the "black box" problem of deep learning models, facilitating clinicians' understanding and adoption of the prediction basis.

[0018] 4. This invention not only outputs the probability of myopia occurrence and the predicted degree of myopia, but also unifies the multimodal feature matrix through a left-connection method and removes features such as spherical and cylindrical diopters that may cause label leakage, ensuring the model's generalization ability in real clinical environments. The entire method has clear steps and standardized data input, and can be directly integrated into hospital information systems or community vision screening platforms, providing decision support for early warning and personalized intervention of myopia in children and adolescents. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion, provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of the invention, it should be understood that the terms "upper," "lower," "left," "right," "front," "rear," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or relative positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Unless otherwise specified, the above-mentioned orientational descriptions can be flexibly set in practical applications, provided that the relative positional relationships shown in the accompanying drawings are satisfied.

[0022] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "communication" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection. They can refer to a direct connection or an indirect connection through an intermediate medium, or a communication between the internal components of two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] In embodiments of the invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, article, or apparatus that includes that element.

[0025] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0026] It should be noted that, in the embodiments provided in this application, "raw gene sequencing data file" refers to nucleotide sequence data obtained directly through high-throughput sequencing technology without any processing or analysis. Its standard format is FASTQ format, which simultaneously contains nucleotide sequence information (usually represented by the letters A, T, C, and G to represent the four bases) and the corresponding sequencing quality score. "Sequence alignment" refers to the process of locating short fragments (also called "reads") obtained from sequencing onto the human reference genome. A commonly used algorithm is the BurrowsWheeler Aligner (BWA)-Maximal Exact Matches (MEM). "Variation detection" refers to identifying sites that differ from the reference genome from the alignment results, including single nucleotide polymorphisms (SNPs) and insertion / deletion mutations (InDels). "ACMG genetic variation classification criteria" refers to the rule system published by the American College of Medical Genetics and Genomics (ACMG) that guides the rating of the pathogenicity of gene variations. "Horizontal concatenation" refers to merging multiple data tables along the column direction according to the same sample identifier (such as sample number) to form a wide table containing all feature information. "Ensemble learning model" refers to an algorithm that completes a learning task by constructing and combining multiple learners. Common examples include Random Forest (RF), Extremely Randomized Trees (ET), Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost), eXtreme Gradient Boosting (XGBoost), and Tabular Network (TabNet) based on attention mechanisms. This invention preferably uses the LightGBM model.

[0027] For example, such as Figure 1 As shown in the embodiments of this application, a method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion is provided, including: S1: Obtain the raw gene sequencing data file, which is in FASTQ format.

[0028] In some embodiments, whole-genome sequencing of peripheral blood samples from subjects is performed using a high-throughput sequencing platform (e.g., Illumina NovaSeq 6000) to obtain raw image data, which is then converted into a standard FASTQ format file using base calling software. Each record in a FASTQ file contains four lines: the first line begins with "@" followed by a sequence identifier; the second line is the nucleotide sequence (composed of A, T, C, and G); the third line begins with "+" and may be followed by the same sequence identifier; and the fourth line is the quality score string for the corresponding base.

[0029] S2: Sequence alignment and variant detection are performed sequentially on the original gene sequencing data file to generate a variant result file. Then, each gene variant site in the variant result file is annotated according to the ACMG genetic variant classification standard for pathogenicity. The number of variant sites corresponding to each type in the pathogenicity classification is counted to obtain the ACMG classification statistical characteristics. The ACMG genetic variant classification standard classifies gene variant sites into five types: pathogenic, possibly pathogenic, of uncertain significance, possibly benign, and benign.

[0030] For example, step S2 further includes: calculating the ratio of the number of variant sites with unclear significance to the total number of variant sites to obtain the proportion of variants with unclear significance, and using the proportion of variants with unclear significance as a component of the statistical features of ACMG classification.

[0031] In some embodiments, the FASTQ file obtained in step S1 is used as input. First, the sequence alignment tool BWA-MEM (BurrowsWheeler Aligner with Maximal Exact Matches) is used to align each read to the human reference genome GRCh38, generating an intermediate file in SAM format (Sequence Alignment Map). This intermediate file is then sorted and deduplicated before being converted to BAM format (Binary Alignment Map). Next, the HaplotypeCaller module of the GATK (Genome Analysis Toolkit) is used to detect variants, identifying single nucleotide polymorphism sites and insertion / deletion mutation sites in each sample relative to the reference genome. The original VCF format (VariantCall Format) file is output, and then low-quality variants are filtered out using Variant Quality Score Recalibration (VQSR) to obtain the final variant result file (VCF).

[0032] For each variant site in the VCF file, annotation was performed using the Variant Effect Predictor (VEP), specifically applying the 28 criteria outlined in the ACMG / AMP (American College of Medical Genetics and Genomics / Association for Molecular Pathology) joint guidelines. After annotation, each variant site was categorized into one of five classes: Pathogenic (P), Likely Pathogenic (LP), Variant of Uncertain Significance (VUS), Likely Benign (LB), and Benign (B). The number of variant sites belonging to each of these classes in each sample was counted and denoted as follows: , , , , These together constitute the statistical features of ACMG classification. For example, for a given sample, =2, =5, =123, =8, =45 For example, after calculating the statistical features of ACMG classification in step S2, the method further includes: extracting high-frequency mutant gene features and human phenotype ontology keyword features, and using the ACMG classification statistical features, high-frequency mutant gene features, and human phenotype ontology keyword features together as input features for gene variation data.

[0033] As one possible approach, the extraction method for high-frequency mutant gene features is as follows: all detected genes are sorted in descending order according to the mutation site frequency, and the top three genes are selected as high-frequency mutant gene features; the extraction method for human phenotype ontology keyword features is as follows: the total number of occurrences of human phenotype ontology keywords in each sample is counted.

[0034] In some embodiments, after completing the ACMG classification statistics, the following operations are performed: For all gene loci identified by variant detection, calculate the total number of mutations occurring at each gene (cumulative summation of mutations at different locations within the same gene). Sort all genes by this total number of occurrences from highest to lowest, and select the top three gene names. These three genes are the high-frequency mutation gene features, for example, TOP1 = "ACTN4", TOP2 = "COL11A1", TOP3 = "ZNF469". If there are ties, prioritize genes known to be relevant to ophthalmology (this can be determined using databases such as OMIM and ClinVar). For machine learning modeling, convert these three gene names into numerical vectors using one-hot encoding or label encoding.

[0035] More specifically, let the set of all detected variant sites in a sample be . Each mutation site Belongs to a certain gene ,in This is the human reference gene set (approximately 20,000 genes). (Definition of a gene) mutation frequency ,in This is an indicator function (1 if the condition is true, 0 otherwise). All genes are sorted... Sort in descending order to obtain a sorted list Take the first three As a characteristic of high-frequency mutated genes. If If there are ties, the association scores of these genes with ophthalmic diseases in the ClinVar database are compared, and the gene with the highest score is selected first.

[0036] Let each variant site be annotated Associate a set of HPO terms ,in This is a standard HPO identifier (e.g., HP:0000545). The set of genome-wide associated HPO terms is defined as follows: HPO keyword characteristics, i.e. This represents the total number of HPO terms after deduplication. To avoid duplicate counting, HPO terms from all loci are first merged, and then the number of unique terms is calculated. For example, if two different variant loci are both associated with "HP:0000545", they will only be counted once in the total count. This value reflects the diversity of abnormal phenotypes that an individual may exhibit at the genetic level and is correlated with the degree of myopia.

[0037] When using VEP annotation, HPO terms associated with variant sites are collected. HPO is a standardized vocabulary used to describe abnormal human phenotypes, such as HP:0000545 (myopia). The total number of unique HPO keywords associated with all variant sites in each sample is counted, denoted as . For example, if the variation in a sample involves three different HPO terms: "myopia," "cataract," and "retinitis pigmentosa," then... =3.

[0038] Classification statistical characteristics of ACMG ( , , , , The high-frequency mutant gene features (the coding values ​​of TOP1, TOP2, and TOP3) and the HPO_Count feature together serve as the complete input features for the gene variation data. These features are then combined with data on traditional Chinese medicine constitution, refractive parameters, and fundus structure to enter the stitching process.

[0039] S3: The ACMG classification statistical features are horizontally spliced ​​with the preprocessed TCM constitution data, refractive parameter data, and fundus structure data to construct a multimodal feature matrix.

[0040] The TCM constitution data in step S3 includes nine constitution types: balanced constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution, and special constitution. The preprocessing of the TCM constitution data includes converting the nine constitution types into numerical categorical features using label coding.

[0041] In some embodiments, a standardized questionnaire was used to collect the subjects' constitution types according to the "Classification and Determination of Traditional Chinese Medicine Constitutions" standard published by the China Association of Traditional Chinese Medicine. The questionnaire contained 60 items, corresponding to nine constitutions: balanced, qi deficiency, yang deficiency, yin deficiency, phlegm-dampness, damp-heat, blood stasis, qi stagnation, and special constitution. Based on the converted score (raw score - number of items) / (number of items × 4) × 100, a score ≥ 40 and a converted score ≥ 60 for balanced constitution was determined as balanced constitution; a score ≥ 40 and a converted score ≥ 40 for other unbalanced constitutions were determined as the corresponding unbalanced constitution. The nine constitutions were mapped to numerical values: balanced constitution as 0, qi deficiency as 1, yang deficiency as 2, yin deficiency as 3, phlegm-dampness as 4, damp-heat as 5, blood stasis as 6, qi stagnation as 7, and special constitution as 8. If a subject possessed multiple constitutions (mixed constitutions), the highest score was taken as the primary constitution. When missing, the default filler is a neutral quality (coded 0).

[0042] The refractive parameter data in step S3 includes equivalent spherical power, axial length, and corneal curvature; the preprocessing of the refractive parameter data includes filling missing values ​​with the median and performing Z-score standardization on all filled numerical features.

[0043] In some embodiments, an automated computer refractometer (such as TOPCON KR-800) is used to measure the spherical power (S) and cylindrical power (C) of the right and left eyes, and the spherical equivalent (SE) is calculated as S + C / 2. Axial length (AL, in mm) and corneal curvature (K1, K2, in D) are measured using an optical coherence biometer (such as ZEISS IOLMaster 700). Uncorrected visual acuity (UVA) is measured using a visual acuity chart. For numerical features, if missing values ​​exist (e.g., a subject did not cooperate with the axial length measurement), the median of that feature in the entire sample is used to fill in the missing values; subsequently, all numerical features (including SE, AL, K1, K2, and UVA) are Z-score standardized using the formula: in, These are the original values. This is the sample mean of this feature. This represents the sample standard deviation.

[0044] The fundus structure data in step S3 includes retinal thickness and optic disc area; the preprocessing of the fundus structure data includes filling missing values ​​with the median and performing Z-score standardization on all filled numerical features.

[0045] In some embodiments, fundus cameras such as the Canon CR-2 are used to acquire images of the posterior pole of the eye, and built-in software is used to automatically segment and extract retinal thickness (RT, in μm) and optic disc area (ODA, in mm). 2 Missing values ​​were also filled with the median, and the filled data were Z-score standardized.

[0046] As one possible implementation, the horizontal stitching in step S3 is as follows: using the basic information table as the main table, which contains sample number, age, gender and refractive parameter data, ACMG classification statistical features, traditional Chinese medicine constitution data and fundus structure data are connected to the basic information table by left join through sample number to generate a unified multimodal feature matrix. Then, the two features of spherical power and cylindrical power are removed from the multimodal feature matrix.

[0047] For example, a basic information table is established, which includes sample number (unique identifier), age (years, integer), gender (male code 0, female code 1), and standardized values ​​of refractive parameters obtained in step 2. Using this basic information table as the main table, the ACMG classification statistical features obtained in step S2 are... , , , , The TCM constitution codes (integers from 0 to 8) and standardized fundus structure features (RT, ODA) obtained in the previous steps are joined to the main table using a left join (LEFT JOIN) in SQL, based on the sample numbers. For samples that cannot be matched, the feature values ​​are filled with 0 or the mean. After joining, the original features of spherical and cylindrical power are removed because they already partially express refractive information through equivalent spherical power, and directly using the original spherical and cylindrical power may lead to label leakage (since equivalent spherical power is one of the prediction targets). Finally, a unified multimodal feature matrix is ​​generated, where rows represent samples and columns represent features (including age, gender, standardized refractive parameters, TCM constitution codes, ACMG five-class counts, and standardized fundus structure parameters).

[0048] For example, if spherical and cylindrical diopter exist in the basic information table, after generating the initial feature matrix, the two original features, spherical diopter (S) and cylindrical diopter (C), are removed. Since the equivalent spherical diopter already contains the information of S and C, and the equivalent spherical diopter is the target variable for regression prediction in this invention, if S and C are retained, the model may directly use a linear combination of S and C to approximate the equivalent spherical diopter, causing label leakage. This leads to overfitting on the training set and a decrease in generalization ability on new data.

[0049] The final multimodal feature matrix has the following dimensions: the actual number of columns after removing two features. It consists of: 1 column for sample ID, 2 columns for age and gender, 4 columns for standardized refractive parameters (SE, AL, K1, K2), 1 column for TCM constitution, 5 columns for ACMG five-category counts, 3 columns for high-frequency gene features, 1 column for HPO_Count, and 2 columns for fundus structures. This multimodal feature matrix can be directly used for training machine learning models.

[0050] S4: Input the multimodal feature matrix into the pre-trained ensemble learning model and output the probability of myopia occurrence and the predicted degree of myopia.

[0051] As one possible implementation, the ensemble learning model in step S4 is the Light Gradient Boosting Machine model; the hyperparameters of the Light Gradient Boosting Machine model are optimized on the validation set using a grid search method, and the validation set is obtained by stratified sampling of the original dataset.

[0052] The multimodal feature matrix generated in step S3 is used as training data. First, a stratified sampling method is used to divide the dataset into a training set (60%), a validation set (20%), and a test set (20%) based on the myopia label (whether myopic or not), ensuring that the ratio of myopic to emmetropic samples in each subset is consistent with the original dataset. Then, the LightGBM (Light Gradient Boosting Machine) model is used for training. LightGBM is an efficient implementation of Gradient Boosting Decision Tree (GBDT), featuring histogram algorithm and Gradient-based One-Side Sampling (GOSS). Hyperparameter optimization uses grid search to evaluate candidate hyperparameter combinations on the validation set. Example hyperparameter search range: learning rate ∈ {0.01, 0.05, 0.1}, tree depth (max_depth) ∈ {3, 5, 7}, number of leaf nodes (num_leaves) ∈ {31, 63, 127}. ROC-AUC (Area Under the Receiver Operating Characteristic) was used as the evaluation metric for the classification task, and Mean Squared Error (MSE) was used as the evaluation metric for the regression task. After training, the model performance was evaluated on the test set: the classification accuracy reached 96.67%, the ROC-AUC reached 0.9980, the regression MSE was 0.0862, and the Mean Absolute Error (MAE) was 0.1694.

[0053] For a new sample to be predicted, the same multimodal feature vector is obtained by repeating steps S1 to S3 above. This vector is then input into the trained LightGBM model. The model outputs a probability value of myopia occurrence between 0 and 1 (e.g., 0.86 indicates an 86% chance of developing myopia in the future) and a continuously represented predicted equivalent spherical power (e.g., -1.25 D). Risk levels are categorized based on the probability value: <0.3 indicates low risk, 0.3~0.7 indicates medium risk, and >0.7 indicates high risk.

[0054] As one possible implementation, after outputting the probability of myopia occurrence and the predicted degree of myopia in step S4, the method further includes: calculating the importance score of each feature in the multimodal feature matrix using the average impurity reduction method, and identifying key risk factors based on the importance score.

[0055] The Mean Decrease Impurity (MDI) method is used to calculate the importance score of each feature. For decision tree ensemble models such as LightGBM, each node uses an impurity index when splitting (Gini impurity is commonly used for classification tasks, and mean squared error (MSE) gain is commonly used for regression tasks). The importance of a feature is equal to the sum of the impurity reductions brought about by that feature participating in splits across all trees, divided by the total impurity reduction of all features, to obtain a normalized score (range 0-1).

[0056] For example, for a node in a tree The dataset before splitting was After splitting, it is divided into a left subset. and right subset The amount of impurity reduction is: in, Impurity function (Classification: ,return: ); These represent the number of samples for node t, its left child node, and its right child node, respectively. Features Importance: in, Indicates all used features The set of nodes to be split.

[0057] After calculating the importance of all features and sorting them in descending order, this invention found that features such as uncorrected visual acuity, axial length, corneal curvature, percentage of unsigned variation (VUS), and TCM constitution category scored higher. Clinicians can use this list to quickly identify the factors that have the greatest impact on an individual's myopia risk, enabling precise intervention.

[0058] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0059] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion, characterized in that, include: S1: Obtain the raw gene sequencing data file, wherein the raw gene sequencing data file is a FASTQ format file; S2: Sequence alignment and variant detection are performed sequentially on the original gene sequencing data file to generate a variant result file. Then, each gene variant site in the variant result file is annotated according to the ACMG genetic variant classification standard for pathogenicity. The number of variant sites corresponding to each type in the pathogenicity classification is counted to obtain ACMG classification statistical features. The ACMG genetic variant classification standard classifies gene variant sites into five types: pathogenic, possibly pathogenic, of unclear significance, possibly benign, and benign. S3: The ACMG classification statistical features are horizontally concatenated with the preprocessed TCM constitution data, refractive parameter data, and fundus structure data to construct a multimodal feature matrix. S4: The multimodal feature matrix is ​​input into a pre-trained ensemble learning model to output the probability of myopia occurrence and the predicted degree of myopia.

2. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, After calculating the statistical features of ACMG classification in step S2, the method further includes: extracting high-frequency mutant gene features and human phenotype ontology keyword features, and using the ACMG classification statistical features, the high-frequency mutant gene features, and the human phenotype ontology keyword features as input features for gene variation data.

3. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 2, characterized in that, The method for extracting the high-frequency mutant gene features is as follows: all detected genes are sorted in descending order according to the mutation site frequency, and the top three genes are selected as high-frequency mutant gene features. The method for extracting the human phenotype ontology keyword features is as follows: count the total number of occurrences of human phenotype ontology keywords in each sample.

4. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, The TCM constitution data in step S3 includes nine constitution types: balanced constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution, and special constitution. The preprocessing of the TCM constitution data includes: using label encoding to convert the nine constitution types into numerical categorical features.

5. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, The refractive parameter data in step S3 includes equivalent spherical power, axial length, and corneal curvature; the preprocessing of the refractive parameter data includes filling missing values ​​with the median and performing Z-score standardization on all filled numerical features.

6. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, The fundus structure data in step S3 includes retinal thickness and optic disc area; The preprocessing of the fundus structure data includes: filling missing values ​​with the median and performing Z-score standardization on all the filled numerical features.

7. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, The horizontal stitching in step S3 specifically involves: using the basic information table as the main table, which contains sample number, age, gender, and refractive parameter data, connecting the ACMG classification statistical features, the TCM constitution data, and the fundus structure data to the basic information table via sample number using a left join method to generate a unified multimodal feature matrix, and then removing the spherical power and cylindrical power features from the multimodal feature matrix.

8. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, The ensemble learning model in step S4 is the Light Gradient Boosting Machine model; the hyperparameters of the Light Gradient Boosting Machine model are optimized on the validation set using a grid search method, and the validation set is obtained by stratified sampling of the original dataset.

9. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, After outputting the probability of myopia occurrence and the predicted degree of myopia in step S4, the method further includes: calculating the importance score of each feature in the multimodal feature matrix using the average impurity reduction method, and identifying key risk factors based on the importance score.

10. The method for predicting the development trend of myopia in children and adolescents based on multimodal data fusion according to claim 1, characterized in that, Step S2 further includes: calculating the ratio of the number of variant sites with unclear significance to the total number of variant sites to obtain the proportion of variants with unclear significance, and using the proportion of variants with unclear significance as a component of the statistical features of ACMG classification.

Images