Myopia prediction method and system based on multi-subject interaction

Abstract

The invention discloses a myopia prediction method and system based on multi-subject interaction, and relates to the technical field of medical health information, and the method comprises the steps: synchronously collecting multi-dimensional data through three types of heterogeneous subjects, namely ophthalmology diagnosis and treatment equipment, a wearable monitoring terminal and environment sensing equipment, building a multi-source heterogeneous data set with time-space alignment, and carrying out the multi-source heterogeneous data set; and extracting features of optic disc inclination, macular region choroid thickness, refractive fluctuation amplitude, short-distance eye using duration and sunlight intensity mean value from the multi-source heterogeneous data set. According to the method, by integrating ophthalmology diagnosis and treatment equipment, a wearable monitoring terminal and environment sensing equipment, synchronous collection and space-time alignment of multi-source heterogeneous data are achieved, and multi-dimensional features such as optic disc inclination, choroid thickness, refractive fluctuation, eye using duration and sunlight intensity are extracted; the comprehensive perception ability of the system on structure, physiology, behavior and environmental factors is improved, and the accuracy and dynamic prediction ability of pathological myopia risk identification are enhanced.

Description

Technical Field

[0001] This invention relates to medical and health information technology, and in particular to a method and system for myopia prediction based on multi-subject interaction. Background Technology

[0002] In recent years, with the widespread use of electronic products in daily life and learning, the incidence of myopia among teenagers has continued to rise due to factors such as frequent close-range eye use and insufficient ambient light. In particular, the incidence of pathological myopia is showing a trend towards younger age groups. To address the increasingly severe public health challenges, myopia risk prediction methods that combine ophthalmological clinical practice with artificial intelligence technology have received high attention. Current mainstream research focuses on using single types of data sources (such as fundus images or refractive changes) for modeling and using deep learning algorithms to make preliminary judgments on the progression of myopia in individuals. However, this approach is insufficient to fully capture the complex mechanisms of myopia development under the combined effects of multiple factors.

[0003] However, the occurrence of myopia, especially pathological myopia, is often influenced by the dynamic interaction of multiple factors, including structural indicators of the visual system, functional fluctuations, behavioral characteristics, and environmental factors. Existing technologies have not yet been able to fully integrate data from multiple heterogeneous sources and achieve high-precision spatiotemporal alignment and feature fusion, resulting in limited model predictive capabilities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a myopia prediction method and system based on multi-subject interaction, which solves the problem of spatiotemporal alignment and fusion of multi-source heterogeneous data, and improves the accuracy of pathological myopia risk prediction and real-time early warning capability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a myopia prediction method based on multi-subject interaction, comprising:

[0008] By using three heterogeneous entities—ophthalmic diagnostic and treatment equipment, wearable monitoring terminals, and environmental sensing devices—multi-dimensional data is collected simultaneously to establish a spatiotemporally aligned multi-source heterogeneous dataset.

[0009] Features such as optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity are extracted from multi-source heterogeneous datasets. Multimodal temporal alignment is achieved through dynamic time warping, and normalized fusion feature vectors are generated by combining spatiotemporal coding.

[0010] A pathological myopia prediction model was constructed, and the risk probability of pathological myopia was obtained by normalizing and fusing feature vectors.

[0011] Data is divided into hierarchical time windows, and the weighted cross-entropy loss function is optimized. During training, the Adam optimizer is combined with L2 regularization to update the parameters step by step. Based on the optimal threshold of the validation set and the clinical follow-up period, a continuous warning time step is set. When the risk probability continuously exceeds the threshold, the warning signal is triggered.

[0012] As a preferred embodiment of the myopia prediction method based on multi-subject interaction described in this invention, the method involves simultaneously collecting multi-dimensional data from three heterogeneous subjects—ophthalmic diagnostic and treatment equipment, wearable monitoring terminals, and environmental sensing devices—to establish a spatiotemporally aligned multi-source heterogeneous dataset, including:

[0013] The NTP master clock server is used as the unified synchronization clock source for the three types of devices. The master control terminal uniformly initiates the start acquisition signal, and the synchronization command is broadcast using the MQTT message mechanism. The device data is uniformly in JSON format.

[0014] By using a fully automatic fundus camera with high-definition shooting and OCT functions, fundus image data is collected. Smart glasses that support refractive power, eye movement and IMU are used to collect refractive power and eye behavior data. An ambient light sensor monitors indoor and outdoor light intensity. The head posture data collected by the IMU is used to calculate the direction of gaze by converting quaternions to Euler angles. The position of the ambient light sensor is fused to achieve spatial alignment between the shooting direction and the light direction.

[0015] When a device has no data within a time period, data compensation is performed using linear interpolation based on adjacent valid sampling points.

[0016] A multi-source heterogeneous dataset was constructed based on fundus image data, refractive error, eye use behavior, and light intensity.

[0017] As a preferred embodiment of the myopia prediction method based on multi-subject interaction described in this invention, the following steps are performed: Features such as optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity are extracted from the multi-source heterogeneous dataset; multimodal temporal alignment is achieved through dynamic time warping; and a normalized fusion feature vector is generated by combining spatiotemporal coding.

[0018] The fundus image data acquired by the fully automated fundus camera were used to extract features of optic disc tilt and choroid thickness in the macular region. Canny edge detection was used to automatically segment the optic disc, extract the major and minor axes of the optic disc, and calculate the optic disc tilt.

[0019] After performing RPE layer segmentation on the OCT image, the upper and lower boundaries of the choroid are identified, and the thickness of the subfoveal choroid in the macular region is calculated using the boundary integral method. The refractive fluctuation amplitude features of the refractive power acquired by the smart glasses are extracted.

[0020] Based on smart glasses, the real-time distance from the human eye to the gaze target is measured, and the characteristics of near-field eye use duration are extracted. The average characteristics of sunlight intensity are obtained through an ambient light sensor.

[0021] The features of optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration and average sunlight intensity are used with the fundus image acquisition time as the standard timestamp. The optimal alignment path of the features is calculated by using a dynamic time warping algorithm. Spatiotemporal markers are added to each aligned feature to generate a fused feature vector.

[0022] The fused feature vectors are vertically stacked to form a fused feature matrix, and the fused feature matrix is ​​then normalized according to its standard deviation.

[0023] As a preferred embodiment of the myopia prediction method based on multi-subject interaction described in this invention, the construction of the pathological myopia prediction model, which obtains the risk probability of pathological myopia by normalizing and fusing feature vectors, includes:

[0024] A pathological myopia prediction model is formed by combining a bidirectional GRU network and a causal temporal convolutional network, including an input layer, a causal convolutional layer, a bidirectional GRU layer, a fully connected layer, and an output layer.

[0025] The input layer receives the normalized fused feature vector, and the normalized fused feature vector is input into the causal convolutional layer. Dilated causal convolution is used to stack multiple TCN blocks. Each layer contains dilated convolution, weight normalization, ReLU activation function and residual connection to obtain the convolutional feature vector at each time step.

[0026] The convolutional feature vector is input into a bidirectional GRU layer, which is then processed by the forward and backward GRU layers. Processing yields bidirectional feature vectors;

[0027] The bidirectional feature vectors are input to the output layer, and the risk probability of pathological myopia is calculated through a fully connected layer and a sigmoid activation function.

[0028] As a preferred embodiment of the myopia prediction method based on multi-subject interaction described in this invention, the method involves: dividing the data using hierarchical time windows, optimizing the weighted cross-entropy loss function, employing an Adam optimizer combined with L2 regularization during training to progressively update parameters, setting a continuous warning time step based on the optimal threshold of the validation set and the clinical follow-up period, and triggering a warning signal when the risk probability continuously exceeds the threshold.

[0029] The model training uses a hierarchical time window to divide the data into training, validation, and test sets, and dynamically adjusts the weights based on the distribution of the training set samples.

[0030] The weighted cross-entropy loss is calculated on the samples. The Adam optimizer is combined with L2 regularization. The model parameters are optimized by gradient descent through the weighted cross-entropy loss. After each training cycle, the weighted cross-entropy loss value is calculated on the validation set. If the loss does not decrease in consecutive training cycles, the iteration stops and the model parameters are output.

[0031] Based on the model's predicted probability of pathological myopia risk for each sample in the validation set, the actual clinical diagnosis results and the warning threshold (human-set), the correct warning rate and the correct exclusion rate are calculated, and the maximum threshold of the sum of the correct warning rate and the correct exclusion rate is used as the optimal threshold.

[0032] The number of consecutive warning time steps is obtained based on the clinical follow-up period and data sampling time. When the number of consecutive warning time steps is met, the warning signal is triggered.

[0033] Secondly, the present invention provides a myopia prediction system based on multi-subject interaction, comprising:

[0034] The multi-source heterogeneous data acquisition module simultaneously collects multi-dimensional data from three heterogeneous entities: ophthalmic diagnostic and treatment equipment, wearable monitoring terminals, and environmental sensing devices, and establishes a spatiotemporally aligned multi-source heterogeneous dataset.

[0035] The feature extraction module extracts features such as optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity from multi-source heterogeneous datasets.

[0036] The feature fusion module achieves multimodal temporal alignment through dynamic time warping and generates normalized fusion feature vectors by combining spatiotemporal coding.

[0037] The pathological myopia prediction module constructs a pathological myopia prediction model and obtains the risk probability of pathological myopia by normalizing and fusing feature vectors.

[0038] The training module uses a hierarchical time window to divide the data and optimizes the weighted cross-entropy loss function. During model training, the Adam optimizer is combined with L2 regularization to update the parameters step by step.

[0039] The dynamic early warning module sets a continuous early warning time step based on the optimal threshold of the validation set and the clinical follow-up period. When the risk probability continuously exceeds the threshold, an early warning signal is triggered.

[0040] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the myopia prediction method based on multi-agent interaction as described in the first aspect of the present invention.

[0041] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the myopia prediction method based on multi-agent interaction as described in the first aspect of the present invention.

[0042] The beneficial effects of this invention are as follows: By integrating ophthalmic equipment, wearable terminals, and environmental sensing devices, a multi-source heterogeneous data system capable of synchronously collecting structural, behavioral, and environmental data is constructed; through feature extraction and dynamic time warping, a unified expression and temporal alignment of key indicators are achieved, improving the completeness and accuracy of feature fusion; by combining causal modeling and deep learning algorithms, a pathological myopia risk prediction model is constructed, achieving individualized high-precision risk assessment; the training process is optimized using layered time windows and weighted loss functions, enhancing the model's stability and generalization ability under imbalanced data and temporal features; and the dynamic early warning strategy designed based on a continuous threshold judgment mechanism effectively improves the accuracy and clinical applicability of early warning responses. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of the myopia prediction method based on multi-subject interaction in Example 1.

[0045] Figure 2 This is a schematic diagram of the myopia prediction system based on multi-subject interaction in Example 1. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a myopia prediction method based on multi-subject interaction, including the following steps:

[0050] S1 uses three heterogeneous entities—ophthalmic diagnostic and treatment equipment, wearable monitoring terminals, and environmental sensing devices—to simultaneously collect multi-dimensional data and establish a spatiotemporally aligned multi-source heterogeneous dataset.

[0051] Preferably, multi-dimensional data is collected simultaneously from three heterogeneous entities: ophthalmic diagnostic and treatment equipment, wearable monitoring terminals, and environmental sensing devices, to establish a spatiotemporally aligned multi-source heterogeneous dataset, including...

[0052] Using the NTP master clock server as the unified synchronization clock source for the three types of devices is represented as follows:

[0053]

[0054] in, The standard timestamp after device alignment. This refers to the local data collection time of the i-device. To account for the time offset from the master clock (NTP automatically calculates and adjusts);

[0055] The main control terminal uniformly initiates the start acquisition signal, uses the MQTT message mechanism to broadcast synchronization commands, and uses JSON format for all device data.

[0056] By using a fully automatic fundus camera with high-definition shooting and OCT functions, fundus image data is collected. Smart glasses that support refractive power, eye movement and IMU are used to collect refractive power and eye behavior data. An ambient light sensor is used to monitor indoor and outdoor light intensity.

[0057] By using head pose data collected by the IMU, the direction of gaze is calculated by converting quaternions to Euler angles, and the position of the ambient light sensor is fused to achieve spatial alignment between the shooting direction and the lighting direction.

[0058] When device i has no data at time t, data compensation is performed using linear interpolation based on adjacent valid sampling points, expressed as:

[0059]

[0060] in, The interpolation result at time t. For the time that needs to be interpolated, The effective time of the time step preceding time t. The effective time of the next time step after time t. for The actual values ​​collected above, for The actual value collected above;

[0061] A multi-source heterogeneous dataset was constructed based on fundus image data, refractive error, eye use behavior, and light intensity.

[0062] By constructing a unified time synchronization mechanism with an NTP master clock server at its core, and combining it with MQTT message broadcasting to achieve collaborative data acquisition from multiple devices, and by standardizing the data structure using JSON format, the timing consistency and data compatibility across heterogeneous devices are achieved. This improves the spatiotemporal alignment accuracy of multimodal data, provides a stable and reliable data foundation for pathological myopia risk modeling, and ultimately enhances the accuracy and practicality of prediction results in clinical applications.

[0063] S2 extracts features from multi-source heterogeneous datasets, including optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity. Multimodal temporal alignment is achieved through dynamic time warping, and normalized fusion feature vectors are generated by combining spatiotemporal coding.

[0064] Preferably, features are extracted from the multi-source heterogeneous dataset, including optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity.

[0065] From the fundus image data acquired by the fully automated fundus camera, features of optic disc tilt and choroidal thickness in the macular region are extracted. Canny edge detection is used for automatic optic disc segmentation, the major and minor axes of the optic disc are extracted, and the optic disc tilt is calculated and expressed as:

[0066]

[0067] in, This refers to the tilt angle of the viewing plate. The long axis of the display disc, This is the short axis of the display screen;

[0068] After performing RPE layer segmentation on the OCT image, the upper and lower boundaries of the choroid are identified, and the thickness of the subfoveal choroid in the macular region is calculated using the boundary integral method, expressed as:

[0069]

[0070] in, The thickness of the choroid in the macular region. The width in pixels. Let the coordinates of the upper boundary of the i-th column be . Let be the coordinates of the lower boundary of the i-th column;

[0071] The refractive error amplitude features of the refractive power collected by the smart glasses are extracted and represented as follows:

[0072]

[0073] in, This refers to the amplitude of refractive fluctuation. This represents the number of diopter sampling points. For the k-th refractive power measurement, This represents the average value of the refractive power measurements.

[0074] Based on smart glasses measuring the real-time distance from the human eye to the gaze target, the near-field eye usage time feature is extracted and represented as:

[0075]

[0076] in, For the duration of close-range eye use, Sampling frequency, The number of sampling points for eye distance. Let k be the eye-object distance in the kth sample. For near-distance threshold, This is an indicator function; it takes the value 1 when the condition is true, and 0 otherwise.

[0077] The average characteristics of sunlight intensity obtained through an ambient light sensor are represented as follows:

[0078]

[0079] in, The average solar intensity This represents the number of light sampling points. Let be the light intensity at time t;

[0080] By automatically extracting features of optic disc tilt and choroidal thickness in fundus images, and combining OCT image layering and boundary integration, accurate modeling of macular microstructural parameters is achieved, effectively capturing fundus anatomical features and microcirculatory changes related to pathological myopia. Furthermore, refractive fluctuation amplitude is extracted based on refractive power data collected by smart glasses to reflect the accommodative ability and stability of the visual system, improving the ability to identify dynamic abnormalities in individual refractive states. At the same time, by combining ranging function and light sensor, the average duration of near-vision use and sunlight intensity are extracted, realizing the objective quantification of eye use behavior and environmental factors, and enhancing the model's comprehensive perception of external inducing factors.

[0081] Furthermore, multimodal temporal alignment is achieved through dynamic time warping, and normalized fusion feature vectors are generated by combining spatiotemporal coding, including...

[0082] Features such as optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity are used with the fundus image acquisition time as the standard timestamp. A dynamic time warping algorithm is employed to calculate the optimal alignment path for these features. A spatiotemporal marker is then added to each aligned feature to generate a fused feature vector, represented as follows:

[0083]

[0084] in, Let i be the fused feature vector of the i-th feature at time step k. For the i-th feature at time step k, For time-coded frequency, For spatial coding frequency, For time step k, the standard timestamp. The spatial coordinates of time step k (IMU positioning data);

[0085] The fused feature vectors are stacked vertically to form a fused feature matrix, represented as follows:

[0086]

[0087] in, To fuse the feature matrix, For the fusion feature of the disc tilt at time k, The fusion characteristics of macular choroid thickness at time k. The refractive fluctuation amplitude fusion characteristics at time k, For near-field eye use duration fusion features at time k, The mean alignment feature of sunlight intensity at time k;

[0088] The fused feature matrix is ​​normalized by standard deviation and expressed as follows:

[0089]

[0090] in, Let be the normalized feature vector at the i-th time step. To fuse the i-th time eigenvector of the feature matrix, Let be the mean vector of all time-step features. The standard deviation vector of all time-step features;

[0091] By using multidimensional features such as optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity as a unified timestamp, and combining them with a dynamic time warping algorithm to calculate the optimal alignment path for each feature, and adding spatiotemporal markers to generate a fused feature vector, precise alignment and fusion of heterogeneous data at both temporal and semantic levels are achieved. This effectively solves the problems of temporal asynchrony and semantic offset between multi-source data. Furthermore, by vertically stacking the fused feature vectors to construct a fused feature matrix and normalizing the standard deviation of each feature dimension, scale consistency and numerical normalization of multimodal inputs are achieved, thereby improving the model's training stability, feature representation ability, and generalization performance, and enhancing the accuracy, robustness, and clinical applicability of the pathological myopia prediction model.

[0092] S3, construct a pathological myopia prediction model, and obtain the risk probability of pathological myopia by normalizing and fusing feature vectors;

[0093] Preferably, a pathological myopia prediction model is constructed, and the risk probability of pathological myopia is obtained by normalizing and fusing feature vectors, including:

[0094] A pathological myopia prediction model is formed by combining a bidirectional GRU network and a causal temporal convolutional network, including an input layer, a causal convolutional layer, a bidirectional GRU layer, a fully connected layer, and an output layer.

[0095] The input layer receives the normalized and fused feature vector, which is then input into a causal convolutional layer. Dilated causal convolution is used, stacking multiple TCN blocks. Each layer contains dilated convolution, weight normalization, ReLU activation function, and residual connections, resulting in the convolutional feature vector for each time step, represented as:

[0096]

[0097] in, Let be the convolutional feature vector at time step t. For the mapping function of causal temporal convolutional networks, Let be the normalized fused feature vector at the t-th time step;

[0098] The convolutional feature vector is input into a bidirectional GRU layer, which is then processed by the forward and backward GRU layers. After processing, a bidirectional feature vector is obtained, represented as:

[0099]

[0100]

[0101]

[0102] in, Forward at time step t Forward eigenvectors Forward GRU unit, Let be the forward eigenvector at time step t-1. Backwards at time step t Backward eigenvectors For backward GRU units, This is the backward eigenvector at time step t+1. This is the bidirectional eigenvector at time t;

[0103] The bidirectional feature vectors are input to the output layer, and the probability of pathological myopia risk is calculated through a fully connected layer and a sigmoid activation function, as follows:

[0104]

[0105]

[0106] in, The pathological myopia risk score at time t. This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, The probability of predicting the risk of pathological myopia at time t;

[0107] By inputting bidirectional feature vectors into the output layer and then sequentially passing them through a fully connected layer and a Sigmoid activation function, the risk probability of pathological myopia is calculated, realizing the transformation from high-dimensional fused features to specific risk scores. This process enables the model to have quantifiable and continuous risk assessment capabilities, facilitating subsequent integration with a threshold-triggered continuous early warning mechanism and providing clinicians with operable reference indicators, effectively assisting in myopia screening and intervention decisions.

[0108] S4 uses a hierarchical time window to divide the data, optimizes the weighted cross-entropy loss function, and uses the Adam optimizer combined with L2 regularization during training to gradually update the parameters. Based on the optimal threshold of the validation set and the clinical follow-up period, it sets the number of consecutive warning time steps and triggers a warning signal when the risk probability continuously exceeds the threshold.

[0109] Preferably, a tiered time window is used to partition the data, the weighted cross-entropy loss function is optimized, and the Adam optimizer combined with L2 regularization is used during training to progressively update the parameters. This includes dividing the data into training, validation, and test sets using tiered time windows during model training, and dynamically adjusting the weights based on the distribution of the training set samples, as expressed below:

[0110]

[0111]

[0112] in, For positive sample weights, For negative sample weights, The total number of samples, The number of positive samples. The number of negative samples;

[0113] The weighted cross-entropy loss of the samples is calculated as follows:

[0114]

[0115] in, For weighted cross-entropy loss, This is a clinical diagnostic result;

[0116] The Adam optimizer is used in combination with L2 regularization. The model parameters are optimized by gradient descent through weighted cross-entropy loss. After each training cycle, the weighted cross-entropy loss value is calculated on the validation set. If the loss does not decrease in consecutive training cycles, the iteration stops and the model parameters are output.

[0117] By employing a hierarchical time window to divide the training, validation, and test sets, and dynamically adjusting sample weights based on the training sample distribution, joint modeling control of the phased changes in time-series features and class imbalance is achieved, improving the model's ability to identify early anomalies and high-risk samples. Furthermore, a weighted cross-entropy loss function is introduced, combined with the Adam optimizer and L2 regularization for parameter optimization, enhancing the model's classification accuracy and generalization ability under complex sample structures. Simultaneously, the loss is dynamically evaluated on the validation set after each training round, and an early stopping mechanism is set to stop training prematurely if the loss does not decrease for several consecutive rounds, achieving stability control and overfitting suppression during the training process, thereby ultimately improving the model's learning efficiency, convergence quality, and prediction stability.

[0118] Furthermore, based on the optimal threshold of the validation set and the clinical follow-up period, a continuous warning time step is set. When the risk probability continuously exceeds the threshold, a warning signal is triggered, including...

[0119] Based on the model's predicted probability of pathological myopia risk for each sample in the validation set, the actual clinical diagnosis results, and the warning threshold (human-set), the correct warning rate and correct exclusion rate are calculated and expressed as follows:

[0120]

[0121]

[0122] in, To ensure accurate early warning rate, For the correct exclusion rate, The total number of samples in the validation set, As the warning threshold, To predict the probability of pathological myopia risk for the j-th sample, This represents the actual clinical diagnosis result of the j-th sample, where 0 indicates healthy and 1 indicates diseased.

[0123] The optimal threshold is defined as the maximum threshold of the sum of the correct warning rate and the correct exclusion rate, expressed as:

[0124]

[0125] in, The optimal threshold;

[0126] The number of consecutive early warning time steps is obtained based on the clinical follow-up period and data sampling time, and is expressed as follows:

[0127]

[0128] in, The number of consecutive warning time steps. For the clinical follow-up period, For data sampling time;

[0129] When the consecutive warning time steps are met, a warning signal is triggered, as shown below:

[0130]

[0131] in, This is a warning signal;

[0132] By combining the actual clinical follow-up period with the model data sampling interval, a reasonable number of continuous warning time steps is set, and a warning signal is triggered when the risk probability continuously exceeds the optimal threshold, thus realizing a dynamic warning mechanism with time continuity criteria. This mechanism eliminates the risk of false alarms or missed alarms caused by fluctuations in single-time point predictions, and introduces "persistent risk performance" as the judgment criterion, which is more in line with the clinical observation period and intervention window characteristics of pathological progression. It filters out misjudgments caused by short-term abnormal values ​​while enhancing the ability to respond to real risk trends, ultimately improving the stability, practicality and forward-looking nature of the warning.

[0133] This embodiment also provides a myopia prediction system based on multi-subject interaction, including:

[0134] The multi-source heterogeneous data acquisition module simultaneously collects multi-dimensional data from three heterogeneous entities: ophthalmic diagnostic and treatment equipment, wearable monitoring terminals, and environmental sensing devices, and establishes a spatiotemporally aligned multi-source heterogeneous dataset.

[0135] The feature extraction module extracts features such as optic disc tilt, macular choroid thickness, refractive fluctuation amplitude, near-vision duration, and mean sunlight intensity from multi-source heterogeneous datasets.

[0136] The feature fusion module achieves multimodal temporal alignment through dynamic time warping and generates normalized fusion feature vectors by combining spatiotemporal coding.

[0137] The pathological myopia prediction module constructs a pathological myopia prediction model and obtains the risk probability of pathological myopia by normalizing and fusing feature vectors.

[0138] The training module uses a hierarchical time window to divide the data and optimizes the weighted cross-entropy loss function. During model training, the Adam optimizer is combined with L2 regularization to update the parameters step by step.

[0139] The dynamic early warning module sets a continuous early warning time step based on the optimal threshold of the validation set and the clinical follow-up period. When the risk probability continuously exceeds the threshold, an early warning signal is triggered.

[0140] This embodiment also provides a computer device applicable to the myopia prediction method based on multi-subject interaction, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the myopia prediction method based on multi-subject interaction as proposed in the above embodiment.

[0141] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0142] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the myopia prediction method based on multi-agent interaction as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0143] In summary, this invention improves the model's ability to model the risk of pathological myopia by: integrating ophthalmic images, eye-use behavior, and environmental data to construct a multi-source heterogeneous data system; achieving unified expression through feature extraction and temporal alignment; and combining optimized training and dynamic early warning strategies to effectively improve prediction accuracy and system stability, demonstrating good clinical applicability and promotional value.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A myopia prediction method based on multi-agent interaction, characterized by: The application relates to a pathological myopia prediction method based on multi-source heterogeneous data set. Synchronous acquisition of multi-dimensional data by three types of heterogeneous subjects, namely, an ophthalmic diagnosis and treatment device, a wearable monitoring terminal and an environment sensing device, to establish a multi-source heterogeneous data set with time and space alignment; Features of disc tilt, macular choroidal thickness, refractive fluctuation amplitude, near-distance eye use time length and average sunlight intensity of the multi-source heterogeneous data set are extracted; Multi-modal time sequence alignment is realized through dynamic time warping, and a normalized fusion feature vector is generated in combination with time and space coding; A pathological myopia prediction model is constructed, and a pathological myopia risk probability is obtained through the normalized fusion feature vector; A layered time window is used to divide data, a weighted cross-entropy loss function is optimized, an Adam optimizer is used in time training in combination with L2 regularization, and parameters are gradually updated; Based on the optimal threshold value of the verification set and the clinical follow-up cycle, the number of continuous early warning time steps is set, and an early warning signal is triggered when the risk probability continuously exceeds the threshold value. 2.The myopia prediction method based on multi-agent interaction of claim 1, wherein: The application relates to a pathological myopia prediction method based on multi-source heterogeneous data set. An NTP master clock server is used as a unified synchronous clock source of the three types of devices, a master control end uniformly initiates a start collection signal, a synchronization command is broadcast using an MQTT message mechanism, and device data is uniformly in JSON format; Through use of a full-automatic fundus camera with high-definition shooting and OCT functions, fundus image data are collected, diopter, eye movement and IMU intelligent glasses are used to collect diopter and eye use behavior data, an ambient light sensor is used to monitor indoor and outdoor light intensity, head posture data collected through the IMU are used to calculate the line-of-sight direction by using a quaternion to Euler angle conversion, and the position of the ambient light sensor is fused to realize space alignment of the shooting direction and the light direction; When there is no data of a device in a time period, linear interpolation based on adjacent effective sampling points is used for data compensation; A multi-source heterogeneous data set is constructed through the fundus image data, the diopter, the eye use behavior and the light intensity. 3.The myopia prediction method based on multi-agent interaction of claim 2, wherein: Features of disc tilt, macular choroidal thickness, refractive fluctuation amplitude, near-distance eye use time length and average sunlight intensity of the multi-source heterogeneous data set are extracted; After the fundus image data collected by the full-automatic fundus camera are processed, features of disc tilt and macular choroidal thickness are extracted, Canny edge detection is used for automatic disc segmentation, the long axis and the short axis of the disc are extracted, and the disc tilt is calculated; After RPE layer segmentation of the OCT image, the upper and lower boundaries of the choroid are identified, the boundary integral method is used to calculate the macular fovea choroidal thickness, and the diopter collected by the intelligent glasses is used for feature extraction of the refractive fluctuation amplitude; The real-time distance from the eye to the fixation target is measured based on the intelligent glasses, the near-distance eye use time length feature is extracted, and the average sunlight intensity feature is obtained through the ambient light sensor. 4.The myopia prediction method based on multi-agent interaction of claim 3, wherein: ​ The features of the disc tilt, macular choroidal thickness, refractive fluctuation amplitude, near work duration and average sunlight intensity are time-stamped with the fundus image collection time as the standard, and the optimal alignment path of the features is calculated by the dynamic time warping algorithm. The space-time label is added to each aligned feature to generate a fusion feature vector; The fusion feature vector is vertically stacked to form a fusion feature matrix, and the fusion feature matrix is subjected to standard deviation normalization processing. 5.The myopia prediction method based on multi-agent interaction of claim 4, wherein: The pathological myopia prediction model is constructed, and the pathological myopia risk probability is obtained by normalizing the fusion feature vector, including, The bidirectional GRU network and the causal time series convolution network are combined to form a pathological myopia prediction model, including an input layer, a causal convolution layer, a bidirectional GRU layer, a full connection layer and an output layer; The input layer receives the normalized fusion feature vector, and the normalized fusion feature vector is input into the causal convolution layer. An expanded causal convolution is used, and multiple layers of TCN blocks are stacked. Each layer includes an expanded convolution, weight normalization, a ReLU activation function and a residual connection to obtain a convolution feature vector at each time step. The convolution feature vector is input into a bidirectional GRU layer, and a forward GRU and a backward GRU are used to obtain a bidirectional feature vector through processing. processing, to obtain a bidirectional feature vector; The bidirectional feature vector is input into the output layer, and the pathological myopia risk probability is calculated by the full connection layer and the Sigmoid activation function. 6.The myopia prediction method based on multi-agent interaction of claim 5, wherein: The data is divided by using a hierarchical time window, the weighted cross-entropy loss function is optimized, the Adam optimizer is used in training combined with L2 regularization, and the parameters are updated step by step, including, The model training divides the data into training set, validation set and test set by using a hierarchical time window, and dynamically adjusts the weights according to the sample distribution of the training set; The sample is subjected to weighted cross-entropy loss calculation, and the Adam optimizer is used combined with L2 regularization. The model parameters are optimized by gradient descent through the weighted cross-entropy loss. After completing a training cycle, the weighted cross-entropy loss value is calculated on the validation set. If the loss does not decrease for consecutive training cycles, the iteration of the model parameters is stopped. 7.The myopia prediction method based on multi-agent interaction of claim 6, wherein: Based on the optimal threshold value of the validation set and the clinical follow-up period, the number of consecutive early warning time steps is set, and the early warning signal is triggered when the risk probability continuously exceeds the threshold value, including, Based on the pathological myopia risk prediction probability given by the model for each sample in the validation set, the actual clinical diagnosis result and the early warning threshold value, the correct early warning rate and the correct exclusion rate are calculated. The maximum threshold value of the sum of the correct early warning rate and the correct exclusion rate is used as the optimal threshold value; According to the clinical follow-up period and the data sampling time, the number of consecutive early warning time steps is obtained, and the early warning signal is triggered when the number of consecutive early warning time steps is met.

8. A myopia prediction system based on multi-agent interaction, based on the myopia prediction method based on multi-agent interaction according to any one of claims 1-7, characterized in that: It includes, A multi-source heterogeneous data acquisition module synchronously acquires multi-dimensional data through three types of heterogeneous subjects, namely ophthalmic diagnosis and treatment equipment, wearable monitoring terminals and environmental perception equipment, and establishes a spatio-temporally aligned multi-source heterogeneous data set; A feature extraction module extracts disc tilt, macular choroidal thickness, refractive fluctuation amplitude, near work duration and average sunlight intensity features from the multi-source heterogeneous data set; A fusion feature module realizes multi-modal time alignment through dynamic time warping and generates a normalized fusion feature vector by combining spatio-temporal coding; A pathological myopia prediction module constructs a pathological myopia prediction model and obtains a pathological myopia risk probability through a normalized fusion feature vector. The training module divides data by using a hierarchical time window, optimizes a weighted cross-entropy loss function, and uses an Adam optimizer combined with L2 regularization to update parameters step by step during model training. The dynamic early warning module sets a number of continuous early warning time steps based on the optimal threshold value of the verification set and the clinical follow-up cycle, and triggers an early warning signal when the risk probability continuously exceeds the threshold value. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor implements the steps of the myopia prediction method based on multi-agent interaction according to any one of claims 1-7 when executing the computer program.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The processor implements the steps of the myopia prediction method based on multi-agent interaction according to any one of claims 1-7 when executing the computer program.

Images