AI-driven early risk assessment method for glaucoma

Abstract

This invention discloses an AI-driven method for early risk prediction and assessment of glaucoma, belonging to the field of early glaucoma prediction technology. This invention overcomes the limitations of traditional single-modality methods due to insufficient information by utilizing multimodal ophthalmological examination information. It introduces cross-modal fusion and multi-head self-attention mechanisms to jointly model in both time and modality dimensions, enabling the model to adaptively identify the differences in the contribution of different time points and examination modalities to risk prediction, fully exploring the deep semantic relationships between multimodal data. Combined with a conditional diffusion decoder for prospective prediction, it can output the changes in retinal nerve fiber layer thickness and glaucoma progression risk scores for the next six and twelve months, providing clinicians with prospective decision support information. Furthermore, by employing a contrastive learning strategy, it enhances the model's ability to distinguish between those with and without progression, effectively improving the model's generalization performance under imbalanced sample conditions.

Description

Technical Field

[0001] This invention relates to the field of early glaucoma prediction technology, specifically to an artificial intelligence-driven method for early glaucoma risk prediction and assessment. Background Technology

[0002] Glaucoma is the leading cause of irreversible blindness worldwide. Its pathological characteristics are primarily progressive apoptosis of retinal ganglion cells and irreversible damage to optic nerve fibers, ultimately leading to visual field defects and even complete blindness. According to the World Health Organization, the number of glaucoma patients worldwide has exceeded 80 million, and is projected to surpass 110 million by 2040. Early stages of glaucoma often present with no obvious symptoms. By the time visual field defects appear, patients have often already lost a significant amount of optic nerve fibers. Therefore, early risk assessment is of paramount clinical importance in delaying disease progression and reducing the rate of blindness.

[0003] Currently, clinical diagnosis and assessment of glaucoma primarily rely on intraocular pressure (IOP) measurement, visual field testing, optic disc morphology assessment, and optical coherence tomography (OCT). IOP, as the most important risk factor for glaucoma, is affected by various factors such as central corneal thickness and the time of measurement, making it difficult to accurately assess an individual's risk based solely on IOP values. While visual field testing directly reflects the degree of visual impairment, its results are significantly influenced by patient cooperation and learning effects, and its sensitivity is insufficient in the early stages of the disease. OCT can quantitatively measure retinal nerve fiber layer thickness and has a high detection rate for early structural changes, but single-modality structural data cannot comprehensively reflect the complex multifactorial pathogenesis of glaucoma. Anterior chamber angle examination can help determine the type of glaucoma, but its results are mostly qualitative and difficult to use directly for quantitative risk assessment. Therefore, existing single-examination methods all have their limitations and cannot provide a comprehensive and accurate quantitative assessment of early glaucoma risk from multiple dimensions, including structure, function, and anatomy. Summary of the Invention

[0004] To address the aforementioned technical issues and provide an AI-driven method for early risk assessment of glaucoma, this technical solution resolves the problems raised in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the invention, an artificial intelligence-driven method for early risk prediction and assessment of glaucoma is also provided, comprising: Acquire multimodal ophthalmic examination data of the target patients, including structural optical coherence tomography image sequences, functional visual field test data sequences, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data; Standardized preprocessing was performed on multimodal ophthalmological examination data to convert data from different sources and formats into a standardized data format that is compatible with the input of artificial intelligence models. Denoising, missing value imputation, and outlier removal were performed on each modality of data. Feature extraction was performed on the preprocessed multimodal ophthalmological examination data to obtain structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features. The extracted modal features are fused across modalities to generate a joint latent space representation vector, which encodes the comprehensive representation information of the target patient's ocular structure and function. The joint latent space representation vector is input into the trained early glaucoma risk prediction model, and the output is the early glaucoma risk assessment result of the target patient. The risk assessment result includes the probability value of glaucoma progression and the risk level determination.

[0006] Preferably, acquiring the multimodal ophthalmological examination data of the target patient specifically includes: Structural optical coherence tomography (SOCT) image sequences of target patients were extracted from the hospital information system. The SOCT image sequences included retinal nerve fiber layer thickness maps, optic disc scan images, and macular scan images. Each image sequence contained follow-up images at no less than three time points, with varying follow-up intervals between the time points. The functional visual field test data sequence of the target patient is exported from the visual field testing device. The functional visual field test data sequence contains the test results of the Humphrey perimeter 24-2 procedure. The test results are stored in the form of a spatial distribution matrix of sensitivity thresholds. Each element in the matrix corresponds to the light sensitivity value of a test point in the visual field test grid. Intraocular pressure (IOP) data were collected from the target patients. The IOP data included IOP values ​​measured by a non-contact tonometer and an IOP value measured by a Goldmann applanation tonometer, as well as central corneal thickness data obtained by an ultrasonic pachymeter. Obtain color fundus images of the target patient, covering the optic disc region and macular region, with a resolution of no less than 2000×2000 pixels; Acquire anterior chamber angle examination data of the target patient. The anterior chamber angle examination data includes the grading data of the degree of angle opening and closing of the anterior chamber angle obtained by gonioscopy. The grading data is recorded as text description and corresponding angle value according to the Shaffer grading standard. ; in, This represents the numerical encoding result of the Shaffer classification, with a value range of [value range missing]. , This indicates the angle of the anterior chamber as recorded by gonioscopy.

[0007] Preferably, the standardization preprocessing of the multimodal ophthalmic examination data specifically includes: The structural optical coherence tomography (OCT) image sequence was format-converted from DICOM to NIfTI format. Speckle noise removal was then applied to the converted images using a wavelet transform-based denoising method. After multi-scale decomposition, soft thresholding was performed on the high-frequency subband coefficients. The formula for soft thresholding is as follows: ; in, Indicates the first Layer decomposition Wavelet coefficients at each position after soft thresholding Represents the original wavelet coefficients. The threshold parameter is determined by the following formula: ; in, The noise standard deviation is estimated using the median absolute deviation of the wavelet coefficients at the finest scale. and These represent the number of rows and columns of the image, respectively. Missing values ​​were detected and imputed in the spatial distribution matrix of sensitivity thresholds in the functional visual field test data sequence. For local missing regions caused by fixation loss in patients, an inverse distance-weighted interpolation method based on eight surrounding test points was used for imputation. The interpolation calculation formula is as follows: ; in, This indicates the sensitivity threshold after filling in the blanks. Indicates the first The original sensitivity threshold of each adjacent test point Indicates the first Euclidean distance from each adjacent test point to the missing location This represents the distance weighting index, with a value of 2. Outlier detection was performed on the intraocular pressure measurement data, and the upper and lower bounds of the intraocular pressure values ​​were calculated using the interquartile range method. The formulas for calculating the upper and lower bounds are as follows: ; ; in, This represents the first quartile of the intraocular pressure value sequence. The third quartile represents the intraocular pressure value sequence, and IQR represents the interquartile range. and These represent the lower and upper limits of intraocular pressure, respectively. and These represent the functions for calculating the 25th and 75th percentiles, respectively, and will exceed [ Intraocular pressure values ​​within a certain range are marked as abnormal, and these abnormal values ​​are corrected using central corneal thickness data. The correction formula is as follows: ; in, This indicates the corrected intraocular pressure value. This represents the original measured intraocular pressure value. This represents the corneal thickness correction factor, with a value of 0.013. This indicates the standard central corneal thickness, with a value of 520 micrometers. This indicates the actual measured central corneal thickness; Brightness normalization was performed on the color fundus photographs. Based on a brightness threshold, the fundus brightness grayscale image was divided into black and white images to separate the optic disc and vascular regions. The brightness of the entire image was then linearly adjusted according to the brightness threshold parameters of the capturing device. The brightness adjustment formula is as follows: ; in, Indicates the normalized position Pixel brightness value at that location, Indicates the original image location Pixel brightness value at that location, Indicates the brightness scaling factor. This indicates the brightness offset. and The value is determined by fitting the brightness threshold parameter of the shooting device using the least squares method; The anterior chamber angle examination data were structured and encoded, converting the textual descriptions of the Shaffer classification into numerical codes from one to four, while retaining the angle values ​​as floating-point data, thus forming a structured anterior chamber angle feature table.

[0008] Preferably, the feature extraction of the preprocessed multimodal ophthalmic examination data specifically includes: A convolutional neural network was used to extract features from the retinal nerve fiber layer thickness map in a structural optical coherence tomography (OCT) image sequence. The set of outputs from the convolutional and pooling layers of three lines was used to extract the color features, edge features, and texture features of the retinal nerve fiber layer, as well as the correlation features between these features. The color features were composed of the output of a pooling layer of the first line and a convolutional layer of the second line. The texture features were composed of the output of a zero convolutional layer of the third line, a convolutional layer of the second line, and a convolutional layer of the third line. The three sets of features were concatenated and mapped through an activation function to obtain the structural image feature vector. Statistical feature calculations were performed on the spatial distribution matrix of sensitivity thresholds in the functional visual field detection data sequence to extract the location features, depth features, and extent features of visual field defects. The slope and fluctuation features of sensitivity values ​​at each test point over time were also calculated. The functional visual field feature vector was obtained by concatenating the above five types of features. The formula for calculating the slope characteristic of the sensitivity value of each test point changing with time is as follows: ; in, Indicates the first The slope of sensitivity change at each test point This indicates the total number of follow-up time points. Indicates the first The test point at the ... Sensitivity threshold at each time point This represents the mean at a given time point. Indicates the first The mean sensitivity of each test point at all time points; The formula for calculating the fluctuation characteristics of sensitivity values ​​at each test point is as follows: ; in, Indicates the first Sensitivity fluctuation values ​​at each test point; Time-domain statistical analysis was performed on the intraocular pressure measurement data to extract five numerical features: daily average, daily peak, daily trough, intraocular pressure fluctuation amplitude, and diurnal intraocular pressure difference. The five numerical features were normalized and then spliced ​​together to obtain the intraocular pressure numerical features. The formula for calculating the amplitude of intraocular pressure fluctuation is as follows: ; in, This indicates the amplitude of intraocular pressure fluctuations. Indicates the first Corrected intraocular pressure values ​​at each time point This indicates the total number of time points used for intraocular pressure measurement; The formula for calculating the diurnal intraocular pressure difference is: ; in, This represents the diurnal intraocular pressure difference. Indicates the number of measurements taken during the day. Indicates the number of measurements taken at night; An image feature extraction network was used to extract features from color fundus photographs. The optic disc cup-to-disc ratio, retinal vessel tortuosity, and retinal hemorrhage density were extracted through multi-layer convolution and pooling operations. The fundus image features were obtained by concatenating the above three features. The formula for calculating the retinal vessel tortuosity characteristics is as follows: ; in, Indicates the curvature of retinal vessels. This indicates the total number of blood vessel segments detected. Indicates the first The actual path length of the segmental blood vessel Indicates the first The straight-line distance between the two ends of a segment of blood vessel; The formula for calculating the density characteristics of retinal hemorrhages is: ; in, Indicates the density of retinal hemorrhages. This indicates the total number of bleeding points detected. This indicates the area of ​​the effective analysis region in the fundus photograph; Feature encoding was performed on the anterior chamber angle examination data, and three structured features were extracted: the angle width value, the trabecular mesh visibility level, and the iris morphology classification. The three features were then individually encoded and spliced ​​to obtain the anterior chamber angle structural features. The conversion formula for one-hot encoding is as follows: ; in, Indicates the first The value of the one-hot encoded bit, The range of values ​​is , This represents the total number of categories for this classification feature.

[0009] Preferably, the step of fusing the extracted modal features across modalities to generate a joint latent space representation vector specifically includes: Structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features are mapped to a feature space of a unified dimension through fully connected layers, so that the dimensions of each modality feature vector are consistent. The formula for calculating the mapping of fully connected layers is: ; in, Indicates the first Feature vectors after modal mapping Indicates the first The original feature vectors of each modality, Indicates the first The weight matrix of the fully connected layer corresponding to each mode. Indicates the first The bias vector of the fully connected layer corresponding to each mode. ; Time alignment is performed on the mapped modal feature vectors. For follow-up data with irregular time intervals, the number of days between adjacent time points is calculated based on the actual collection date of each time point, and a time mask vector is generated. Each element in the time mask vector indicates the validity of the data at the corresponding time point and the length of the time interval. The formula for calculating the number of days in a time interval is: ; in, Indicates the first The number of days in each time interval. Indicates the first The actual acquisition date at the nth time point, the nth time mask vector element The calculation formula is: ; in, This indicates the maximum allowed time interval in days, with a value of 180 days. Multi-head self-attention mechanism is used to model the time-aligned multimodal feature sequence, and the attention weight distribution of each modality feature at different time points is calculated so that the model automatically focuses on the modality and time point that contributes the most to the prediction of glaucoma risk. The formula for calculating the weight of a single attention head in multi-head self-attention is: ; in, Represents the query matrix. Represents the key matrix. Represents a value matrix, Indicates the dimension of the key vector; Attention-weighted multimodal features are concatenated and compressed into a fixed-dimensional vector through multi-layer nonlinear transformation. This fixed-dimensional vector is the joint latent space representation vector. The formula for calculating multilayer nonlinear transformation is:

[0010] in, Denotes the vector representing the joint latent space. This represents a vector concatenation operation. and They represent the first The weight matrix and bias vector of the layer nonlinear transformation. Indicates the first The activation function of the layer, Indicates the number of layers in the nonlinear transformation.

[0011] Preferably, the step of inputting the joint latent space representation vector into the trained early glaucoma risk prediction model and outputting the early glaucoma risk assessment result for the target patient specifically includes: The joint latent space representation vector is input into a risk prediction network constructed based on a spatiotemporal attention mechanism. The risk prediction network consists of two components: a spatiotemporal encoder and a conditional diffusion decoder. A spatiotemporal encoder is used to perform temporal modeling of the historical joint latent space representation sequence. The spatiotemporal encoder uses a recurrent attention mechanism in the time dimension to capture the evolution trend of features at each time point, and a cross-attention mechanism in the modality dimension to capture the interaction relationship between features of different modalities. The historical sequence encoding result output by the spatiotemporal encoder is used as a conditional input to the conditional diffusion decoder. The conditional diffusion decoder generates the target latent vector for the future preset time point through a step-by-step denoising process. The preset time points are two time nodes: the next six months and the next twelve months. The target latent vector is parsed into the predicted future change in retinal nerve fiber layer thickness and the glaucoma progression risk score using a multi-task decoder. The multi-task decoder contains two independent output heads: one output head regresses and predicts the change in retinal nerve fiber layer thickness, and the other output head classifies and outputs the glaucoma progression risk score. Based on the predicted future changes in retinal nerve fiber layer thickness and glaucoma progression risk score, the target patient is determined to be low-risk, medium-risk, or high-risk according to the preset risk level classification rules, and the corresponding risk probability value is output. The calculation formula for the conditional diffusion decoder in the denoising process at step t is as follows: ; in, Indicates the first Denoising latent vector of step, Indicates the first The noisy latent vector of the step, Indicates the first The noise scheduling coefficient of the step, Represents the cumulative noise scheduling coefficient. This represents a conditional denoising neural network. This represents the encoded result of the historical sequence output by the spatiotemporal encoder. Indicates the first The noise standard deviation of the step, This represents random noise sampled from a standard normal distribution; The formula for calculating the change in retinal nerve fiber layer thickness predicted by the regression output head in the multi-task decoder is as follows: ; in, This represents the predicted change in the thickness of the retinal nerve fiber layer. This represents the target latent vector generated by the conditional diffusion decoder. This represents the weight vector of the regression output head. This represents the bias term of the regression output header; The formula for determining the risk level classification rule is as follows: : in, Indicates the risk score for glaucoma progression. and These represent the probability thresholds for low risk and medium risk, and medium risk and high risk, respectively. and These represent the threshold values ​​for changes in retinal nerve fiber layer thickness at low and medium risk, and medium and high risk, respectively.

[0012] Preferably, the method further includes a training step for an early glaucoma risk prediction model, specifically including: A large amount of multimodal ophthalmological examination data of diagnosed glaucoma patients and healthy controls were collected as training samples. The glaucoma patients in the training samples included patients with primary open-angle glaucoma and patients with primary angle-closure glaucoma, and the healthy controls included individuals with normal intraocular pressure and no visual field defects. Perform the same standardized preprocessing and feature extraction operations as described above on the training samples to obtain the joint latent space representation vector of each training sample; The joint latent space representation vector of the training samples and the corresponding glaucoma diagnostic labels are input into the initial risk prediction model for training. The diagnostic labels include two binary labels: whether glaucoma is diagnosed and whether there is visual field progression. During training, a contrastive learning strategy is used to construct positive and negative sample pairs, enabling the model to learn to distinguish the feature differences between glaucoma progression patients and non-progression patients in the joint latent space. Positive sample pairs are feature vectors of the same patient at different time points, while negative sample pairs are feature vectors of different patients. The formula for calculating the loss function in contrastive learning is: ; in, Indicates the contrast learning loss. Indicates the total number of samples in the batch. Indicates the first Anchor feature vector of each sample, Indicates the relationship with the first Positive sample feature vectors paired with each sample Indicates the first negative sample feature vectors This represents the function for calculating cosine similarity. The temperature coefficient represents the temperature coefficient used in comparative learning. The model performance was evaluated using the five-fold cross-validation method, with the area under the receiver operating characteristic (ROC) curve as the evaluation metric. The hyperparameters of the number of attention heads and the number of diffusion steps were adjusted until the ROC curve reached a preset threshold. The formula for calculating the area under the receiver operating characteristic curve is: ; in, The area under the receiver operating characteristic curve (ROC) is represented by the curve. Indicates the true rate, This indicates the false positive rate.

[0013] Preferably, the step of using a spatiotemporal encoder to perform temporal modeling of the historical joint latent space representation sequence specifically includes: Arrange the historical joint latent space representation sequence in chronological order to form a temporal feature sequence; In the time dimension, attention weights are calculated for the feature vector of each time point in the time feature sequence and for the feature vectors of all other time points. The weight calculation is based on the cosine similarity between feature vectors, enabling the model to identify key turning points in feature evolution. The formula for calculating the attention weight in the time dimension is: ; in, Indicates the first The time point for the first Attention weights at each point in time, time dimension and They represent the first The and the first Feature vectors at each time point The dimension of the feature vector. Indicates the total length of the time series; The formula for calculating attention output in the time dimension is: ; in, Indicates the first The output after time-dimensional attention weighting at each time point The value transformation matrix representing attention over time; In the modal dimension, cross-modal attention weights are calculated for the multimodal feature vectors at each time point, so that structural image features can focus on the corresponding regions in functional visual field features, and intraocular pressure numerical features can establish a correlation with optic disc region features in fundus image features. The formula for calculating the attention weights in the modality dimension is: ; in, Indicates the first The modality pair of the first Attention weights for each modality and They represent the first The and the first Feature vectors of each modality and These represent the query and key transformation matrices for modal dimension attention, respectively. Represents the dimension of the key vector. Indicates the total number of modes; The temporal and modal attention outputs are subjected to residual connections and layer normalization to obtain the spatiotemporally encoded sequence representation. The formula for calculating layer normalization is: ; in, The output after normalization of the representation layer, Represents the input vector. This represents the mean of the input vector. This represents the variance of the input vector. This represents a small constant to prevent division by zero, and its value is... , and This represents learnable scaling and translation parameters; The formula for calculating residual connectivity is: ; in, Indicates the first The final sequence representation after spatiotemporal encoding at each time point The value transformation matrix represents the attention dimension of the modality.

[0014] Preferably, after outputting the early glaucoma risk assessment results for the target patient, the method further includes: Based on the risk level determination in the risk assessment results, a corresponding clinical follow-up recommendation plan is generated. For patients identified as high-risk, recommendations were made to shorten the follow-up interval to three months, and to add retinal nerve fiber layer thickness monitoring and optic disc photography. For patients identified as medium risk, recommendations are made to maintain a regular follow-up frequency, with a follow-up interval of six months, and to focus on monitoring intraocular pressure fluctuations and trends in visual field defects. For patients identified as low-risk, recommendations were made to extend the follow-up interval to twelve months and to perform routine fundus examinations. The clinical follow-up recommendations are output in a structured text format, which includes a risk level field, a follow-up interval field, and a recommended examination item field. The quantitative assessment formula for intraocular pressure fluctuation is as follows: ; in, Indicates intraocular pressure fluctuation score. This indicates the highest intraocular pressure value during the follow-up period. This indicates the lowest intraocular pressure value during the follow-up period. This represents the average intraocular pressure during the follow-up period; The quantitative assessment formula for the changing trend of visual field defects is as follows: ; in, This indicates the percentage change trend of visual field defects. This represents the average visual field defect value at the most recent follow-up visit. This represents the average visual field defect value during baseline follow-up.

[0015] In a second aspect of the invention, an artificial intelligence-driven early risk prediction and assessment system for glaucoma is also provided, comprising: Acquisition Module: Acquires multimodal ophthalmic examination data of the target patient, including structural optical coherence tomography image sequences, functional visual field test data sequences, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data; Preprocessing module: Performs standardized preprocessing on multimodal ophthalmological examination data, converts data from different sources and in different formats into a standardized data format that is compatible with the input of artificial intelligence models, and performs noise reduction, missing value imputation and outlier removal operations on each modality of data respectively; Feature extraction module: Performs feature extraction on the preprocessed multimodal ophthalmological examination data to obtain structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features; Fusion module: The extracted modal features are fused across modalities to generate a joint latent space representation vector, which encodes the comprehensive representation information of the target patient's ocular structure and function; Output module: Input the joint latent space representation vector into the trained early glaucoma risk prediction model, and output the early glaucoma risk assessment results for the target patient. The risk assessment results include the probability value of glaucoma progression and the risk level determination.

[0016] Compared with existing technologies, this invention provides an artificial intelligence-driven method for early risk prediction and assessment of glaucoma, which has the following beneficial effects: This invention overcomes the limitations of traditional single-modality methods by acquiring and fusing multimodal ophthalmic examination information, including structural optical coherence tomography (OCT) images, functional visual field testing data, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data. This enables a comprehensive representation of the patient's ocular structure and function, effectively improving the comprehensiveness and accuracy of early risk prediction. By introducing cross-modal fusion and multi-head self-attention mechanisms, it jointly models across both time and modality dimensions, allowing the model to adaptively identify the differences in the contribution of different time points and examination modalities to risk prediction, fully exploring the deep semantic relationships between multimodal data. Combined with a conditional diffusion decoder for prospective prediction, it can output changes in retinal nerve fiber layer thickness and glaucoma progression risk scores for the next six and twelve months, providing clinicians with forward-looking decision support information. Simultaneously, by employing a contrastive learning strategy to enhance the model's ability to distinguish between patients with and without progression, it effectively improves the model's generalization performance under imbalanced sample conditions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow for S101-S105 in this invention; Figure 2 This is a schematic diagram of the method flow for S201-S205 in this invention; Figure 3 This is a schematic diagram of the method flow for S301-S305 in this invention; Figure 4 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0018] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. Example

[0019] Please refer to Figure 1 As shown, in a first aspect of the present invention, an artificial intelligence-driven method for early risk prediction and assessment of glaucoma is also provided, comprising: S101. Obtain multimodal ophthalmic examination data of the target patient. The multimodal ophthalmic examination data includes structural optical coherence tomography image sequence, functional visual field test data sequence, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data. S102. Perform standardized preprocessing on multimodal ophthalmic examination data, convert data from different sources and in different formats into a standardized data format that is compatible with the input of artificial intelligence models, and perform noise reduction, missing value imputation and outlier removal operations on each modality of data respectively. S103. Perform feature extraction on the preprocessed multimodal ophthalmological examination data to obtain structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features. S104. The extracted modal features are fused across modalities to generate a joint latent space representation vector, which encodes the comprehensive representation information of the target patient's ocular structure and function. S105. Input the joint latent space representation vector into the trained early glaucoma risk prediction model, and output the early glaucoma risk assessment results for the target patient. The risk assessment results include the probability value of glaucoma progression and the risk level determination.

[0020] As will be understood by those skilled in the art, this invention overcomes the limitations of traditional single-modal methods by acquiring and fusing multimodal ophthalmic examination information, such as structural optical coherence tomography (OCT) images, functional visual field testing data, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data. This achieves a comprehensive representation of the patient's ocular structure and function, effectively improving the comprehensiveness and accuracy of early risk prediction. The invention introduces cross-modal fusion and multi-head self-attention mechanisms, jointly modeling in both time and modality dimensions. This enables the model to adaptively identify the differences in the contribution of different time points and examination modalities to risk prediction, fully exploring the deep semantic relationships between multimodal data. Combined with a conditional diffusion decoder for prospective prediction, it can output retinal nerve fiber layer thickness changes and glaucoma progression risk scores for the next six and twelve months, providing clinicians with prospective decision support information. Simultaneously, the contrastive learning strategy enhances the model's ability to distinguish features between those with and without progression, effectively improving the model's generalization performance under imbalanced sample conditions.

[0021] Please refer to Figure 2 As shown, multimodal ophthalmological examination data of the target patient was obtained, specifically including: S201. Extract the structural optical coherence tomography (SOCT) image sequence of the target patient from the hospital information system. The SOCT image sequence includes retinal nerve fiber layer thickness map, optic disc scan image and macular scan image, and each image sequence includes follow-up images at no less than three time points with different follow-up intervals between time points. S202. Export the functional visual field test data sequence of the target patient from the visual field testing device. The functional visual field test data sequence contains the 24-2 procedure test results of the Humphrey perimeter. The test results are stored in the form of a spatial distribution matrix of sensitivity thresholds. Each element in the matrix corresponds to the light sensitivity value of a test point in the visual field test grid. S203. Collect intraocular pressure measurement data of the target patient. The intraocular pressure measurement data includes intraocular pressure values ​​measured by a non-contact tonometer and intraocular pressure values ​​measured by a Goldmann applanation tonometer, as well as central corneal thickness data obtained by an ultrasonic pachymeter. S204. Obtain color photos of the fundus of the target patient, covering the optic disc area and the macula area, with a resolution of not less than 2000×2000 pixels; S205. Obtain the anterior chamber angle examination data of the target patient. The anterior chamber angle examination data includes the grading data of the degree of opening and closing of the anterior chamber angle by gonioscopy. The grading data is recorded as a text description and the corresponding angle value according to the Shaffer grading standard. ; in, This represents the numerical encoding result of the Shaffer classification, with a value range of [value range missing]. , This indicates the angle of the anterior chamber as recorded by gonioscopy.

[0022] Please refer to Figure 3 As shown, standardized preprocessing of multimodal ophthalmic examination data includes: S301. The structured optical coherence tomography (OCT) image sequence is format-converted from DICOM to NIfTI format. Speckle noise removal is then applied to the converted images using a wavelet transform-based denoising method. After multi-scale decomposition, soft thresholding is performed on the high-frequency subband coefficients. The formula for soft thresholding is: ; in, Indicates the first Layer decomposition Wavelet coefficients at each position after soft thresholding Represents the original wavelet coefficients. The threshold parameter is determined by the following formula: ; in, The noise standard deviation is estimated using the median absolute deviation of the wavelet coefficients at the finest scale. and These represent the number of rows and columns of the image, respectively. S302. Missing values ​​are detected and filled in the spatial distribution matrix of the sensitivity threshold in the functional visual field test data sequence. For local missing areas caused by the patient's fixation loss, an inverse distance weighted interpolation method based on eight surrounding test points is used for filling. The interpolation calculation formula is as follows: ; in, This indicates the sensitivity threshold after filling in the blanks. Indicates the first The original sensitivity threshold of each adjacent test point Indicates the first Euclidean distance from each adjacent test point to the missing location This represents the distance weighting index, with a value of 2. S303. Perform outlier detection on the intraocular pressure measurement data, and calculate the upper and lower bounds of the intraocular pressure values ​​using the interquartile range method. The formula for calculating the upper and lower bounds is as follows: ; ; in, This represents the first quartile of the intraocular pressure value sequence. The third quartile represents the intraocular pressure value sequence, and IQR represents the interquartile range. and These represent the lower and upper limits of intraocular pressure, respectively. and These represent the functions for calculating the 25th and 75th percentiles, respectively, and will exceed [ Intraocular pressure values ​​within a certain range are marked as abnormal, and these abnormal values ​​are corrected using central corneal thickness data. The correction formula is as follows: ; in, This indicates the corrected intraocular pressure value. This represents the original measured intraocular pressure value. This represents the corneal thickness correction factor, with a value of 0.013. This indicates the standard central corneal thickness, with a value of 520 micrometers. This indicates the actual measured central corneal thickness; S304. Perform brightness normalization processing on the fundus color photograph. Based on the brightness threshold, divide the fundus brightness grayscale image into black and white images to separate the optic disc and vascular regions. Then, linearly adjust the brightness of the entire photograph according to the brightness threshold parameters of the shooting device. The brightness adjustment formula is as follows: ; in, Indicates the normalized position Pixel brightness value at that location, Indicates the original image location Pixel brightness value at that location, Indicates the brightness scaling factor. This indicates the brightness offset. and The value is determined by fitting the brightness threshold parameter of the shooting device using the least squares method; S305. Perform structured coding on the anterior chamber angle examination data, convert the textual description of Shaffer classification into numerical codes from one to four, retain the angle values ​​as floating-point data, and form a structured anterior chamber angle feature table.

[0023] Feature extraction was performed on the preprocessed multimodal ophthalmological examination data, specifically including: A convolutional neural network was used to extract features from the retinal nerve fiber layer thickness map in a structural optical coherence tomography (OCT) image sequence. The set of outputs from the convolutional and pooling layers of three lines was used to extract the color features, edge features, and texture features of the retinal nerve fiber layer, as well as the correlation features between these features. The color features were composed of the output of a pooling layer of the first line and a convolutional layer of the second line. The texture features were composed of the output of a zero convolutional layer of the third line, a convolutional layer of the second line, and a convolutional layer of the third line. The three sets of features were concatenated and mapped through an activation function to obtain the structural image feature vector. Statistical feature calculations were performed on the spatial distribution matrix of sensitivity thresholds in the functional visual field detection data sequence to extract the location features, depth features, and extent features of visual field defects. The slope and fluctuation features of sensitivity values ​​at each test point over time were also calculated. The functional visual field feature vector was obtained by concatenating the above five types of features. The formula for calculating the slope characteristic of the sensitivity value of each test point changing with time is as follows: ; in, Indicates the first The slope of sensitivity change at each test point This indicates the total number of follow-up time points. Indicates the first The test point at the ... Sensitivity threshold at each time point This represents the mean at a given time point. Indicates the first The mean sensitivity of each test point at all time points; The formula for calculating the fluctuation characteristics of sensitivity values ​​at each test point is as follows: ; in, Indicates the first Sensitivity fluctuation values ​​at each test point; Time-domain statistical analysis was performed on the intraocular pressure measurement data to extract five numerical features: daily average, daily peak, daily trough, intraocular pressure fluctuation amplitude, and diurnal intraocular pressure difference. The five numerical features were normalized and then spliced ​​together to obtain the intraocular pressure numerical features. The formula for calculating the amplitude of intraocular pressure fluctuation is as follows: ; in, This indicates the amplitude of intraocular pressure fluctuations. Indicates the first Corrected intraocular pressure values ​​at each time point This indicates the total number of time points used for intraocular pressure measurement; The formula for calculating the diurnal intraocular pressure difference is: ; in, This represents the diurnal intraocular pressure difference. Indicates the number of measurements taken during the day. Indicates the number of measurements taken at night; An image feature extraction network was used to extract features from color fundus photographs. The optic disc cup-to-disc ratio, retinal vessel tortuosity, and retinal hemorrhage density were extracted through multi-layer convolution and pooling operations. The fundus image features were obtained by concatenating the above three features. The formula for calculating the retinal vessel tortuosity characteristics is as follows: ; in, Indicates the curvature of retinal vessels. This indicates the total number of blood vessel segments detected. Indicates the first The actual path length of the segmental blood vessel Indicates the first The straight-line distance between the two ends of a segment of blood vessel; The formula for calculating the density characteristics of retinal hemorrhages is: ; in, Indicates the density of retinal hemorrhages. This indicates the total number of bleeding points detected. This indicates the area of ​​the effective analysis region in the fundus photograph; Feature encoding was performed on the anterior chamber angle examination data, and three structured features were extracted: the angle width value, the trabecular mesh visibility level, and the iris morphology classification. The three features were then individually encoded and spliced ​​to obtain the anterior chamber angle structural features. The conversion formula for one-hot encoding is as follows: ; in, Indicates the first The value of the one-hot encoded bit, The range of values ​​is , This represents the total number of categories for this classification feature.

[0024] The extracted modal features are fused across modalities to generate a joint latent space representation vector, specifically including: Structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features are mapped to a feature space of a unified dimension through fully connected layers, so that the dimensions of each modality feature vector are consistent. The formula for calculating the mapping of fully connected layers is: ; in, Indicates the first Feature vectors after modal mapping Indicates the first The original feature vectors of each modality, Indicates the first The weight matrix of the fully connected layer corresponding to each mode. Indicates the first The bias vector of the fully connected layer corresponding to each mode. ; Time alignment is performed on the mapped modal feature vectors. For follow-up data with irregular time intervals, the number of days between adjacent time points is calculated based on the actual collection date of each time point, and a time mask vector is generated. Each element in the time mask vector indicates the validity of the data at the corresponding time point and the length of the time interval. The formula for calculating the number of days in a time interval is: ; in, Indicates the first The number of days in each time interval. Indicates the first The actual acquisition date at the nth time point, the nth time mask vector element The calculation formula is: ; in, This indicates the maximum allowed time interval in days, with a value of 180 days. Multi-head self-attention mechanism is used to model the time-aligned multimodal feature sequence, and the attention weight distribution of each modality feature at different time points is calculated so that the model automatically focuses on the modality and time point that contributes the most to the prediction of glaucoma risk. The formula for calculating the weight of a single attention head in multi-head self-attention is: ; in, Represents the query matrix. Represents the key matrix. Represents a value matrix, Indicates the dimension of the key vector; Attention-weighted multimodal features are concatenated and compressed into a fixed-dimensional vector through multi-layer nonlinear transformation. This fixed-dimensional vector is the joint latent space representation vector. The formula for calculating multilayer nonlinear transformation is:

[0025] in, Denotes the vector representing the joint latent space. This represents a vector concatenation operation. and They represent the first The weight matrix and bias vector of the layer nonlinear transformation. Indicates the first The activation function of the layer, Indicates the number of layers in the nonlinear transformation.

[0026] The joint latent space representation vector is input into the trained early glaucoma risk prediction model, which outputs the early glaucoma risk assessment results for the target patient, specifically including: The joint latent space representation vector is input into a risk prediction network constructed based on a spatiotemporal attention mechanism. The risk prediction network consists of two components: a spatiotemporal encoder and a conditional diffusion decoder. A spatiotemporal encoder is used to perform temporal modeling of the historical joint latent space representation sequence. The spatiotemporal encoder uses a recurrent attention mechanism in the time dimension to capture the evolution trend of features at each time point, and a cross-attention mechanism in the modality dimension to capture the interaction relationship between features of different modalities. The historical sequence encoding result output by the spatiotemporal encoder is used as a conditional input to the conditional diffusion decoder. The conditional diffusion decoder generates the target latent vector for the future preset time point through a step-by-step denoising process. The preset time points are two time nodes: the next six months and the next twelve months. The target latent vector is parsed into the predicted future change in retinal nerve fiber layer thickness and the glaucoma progression risk score using a multi-task decoder. The multi-task decoder contains two independent output heads: one output head regresses and predicts the change in retinal nerve fiber layer thickness, and the other output head classifies and outputs the glaucoma progression risk score. Based on the predicted future changes in retinal nerve fiber layer thickness and glaucoma progression risk score, the target patient is determined to be low-risk, medium-risk, or high-risk according to the preset risk level classification rules, and the corresponding risk probability value is output. The calculation formula for the conditional diffusion decoder in the denoising process at step t is as follows: ; in, Indicates the first Denoising latent vector of step, Indicates the first The noisy latent vector of the step, Indicates the first The noise scheduling coefficient of the step, Represents the cumulative noise scheduling coefficient. This represents a conditional denoising neural network. This represents the encoded result of the historical sequence output by the spatiotemporal encoder. Indicates the first The noise standard deviation of the step, This represents random noise sampled from a standard normal distribution; The formula for calculating the change in retinal nerve fiber layer thickness predicted by the regression output head in the multi-task decoder is as follows: ; in, This represents the predicted change in the thickness of the retinal nerve fiber layer. This represents the target latent vector generated by the conditional diffusion decoder. This represents the weight vector of the regression output head. This represents the bias term of the regression output header; The formula for determining the risk level classification rule is as follows: : in, Indicates the risk score for glaucoma progression. and These represent the probability thresholds for low risk and medium risk, and medium risk and high risk, respectively. and These represent the threshold values ​​for changes in retinal nerve fiber layer thickness at low and medium risk, and medium and high risk, respectively.

[0027] The method also includes a training step for an early glaucoma risk prediction model, specifically including: We collected a large amount of multimodal ophthalmological examination data from diagnosed glaucoma patients and healthy controls as training samples. The glaucoma patients in the training samples included patients with primary open-angle glaucoma and patients with primary angle-closure glaucoma, while the healthy controls included individuals with normal intraocular pressure and no visual field defects. Perform the same standardized preprocessing and feature extraction operations as described above on the training samples to obtain the joint latent space representation vector of each training sample; The joint latent space representation vector of the training samples and the corresponding glaucoma diagnostic labels are input into the initial risk prediction model for training. The diagnostic labels include two binary labels: whether glaucoma is diagnosed and whether there is visual field progression. During training, a contrastive learning strategy is used to construct positive and negative sample pairs, enabling the model to learn to distinguish the feature differences between glaucoma progression patients and non-progression patients in the joint latent space. Positive sample pairs are feature vectors of the same patient at different time points, while negative sample pairs are feature vectors of different patients. The formula for calculating the loss function in contrastive learning is: ; in, Indicates the contrast learning loss. Indicates the total number of samples in the batch. Indicates the first Anchor feature vector of each sample, Indicates the relationship with the first Positive sample feature vectors paired with each sample Indicates the first negative sample feature vectors This represents the function for calculating cosine similarity. The temperature coefficient represents the temperature coefficient used in comparative learning. The model performance was evaluated using the five-fold cross-validation method, with the area under the receiver operating characteristic (ROC) curve as the evaluation metric. The hyperparameters of the number of attention heads and the number of diffusion steps were adjusted until the ROC curve reached a preset threshold. The formula for calculating the area under the receiver operating characteristic curve is: ; in, The area under the receiver operating characteristic curve (ROC) is represented by the curve. Indicates the true rate, This indicates the false positive rate.

[0028] Temporal modeling of historical joint latent space representation sequences using a spatiotemporal encoder specifically includes: Arrange the historical joint latent space representation sequence in chronological order to form a temporal feature sequence; In the time dimension, attention weights are calculated for the feature vector of each time point in the time feature sequence and for the feature vectors of all other time points. The weight calculation is based on the cosine similarity between feature vectors, enabling the model to identify key turning points in feature evolution. The formula for calculating the attention weight in the time dimension is: ; in, Indicates the first The time point for the first Attention weights at each point in time, time dimension and They represent the first The and the first Feature vectors at each time point The dimension of the feature vector. Indicates the total length of the time series; The formula for calculating attention output in the time dimension is: ; in, Indicates the first The output after time-dimensional attention weighting at each time point The value transformation matrix representing attention over time; In the modal dimension, cross-modal attention weights are calculated for the multimodal feature vectors at each time point, so that structural image features can focus on the corresponding regions in functional visual field features, and intraocular pressure numerical features can establish a correlation with optic disc region features in fundus image features. The formula for calculating the attention weights in the modality dimension is: ; in, Indicates the first The modality pair of the first Attention weights for each modality and They represent the first The and the first Feature vectors of each modality and These represent the query and key transformation matrices for modal dimension attention, respectively. Represents the dimension of the key vector. Indicates the total number of modes; The temporal and modal attention outputs are subjected to residual connections and layer normalization to obtain the spatiotemporally encoded sequence representation. The formula for calculating layer normalization is: ; in, The output after normalization of the representation layer, Represents the input vector. This represents the mean of the input vector. This represents the variance of the input vector. This represents a small constant to prevent division by zero, and its value is... , and This represents learnable scaling and translation parameters; The formula for calculating residual connectivity is: ; in, Indicates the first The final sequence representation after spatiotemporal encoding at each time point The value transformation matrix represents the attention dimension of the modality.

[0029] After outputting the early glaucoma risk assessment results for the target patients, it also includes: Based on the risk level determination in the risk assessment results, a corresponding clinical follow-up recommendation plan is generated. For patients identified as high-risk, recommendations were made to shorten the follow-up interval to three months, and to add retinal nerve fiber layer thickness monitoring and optic disc photography. For patients identified as medium risk, recommendations are made to maintain a regular follow-up frequency, with a follow-up interval of six months, and to focus on monitoring intraocular pressure fluctuations and trends in visual field defects. For patients identified as low-risk, recommendations were made to extend the follow-up interval to twelve months and to perform routine fundus examinations. The clinical follow-up recommendations will be output in a structured text format, which includes a risk level field, a follow-up interval field, and a recommended examination item field. The quantitative assessment formula for intraocular pressure fluctuation is as follows: ; in, Indicates intraocular pressure fluctuation score. This indicates the highest intraocular pressure value during the follow-up period. This indicates the lowest intraocular pressure value during the follow-up period. This represents the average intraocular pressure during the follow-up period; The quantitative assessment formula for the changing trend of visual field defects is as follows: ; in, This indicates the percentage change trend of visual field defects. This represents the average visual field defect value at the most recent follow-up visit. This represents the average visual field defect value during baseline follow-up.

[0030] Please refer to Figure 4 As shown, in a second aspect of the present invention, an artificial intelligence-driven early risk prediction and assessment system for glaucoma is also provided, comprising: Acquisition Module: Acquires multimodal ophthalmic examination data of the target patient, including structural optical coherence tomography image sequences, functional visual field test data sequences, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data; Preprocessing module: Performs standardized preprocessing on multimodal ophthalmological examination data, converts data from different sources and in different formats into a standardized data format that is compatible with the input of artificial intelligence models, and performs noise reduction, missing value imputation and outlier removal operations on each modality of data respectively; Feature extraction module: Performs feature extraction on the preprocessed multimodal ophthalmological examination data to obtain structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features; Fusion module: The extracted modal features are fused across modalities to generate a joint latent space representation vector, which encodes the comprehensive representation information of the target patient's ocular structure and function; Output module: Input the joint latent space representation vector into the trained early glaucoma risk prediction model, and output the early glaucoma risk assessment results for the target patient. The risk assessment results include the probability value of glaucoma progression and the risk level determination.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An AI-driven method for early risk prediction and assessment of glaucoma, characterized in that, include: Acquire multimodal ophthalmic examination data of the target patients, including structural optical coherence tomography image sequences, functional visual field test data sequences, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data; Standardized preprocessing was performed on multimodal ophthalmological examination data to convert data from different sources and formats into a standardized data format that is compatible with the input of artificial intelligence models. Denoising, missing value imputation, and outlier removal were performed on each modality of data. Feature extraction was performed on the preprocessed multimodal ophthalmological examination data to obtain structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features. The extracted modal features are fused across modalities to generate a joint latent space representation vector, which encodes the comprehensive representation information of the target patient's ocular structure and function. The joint latent space representation vector is input into the trained early glaucoma risk prediction model, and the output is the early glaucoma risk assessment result of the target patient. The risk assessment result includes the probability value of glaucoma progression and the risk level determination.

2. The AI-driven early risk prediction and assessment method for glaucoma according to claim 1, characterized in that, The acquisition of multimodal ophthalmological examination data of the target patient specifically includes: Structural optical coherence tomography (SOCT) image sequences of target patients were extracted from the hospital information system. The SOCT image sequences included retinal nerve fiber layer thickness maps, optic disc scan images, and macular scan images. Each image sequence contained follow-up images at no less than three time points, with varying follow-up intervals between the time points. The functional visual field test data sequence of the target patient is exported from the visual field testing device. The functional visual field test data sequence contains the test results of the Humphrey perimeter 24-2 procedure. The test results are stored in the form of a spatial distribution matrix of sensitivity thresholds. Each element in the matrix corresponds to the light sensitivity value of a test point in the visual field test grid. Intraocular pressure (IOP) data were collected from the target patients. The IOP data included IOP values ​​measured by a non-contact tonometer and an IOP value measured by a Goldmann applanation tonometer, as well as central corneal thickness data obtained by an ultrasonic pachymeter. Obtain color fundus images of the target patient, covering the optic disc region and macular region, with a resolution of no less than 2000×2000 pixels; Acquire anterior chamber angle examination data of the target patient. The anterior chamber angle examination data includes the grading data of the degree of angle opening and closing of the anterior chamber angle obtained by gonioscopy. The grading data is recorded as text description and corresponding angle value according to the Shaffer grading standard. ； in, This represents the numerical encoding result of the Shaffer classification, with a value range of [value range missing]. , This indicates the angle of the anterior chamber as recorded by gonioscopy.

3. The AI-driven early risk prediction and assessment method for glaucoma according to claim 2, characterized in that, The standardization preprocessing of multimodal ophthalmic examination data specifically includes: The structural optical coherence tomography (OCT) image sequence was format-converted from DICOM to NIfTI format. Speckle noise removal was then applied to the converted images using a wavelet transform-based denoising method. After multi-scale decomposition, soft thresholding was performed on the high-frequency subband coefficients. The formula for soft thresholding is as follows: ； in, Indicates the first Layer decomposition Wavelet coefficients at each position after soft thresholding Represents the original wavelet coefficients. The threshold parameter is determined by the following formula: ； in, The noise standard deviation is estimated using the median absolute deviation of the wavelet coefficients at the finest scale. and These represent the number of rows and columns of the image, respectively. Missing values ​​were detected and imputed in the spatial distribution matrix of sensitivity thresholds in the functional visual field test data sequence. For local missing regions caused by fixation loss in patients, an inverse distance-weighted interpolation method based on eight surrounding test points was used for imputation. The interpolation calculation formula is as follows: ； in, This indicates the sensitivity threshold after filling in the blanks. Indicates the first The original sensitivity threshold of each adjacent test point Indicates the first Euclidean distance from each adjacent test point to the missing location This represents the distance weighting index, with a value of 2. Outlier detection was performed on the intraocular pressure measurement data, and the upper and lower bounds of the intraocular pressure values ​​were calculated using the interquartile range method. The formulas for calculating the upper and lower bounds are as follows: ； ； in, This represents the first quartile of the intraocular pressure value sequence. The third quartile represents the intraocular pressure value sequence, and IQR represents the interquartile range. and These represent the lower and upper limits of intraocular pressure, respectively. and These represent the functions for calculating the 25th and 75th percentiles, respectively, and will exceed [ Intraocular pressure values ​​within a certain range are marked as abnormal, and these abnormal values ​​are corrected using central corneal thickness data. The correction formula is as follows: ； in, This indicates the corrected intraocular pressure value. This represents the original measured intraocular pressure value. This represents the corneal thickness correction factor, with a value of 0.

013. This indicates the standard central corneal thickness, with a value of 520 micrometers. This indicates the actual measured central corneal thickness; Brightness normalization was performed on the color fundus photographs. Based on a brightness threshold, the fundus brightness grayscale image was divided into black and white images to separate the optic disc and vascular regions. The brightness of the entire image was then linearly adjusted according to the brightness threshold parameters of the capturing device. The brightness adjustment formula is as follows: ； in, Indicates the normalized position Pixel brightness value at that location, Indicates the original image location Pixel brightness value at that location, Indicates the brightness scaling factor. This indicates the brightness offset. and The value is determined by fitting the brightness threshold parameter of the shooting device using the least squares method; The anterior chamber angle examination data were structured and encoded, converting the textual descriptions of the Shaffer classification into numerical codes from one to four, while retaining the angle values ​​as floating-point data, thus forming a structured anterior chamber angle feature table.

4. The AI-driven early risk prediction and assessment method for glaucoma according to claim 3, characterized in that, The feature extraction of the preprocessed multimodal ophthalmic examination data specifically includes: A convolutional neural network was used to extract features from the retinal nerve fiber layer thickness map in a structural optical coherence tomography (OCT) image sequence. The set of outputs from the convolutional and pooling layers of three lines was used to extract the color features, edge features, and texture features of the retinal nerve fiber layer, as well as the correlation features between these features. The color features were composed of the output of a pooling layer of the first line and a convolutional layer of the second line. The texture features were composed of the output of a zero convolutional layer of the third line, a convolutional layer of the second line, and a convolutional layer of the third line. The three sets of features were concatenated and mapped through an activation function to obtain the structural image feature vector. Statistical feature calculations were performed on the spatial distribution matrix of sensitivity thresholds in the functional visual field detection data sequence to extract the location features, depth features, and extent features of visual field defects. The slope and fluctuation features of sensitivity values ​​at each test point over time were also calculated. The functional visual field feature vector was obtained by concatenating the above five types of features. The formula for calculating the slope characteristic of the sensitivity value of each test point changing over time is as follows: ； in, Indicates the first The slope of sensitivity change at each test point This indicates the total number of follow-up time points. Indicates the first The test point at the ... Sensitivity threshold at each time point This represents the mean at a given time point. Indicates the first The mean sensitivity of each test point at all time points; The formula for calculating the fluctuation characteristics of sensitivity values ​​at each test point is as follows: ； in, Indicates the first Sensitivity fluctuation values ​​at each test point; Time-domain statistical analysis was performed on the intraocular pressure measurement data to extract five numerical features: daily average, daily peak, daily trough, intraocular pressure fluctuation amplitude, and diurnal intraocular pressure difference. The five numerical features were normalized and then spliced ​​together to obtain the intraocular pressure numerical features. The formula for calculating the amplitude of intraocular pressure fluctuation is as follows: ； in, This indicates the amplitude of intraocular pressure fluctuations. Indicates the first Corrected intraocular pressure values ​​at each time point This indicates the total number of time points used for intraocular pressure measurement; The formula for calculating the diurnal intraocular pressure difference is: ； in, This represents the diurnal intraocular pressure difference. Indicates the number of measurements taken during the day. Indicates the number of measurements taken at night; An image feature extraction network was used to extract features from color fundus photographs. The optic disc cup-to-disc ratio, retinal vessel tortuosity, and retinal hemorrhage density were extracted through multi-layer convolution and pooling operations. The fundus image features were obtained by concatenating the above three features. The formula for calculating the retinal vessel tortuosity characteristics is as follows: ； in, Indicates the curvature of retinal vessels. This indicates the total number of blood vessel segments detected. Indicates the first The actual path length of the segmental blood vessel Indicates the first The straight-line distance between the two ends of a segment of blood vessel; The formula for calculating the density characteristics of retinal hemorrhages is: ； in, Indicates the density of retinal hemorrhages. This indicates the total number of bleeding points detected. This indicates the area of ​​the effective analysis region in the fundus photograph; Feature encoding was performed on the anterior chamber angle examination data, and three structured features were extracted: the angle width value, the trabecular mesh visibility level, and the iris morphology classification. The three features were then individually encoded and spliced ​​to obtain the anterior chamber angle structural features. The conversion formula for one-hot encoding is as follows: ； in, Indicates the first The value of the one-hot encoded bit, The range of values ​​is , This represents the total number of categories for this classification feature.

5. The AI-driven early risk prediction and assessment method for glaucoma according to claim 4, characterized in that, The step of fusing the extracted modal features across modalities to generate a joint latent space representation vector specifically includes: Structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features are mapped to a feature space of a unified dimension through fully connected layers, so that the dimensions of each modality feature vector are consistent. The formula for calculating the mapping of fully connected layers is: ； in, Indicates the first Feature vectors after modal mapping Indicates the first The original feature vectors of each modality, Indicates the first The weight matrix of the fully connected layer corresponding to each mode. Indicates the first The bias vector of the fully connected layer corresponding to each mode. ; Time alignment is performed on the mapped modal feature vectors. For follow-up data with irregular time intervals, the number of days between adjacent time points is calculated based on the actual collection date of each time point, and a time mask vector is generated. Each element in the time mask vector indicates the validity of the data at the corresponding time point and the length of the time interval. The formula for calculating the number of days in a time interval is: ； in, Indicates the first The number of days in each time interval. Indicates the first The actual acquisition date at the nth time point, the nth time mask vector element The calculation formula is: ； in, This indicates the maximum allowed time interval in days, with a value of 180 days. Multi-head self-attention mechanism is used to model the time-aligned multimodal feature sequence, and the attention weight distribution of each modality feature at different time points is calculated so that the model automatically focuses on the modality and time point that contributes the most to the prediction of glaucoma risk. The formula for calculating the weight of a single attention head in multi-head self-attention is: ； in, Represents the query matrix. Represents the key matrix, Represents a value matrix, Indicates the dimension of the key vector; Attention-weighted multimodal features are concatenated and compressed into a fixed-dimensional vector through multi-layer nonlinear transformation. This fixed-dimensional vector is the joint latent space representation vector. The formula for calculating multilayer nonlinear transformation is: ； in, Denotes the vector representing the joint latent space. This represents a vector concatenation operation. and They represent the first The weight matrix and bias vector of the layer nonlinear transformation. Indicates the first The activation function of the layer, Indicates the number of layers in the nonlinear transformation.

6. The artificial intelligence-driven early risk prediction and assessment method for glaucoma according to claim 5, characterized in that, The method of inputting the joint latent space representation vector into the trained early glaucoma risk prediction model and outputting the early glaucoma risk assessment results for the target patient specifically includes: The joint latent space representation vector is input into a risk prediction network constructed based on a spatiotemporal attention mechanism. The risk prediction network consists of two components: a spatiotemporal encoder and a conditional diffusion decoder. A spatiotemporal encoder is used to perform temporal modeling of the historical joint latent space representation sequence. The spatiotemporal encoder uses a recurrent attention mechanism in the time dimension to capture the evolution trend of features at each time point, and a cross-attention mechanism in the modality dimension to capture the interaction relationship between features of different modalities. The historical sequence encoding result output by the spatiotemporal encoder is used as a conditional input to the conditional diffusion decoder. The conditional diffusion decoder generates the target latent vector for the future preset time point through a step-by-step denoising process. The preset time points are two time nodes: the next six months and the next twelve months. The target latent vector is parsed into the predicted future change in retinal nerve fiber layer thickness and the glaucoma progression risk score using a multi-task decoder. The multi-task decoder contains two independent output heads: one output head regresses and predicts the change in retinal nerve fiber layer thickness, and the other output head classifies and outputs the glaucoma progression risk score. Based on the predicted future changes in retinal nerve fiber layer thickness and glaucoma progression risk score, the target patient is determined to be low-risk, medium-risk, or high-risk according to the preset risk level classification rules, and the corresponding risk probability value is output. The calculation formula for the conditional diffusion decoder in the denoising process at step t is as follows: ； in, Indicates the first Denoising latent vector of step, Indicates the first The noisy latent vector of the step, Indicates the first The noise scheduling coefficient of the step, Represents the cumulative noise scheduling coefficient. This represents a conditional denoising neural network. This represents the encoded result of the historical sequence output by the spatiotemporal encoder. Indicates the first The noise standard deviation of the step, This represents random noise sampled from a standard normal distribution; The formula for calculating the change in retinal nerve fiber layer thickness predicted by the regression output head in the multi-task decoder is as follows: ； in, This represents the predicted change in the thickness of the retinal nerve fiber layer. This represents the target latent vector generated by the conditional diffusion decoder. This represents the weight vector of the regression output head. This represents the bias term of the regression output header; The formula for determining the risk level classification rule is as follows: ： in, Indicates the risk score for glaucoma progression. and These represent the probability thresholds for low risk and medium risk, and medium risk and high risk, respectively. and These represent the threshold values ​​for changes in retinal nerve fiber layer thickness at low and medium risk, and medium and high risk, respectively.

7. The AI-driven early risk prediction and assessment method for glaucoma according to claim 6, characterized in that, The method also includes a training step for an early glaucoma risk prediction model, specifically including: A large amount of multimodal ophthalmological examination data of diagnosed glaucoma patients and healthy controls were collected as training samples. The glaucoma patients in the training samples included patients with primary open-angle glaucoma and patients with primary angle-closure glaucoma, and the healthy controls included individuals with normal intraocular pressure and no visual field defects. Perform the same standardized preprocessing and feature extraction operations as in claim 1 on the training samples to obtain the joint latent space representation vector of each training sample; The joint latent space representation vector of the training samples and the corresponding glaucoma diagnostic labels are input into the initial risk prediction model for training. The diagnostic labels include two binary labels: whether glaucoma is diagnosed and whether there is visual field progression. During training, a contrastive learning strategy is used to construct positive and negative sample pairs, enabling the model to learn to distinguish the feature differences between glaucoma progression patients and non-progression patients in the joint latent space. Positive sample pairs are feature vectors of the same patient at different time points, while negative sample pairs are feature vectors of different patients. The formula for calculating the loss function in contrastive learning is: ； in, Indicates the contrast learning loss. Indicates the total number of samples in the batch. Indicates the first Anchor feature vector of each sample, Indicates the relationship with the first Positive sample feature vectors paired with each sample Indicates the first negative sample feature vectors This represents the function for calculating cosine similarity. The temperature coefficient represents the temperature coefficient used in comparative learning. The model performance was evaluated using the five-fold cross-validation method, with the area under the receiver operating characteristic (ROC) curve as the evaluation metric. The hyperparameters of the number of attention heads and the number of diffusion steps were adjusted until the ROC curve reached a preset threshold. The formula for calculating the area under the receiver operating characteristic curve is: ； in, The area under the receiver operating characteristic curve (ROC) is represented by the curve. Indicates the true rate, This indicates the false positive rate.

8. The artificial intelligence-driven early risk prediction and assessment method for glaucoma according to claim 7, characterized in that, The temporal modeling of historical joint latent space representation sequences using a spatiotemporal encoder specifically includes: Arrange the historical joint latent space representation sequence in chronological order to form a temporal feature sequence; In the time dimension, attention weights are calculated for the feature vector of each time point in the time feature sequence and for the feature vectors of all other time points. The weight calculation is based on the cosine similarity between feature vectors, enabling the model to identify key turning points in feature evolution. The formula for calculating the attention weight in the time dimension is: ； in, Indicates the first The time point for the first Attention weights at each point in time, time dimension and They represent the first The and the first Feature vectors at each time point The dimension of the feature vector. Indicates the total length of the time series; The formula for calculating attention output in the time dimension is: ； in, Indicates the first The output after time-dimensional attention weighting at each time point The value transformation matrix representing attention over time; In the modal dimension, cross-modal attention weights are calculated for the multimodal feature vectors at each time point, so that structural image features can focus on the corresponding regions in functional visual field features, and intraocular pressure numerical features can establish a correlation with optic disc region features in fundus image features. The formula for calculating the attention weights in the modality dimension is: ； in, Indicates the first The modality pair of the first Attention weights for each modality and They represent the first The and the first Feature vectors of each modality and These represent the query and key transformation matrices for modal dimension attention, respectively. Represents the dimension of the key vector. Indicates the total number of modes; The temporal and modal attention outputs are subjected to residual connections and layer normalization to obtain the spatiotemporally encoded sequence representation. The formula for calculating layer normalization is: ； in, The output after normalization of the representation layer, Represents the input vector. This represents the mean of the input vector. This represents the variance of the input vector. This represents a small constant to prevent division by zero, and its value is... , and This represents learnable scaling and translation parameters; The formula for calculating residual connectivity is: ； in, Indicates the first The final sequence representation after spatiotemporal encoding at each time point The value transformation matrix represents the attention dimension of the modality.

9. The artificial intelligence-driven early risk prediction and assessment method for glaucoma according to claim 8, characterized in that, Following the output of the early glaucoma risk assessment results for the target patient, the following is also included: Based on the risk level determination in the risk assessment results, a corresponding clinical follow-up recommendation plan is generated. For patients identified as high-risk, recommendations were made to shorten the follow-up interval to three months, and to add retinal nerve fiber layer thickness monitoring and optic disc photography. For patients identified as medium risk, recommendations are made to maintain a regular follow-up frequency, with a follow-up interval of six months, and to focus on monitoring intraocular pressure fluctuations and trends in visual field defects. For patients identified as low-risk, recommendations were made to extend the follow-up interval to twelve months and to perform routine fundus examinations. The clinical follow-up recommendations are output in a structured text format, which includes a risk level field, a follow-up interval field, and a recommended examination item field. The quantitative assessment formula for intraocular pressure fluctuation is as follows: ； in, Indicates intraocular pressure fluctuation score. This indicates the highest intraocular pressure value during the follow-up period. This indicates the lowest intraocular pressure value during the follow-up period. This represents the average intraocular pressure during the follow-up period; The quantitative assessment formula for the changing trend of visual field defects is as follows: ； in, This indicates the percentage change trend of visual field defects. This represents the average visual field defect value at the most recent follow-up visit. This represents the average visual field defect value during baseline follow-up.

10. An AI-driven early risk prediction and assessment system for glaucoma, used to implement the AI-driven early risk prediction and assessment method for glaucoma as described in any one of claims 1-9, characterized in that, include: Acquisition Module: Acquires multimodal ophthalmic examination data of the target patient, including structural optical coherence tomography image sequences, functional visual field test data sequences, intraocular pressure measurement data, fundus color photographs, and anterior chamber angle examination data; Preprocessing module: Performs standardized preprocessing on multimodal ophthalmological examination data, converts data from different sources and in different formats into a standardized data format that is compatible with the input of artificial intelligence models, and performs noise reduction, missing value imputation and outlier removal operations on each modality of data respectively; Feature extraction module: Performs feature extraction on the preprocessed multimodal ophthalmological examination data to obtain structural image feature vectors, functional visual field feature vectors, intraocular pressure numerical features, fundus image features, and anterior chamber angle structural features; Fusion module: The extracted modal features are fused across modalities to generate a joint latent space representation vector, which encodes the comprehensive representation information of the target patient's ocular structure and function; Output module: Input the joint latent space representation vector into the trained early glaucoma risk prediction model, and output the early glaucoma risk assessment results for the target patient. The risk assessment results include the probability value of glaucoma progression and the risk level determination.

Images