An ai-based method and system for rapid screening of pseudo-myopia and a medium

Abstract

The application relates to the technical field of data processing, and discloses an AI-based rapid screening method and system for pseudo myopia and a medium. The method comprises the following steps: collecting eye movement and pupil signals, forming an eye movement feedback index set, extracting a multi-modal feature vector, inputting the multi-modal feature vector into a cascade forest classifier to calculate a category probability, screening a target feature combination, and finally generating a pseudo myopia grading report. The application improves the accuracy and efficiency of pseudo myopia screening.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an AI-based method and system for rapid screening of pseudo-myopia and a medium. BACKGROUND

[0002] Myopia is a common vision problem, especially in the adolescent population, with an increasing incidence year by year. According to the clinical classification of ophthalmology, myopia can be divided into two main types: pseudo-myopia and true myopia. Pseudo-myopia refers to temporary myopia caused by ciliary muscle spasm due to long-term close eye use, which causes the lens to adjust excessively. In theory, it can be improved or reversed through reasonable eye use and rest. True myopia is irreversible myopia caused by anatomical changes such as eye axis growth and corneal curvature change. Traditional myopia screening methods mainly rely on visual acuity chart examination and optometry instruments, including mydriatic refraction, A-scan examination of eye axis length, etc. These methods can evaluate the refractive state and structural parameters of the eye to some extent.

[0003] There are various myopia screening methods in the prior art, but most of them have obvious shortcomings. Traditional visual acuity chart examination is simple, but it is difficult to distinguish between different types of myopia and relies on the subjective feedback of the tested person, so its accuracy is limited. Mydriatic refraction can accurately distinguish between pseudo-myopia and true myopia, but it requires the use of a ciliary muscle paralyzing agent, which is complex to operate and time-consuming, and is not suitable for large-scale screening. A-scan examination of eye axis length is objective and reliable, but the equipment is expensive and the operation is professional, so it is also not suitable for widespread application. In addition, some existing computer vision-based screening techniques often only focus on a single physiological indicator, such as fundus photo analysis or corneal topography detection, lacking comprehensive evaluation of multiple visual function indicators, making it difficult to accurately identify early signs of pseudo-myopia. In particular, there is a lack of effective monitoring and analysis of pupil accommodation function and eye movement characteristics, which are two key indicators of pseudo-myopia.

[0004] Since the cause of pseudo-myopia mainly involves ciliary muscle accommodation dysfunction, existing screening methods cannot effectively quantify the actual accommodation ability of the ciliary muscle at different distances, resulting in the inability to accurately assess the difference between the ciliary muscle torque change range and the theoretical normal value. At the same time, traditional methods lack fusion analysis of multi-modal physiological signals such as eye movement tracking and pupil response, and cannot construct a comprehensive evaluation system that reflects the multi-dimensional physiological characteristics of pseudo-myopia. Furthermore, existing technologies lack an artificial intelligence-based cascade classifier to process complex multi-modal feature data, making it impossible to achieve precise classification of pseudo-myopia and generation of personalized intervention recommendations, which directly affects the accuracy of early identification of pseudo-myopia and the efficiency of screening, making it difficult to meet the actual needs of large-scale adolescent vision health screening. SUMMARY

[0005] The application provides an AI-based pseudomyopia rapid screening method, system and medium, which solves the problem that in the prior art, eye movement tracking, pupil response, ciliary muscle adjustment and other multi-modal physiological signals cannot be effectively fused for accurate identification and classification of pseudomyopia, and improves the accuracy and efficiency of pseudomyopia screening.

[0006] In a first aspect, the application provides an AI-based pseudomyopia rapid screening method, which comprises: collecting and processing eye movement original signal sequences and pupil response original signal sequences of a testee to obtain an eye movement feedback index set; performing multi-modal feature extraction and splicing on the eye movement feedback index set to obtain a multi-modal fusion feature vector; constructing a cascade forest classifier based on random forest and complete random forest two forest algorithms, inputting the multi-modal fusion feature vector into the cascade forest classifier to calculate a category probability distribution, and obtaining a sample classification result; performing sample feature screening on the sample classification result to obtain a target feature combination; and classifying the testee according to the target feature combination to obtain a pseudomyopia screening report.

[0007] In a second aspect, the application provides an AI-based pseudomyopia rapid screening system, which comprises:

[0008] A processing module for collecting and processing eye movement original signal sequences and pupil response original signal sequences of a testee to obtain an eye movement feedback index set;

[0009] A splicing module for performing multi-modal feature extraction and splicing on the eye movement feedback index set to obtain a multi-modal fusion feature vector;

[0010] An input module for constructing a cascade forest classifier based on random forest and complete random forest two forest algorithms, inputting the multi-modal fusion feature vector into the cascade forest classifier to calculate a category probability distribution, and obtaining a sample classification result;

[0011] A screening module for performing sample feature screening on the sample classification result to obtain a target feature combination;

[0012] A classification module for classifying the testee according to the target feature combination to obtain a pseudomyopia screening report.

[0013] In a third aspect, an AI-based pseudomyopia rapid screening device is provided, which comprises a memory and at least one processor, the memory stores instructions; the at least one processor invokes the instructions in the memory, so that the AI-based pseudomyopia rapid screening device executes the AI-based pseudomyopia rapid screening method described above.

[0014] In a fourth aspect, a computer readable storage medium is provided, which stores instructions that, when executed on a computer, cause the computer to perform the AI-based rapid screening method for pseudo-myopia described above.

[0015] In the technical scheme provided in the present application, the eye movement original signal sequence and the pupil reaction original signal sequence of the measured person are collected and processed professionally, so that an eye movement feedback index set containing the eye muscle adjustment amplitude, the pupil contraction range, the iris change amplitude, the lens thickness change amplitude and the ciliary muscle torque change range can be obtained. The comprehensive collection of multi-dimensional physiological characteristics overcomes the limitation of traditional methods which rely on only a single index, and provides rich physiological basis for accurate identification of pseudo-myopia. The technical scheme of multi-modal feature extraction and splicing to obtain a multi-modal fusion feature vector effectively fuses physiological information of different sources such as eye muscle features, pupil features, iris features and lens features, eliminates the dimensional difference and numerical range difference through principal component analysis and Z-score standardization processing, so that various features have balanced weight distribution in subsequent analysis, and the completeness and reliability of feature expression are significantly improved. The cascade forest classifier constructed based on the random forest and the complete random forest two forest algorithms adopts a three-layer structure design, each layer contains the configuration of 4 forests, which not only guarantees the complexity of the model but also avoids the risk of overfitting. Through the feature transmission and enhancement mechanism between layers, more abstract feature representations can be learned layer by layer, and compared with the traditional single classifier, the cascade forest classifier has stronger nonlinear feature relationship processing ability and higher classification accuracy. The technical features of the target feature combination are obtained by sample feature screening of the sample classification results, and through the strategy of combining recursive feature elimination and correlation analysis, the key features that contribute most to the classification of pseudo-myopia are retained, and the redundant information between features is eliminated, which significantly reduces the computational complexity under the premise of ensuring the classification performance, so that the method is more suitable for the rapid screening demand in the actual application scene.

[0016] The quantitative evaluation algorithm of the range of ciliary muscle torque change can accurately reflect the real state of the eye accommodation function, and by setting multiple target distance test points and calculating the difference between the theoretical myopia degree and the actual myopia degree, it provides an objective physiological criterion for distinguishing between pseudomyopia and true myopia. The random forest in the cascade forest classifier optimizes the feature selection strategy through the Gini coefficient, and the completely random forest enhances the model generalization ability through high randomization. The synergistic work of the two algorithms is particularly suitable for dealing with the actual problems existing in pseudomyopia screening, such as large individual differences, much feature noise, and unbalanced samples. Compared with traditional machine learning methods, it has obvious advantages in the reliability and stability of medical diagnosis. The multi-modal feature fusion algorithm is specially designed for the diversity and complexity of eye physiological signals. The time-frequency features of the pupil signal are extracted through wavelet transform, and the dynamic characteristics of the eye muscle accommodation are captured through frequency domain analysis. These targeted signal processing methods fully consider the performance characteristics of pseudomyopia in different physiological dimensions, so that the algorithm can more accurately capture the early signs and subtle changes of pseudomyopia, and provide strong technical support for early clinical intervention. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0018] Figure 1 An embodiment schematic diagram of the AI-based pseudomyopia rapid screening method in the embodiments of the present application;

[0019] Figure 2 An embodiment schematic diagram of the AI-based pseudomyopia rapid screening system in the embodiments of the present application. DETAILED DESCRIPTION

[0020] The embodiment of the present application provides a kind of based on AI's false myopia rapid screening method, system and medium.The terms "first", "second", "third", "fourth" and the like (if exist) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence.It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below, please refer to Figure 1 One embodiment of the false myopia rapid screening method based on AI in the embodiment of the present application includes:

[0022] Step S101, collect the eye movement original signal sequence and the pupil response original signal sequence of the measured person and process to obtain an eye movement feedback index set;

[0023] Step S102, multi-modal feature extraction and splicing are carried out on the eye movement feedback index set to obtain a multi-modal fusion feature vector;

[0024] Step S103, a cascade forest classifier is constructed based on random forest and complete random forest two kinds of forest algorithm, the multi-modal fusion feature vector is input into the cascade forest classifier to calculate the category probability distribution, and the sample classification result is obtained;

[0025] Step S104, sample feature screening is carried out on the sample classification result to obtain a target feature combination;

[0026] Step S105, according to the target feature combination, the measured person is classified as false myopia, and a false myopia screening report is obtained.

[0027] It can be understood that the execution subject of the present application can be a false myopia rapid screening system based on AI, and can also be a terminal or a server, and the specific place is not limited.The embodiment of the present application takes the server as the execution subject for example.

[0028] Specifically, the eye movement raw signal sequence and the pupil response raw signal sequence of the measured person are collected and processed to obtain an eye movement feedback index set. The horizontal direction coordinate x-axis and the vertical direction coordinate y-axis of the eyeball of the measured person are recorded by the eye tracking device at a frequency of 250 frames per second, and the pupil diameter size change is collected at a frequency of 100 frames per second. Since the collection frequencies of the two signals are different, a linear interpolation algorithm is used to resample the two signals to a uniform sampling rate of 200 Hz, ensuring that the data points correspond one by one in time. A Butterworth band-pass filter is used to set the upper boundary frequency to 45 Hz and the lower boundary frequency to 0.5 Hz, to filter out high-frequency electronic noise and low-frequency baseline drift. Then, normalization processing is performed, the eye movement signal is normalized to the 0 to 1 interval using the maximum and minimum value normalization method, and the pupil signal is converted to a distribution with a mean of 0 and a standard deviation of 1 using the Z-score standardization method. Threshold method is used to detect blinking artifacts, and when the signal value is greater than 95% or less than 5%, it is marked as a missing value. A cubic spline interpolation method is used to construct a cubic polynomial to smoothly transition between adjacent data points to complete data completion. The eye muscle adjustment amplitude is obtained by calculating the contraction and relaxation change amplitude of the extraocular muscle during eye movement, and the pupil contraction range is obtained by analyzing the maximum change in pupil diameter under standard light conditions. The ciliary muscle torque change range measurement needs to set a 30 cm near test point and 3 m, 5 m, 10 m, 100 m far test points, record the ciliary muscle contraction 3.00D adjustment state when looking at 30 cm, and the ciliary muscle expansion 1.00D, 1.50D, 2.00D, 3.00D adjustment state when looking far, calculate the total torque change of the ciliary muscle in the range from 10 meters to 30 cm, which should reach 5D torque change in normal state, and the difference between the measured value and the theoretical value is used to evaluate the degree of ciliary muscle fatigue.

[0029] The eye movement feedback index set is subjected to multi-modal feature extraction and splicing to obtain a multi-modal fusion feature vector. The eye muscle adjustment amplitude is calculated by the difference method to obtain the change amount of adjacent sampling points. The difference between the two adjustment amplitude values on the time sequence is calculated and then divided by the time interval of 5 milliseconds to obtain the instantaneous adjustment speed, forming an eye muscle adjustment speed sequence. The statistical features of the sequence include maximum, minimum, average, and standard deviation. Meanwhile, a fast Fourier transform is performed to convert the time domain signal to the frequency domain, and the energy distribution features of the 0.1-1 Hz slow adjustment, 1-5 Hz medium adjustment, and 5-10 Hz fast adjustment sub-frequency bands are extracted. The percentage of energy in each sub-band to the total energy constitutes the eye muscle feature set. The pupil contraction range analysis calculates the pupil diameter change rate defined as the average change amount of the pupil diameter per unit time. The ratio of the maximum diameter to the minimum diameter of the pupil before and after the light change is measured to eliminate the difference in the basic size of the pupil between individuals. According to the time sequence characteristics, the process is divided into the pupil contraction period and the pupil recovery period, and the time length and change rate of the two stages are calculated to form the pupil feature set. The Daubechies wavelet basis function is selected to perform 5-layer wavelet decomposition on the pupil signal, and the detail coefficients D1-D5 and the approximation coefficient A5 corresponding to different frequency ranges are obtained. The energy entropy value, mean square deviation, and extreme value ratio of each layer of wavelet coefficients are calculated, and special attention is paid to the 0.3-0.5 Hz frequency band energy ratio related to visual adjustment function to constitute the pupil wavelet feature set. The eye muscle feature set, pupil wavelet feature set, iris change amplitude feature, and lens thickness change feature are spliced in order of feature dimension to form a 17-20 dimensional feature vector. The principal component analysis algorithm is applied to find the main variation direction in the data, and the related variables are converted to linearly uncorrelated principal components. The principal components with a cumulative contribution rate of more than 95% are selected to reduce to 12-15 dimensions, and Z-score standardization processing is performed to convert each feature component to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0030] Cascade forest classifier is constructed based on two kinds of forest algorithms, random forest and complete random forest. Three layers are set, each layer contains 4 forests, 2 random forests and 2 complete random forests, each forest contains 100 decision trees, forming a powerful ensemble of 1200 decision trees. Random forest randomly selects about 5 features from the 25-dimensional multi-modal fusion feature vector as candidate features at each node of each decision tree, calculates the Gini coefficient of each candidate feature, and selects the feature with the smallest Gini coefficient as the split feature and determines the optimal split point. Complete random forest directly selects a feature from the 25-dimensional feature vector as the split feature, and selects the median of the feature as the split point to increase model diversity and prevent overfitting. The multi-modal fusion feature vector is input into the first layer of the cascade forest, and the training data is divided into 5 parts using five-fold cross-validation to maintain the same proportion of each class of samples. The remaining 1 part is used to predict the class probability, and each forest outputs the probability values of 3 classes, normal vision, pseudo myopia and true myopia. Four forests generate 12 probability features to form the first layer output feature. The 12 probability features output by the first layer are concatenated with the original 25-dimensional multi-modal fusion feature vector to form a 37-dimensional enhanced feature vector, which is input into the second layer of the cascade forest. The cross-validation prediction process is repeated to obtain the second layer output feature, which is again concatenated and input into the third layer of the cascade forest. The final layer forest output is obtained by weighted averaging the feature transfer and enhancement between layers to obtain the classification results of three types of samples.

[0031] The sample classification results are screened for sample features to obtain the target feature combination. The average impurity reduction method is used to calculate the feature contribution, and each decision tree in the cascade forest classifier is traversed to count the nodes using a specific feature as a split condition. The impurity change before and after the split is calculated using the Gini index, and the average impurity reduction of the same feature in all trees is calculated to obtain the importance score of the feature. According to the feature importance score, all 25 features are ranked in descending order to establish a feature importance ranking list. The recursive feature elimination algorithm is executed to remove the least important feature from the complete feature set each time the iteration starts, retrain the cascade forest classifier with the remaining features, and evaluate the performance. The number of features and the corresponding classification performance are recorded to plot the relationship between the number of features and the classification effect. The inflection point where the slope of the curve changes significantly indicates that further reducing the number of features will significantly reduce the performance. The number of features corresponding to the inflection point is determined, and the corresponding number of features with the highest importance from the feature priority sequence is taken to form the initial feature set. Correlation analysis is performed on the initial feature set to calculate the Pearson correlation coefficient between features. When a feature pair with an absolute correlation coefficient greater than 0.8 is found, it is considered to contain a large amount of redundant information, and the feature with higher importance is retained and the other is removed to obtain a low-redundancy feature set as the target feature combination.

[0032] The target feature combination is used to classify the subjects with pseudo myopia, and a pseudo myopia screening report is obtained. The pseudo myopia degree index is calculated based on the target feature combination. The pseudo myopia severity is quantified by weighted fusion of the pupil accommodation flexibility feature, the eye movement fixation stability feature and the visual fatigue degree feature. Each feature is first normalized to have a value range of 0 to 1, then multiplied by the corresponding weight and summed to obtain the final index value. The theoretical myopia degree is calculated according to the ciliary muscle torque change range. The actual visual state is evaluated by the actual accommodation ability of the ciliary muscle. The actual myopia degree of the subject is measured for comparative analysis. When the actual myopia degree is in the range of -0.25D to -1.25D and the ciliary muscle torque change range is less than the normal 5D change, it is determined as pseudo myopia. When the actual myopia degree is in the range of -0.25D to -1.50D and the theoretical myopia degree is calculated as 0.00D or above, it is also determined as pseudo myopia. When the actual myopia degree is less than the theoretical myopia degree, it is determined as ciliary muscle fatigue with rapid development trend of myopia. According to the pseudo myopia degree index, the pseudo myopia is subdivided into three grades: mild 0 to 0.3, moderate 0.3 to 0.7 and severe 0.7 to 1.0. The difference between the highest class probability and the second highest class probability output by the cascade forest classifier is taken as the classification confidence. Samples with a confidence lower than 0.2 are marked for further inspection. Personalized eye use recommendations are generated for different degrees of pseudo myopia. For mild pseudo myopia, it is recommended to rest for 10 minutes after using eyes for 30-40 minutes, and the ambient light should be kept at 300-500 lux. For moderate pseudo myopia, the continuous eye use time is suggested to be shortened to 20-30 minutes, and the rest time is increased to 15 minutes. For severe pseudo myopia, the eye use time needs to be more strictly controlled, and the subject must rest for 15 minutes every 15-20 minutes. The confidence of the pseudo myopia grading result is marked, and the ciliary muscle fatigue state evaluation and the visual health guidance scheme are integrated into a comprehensive screening report. The screening result interpretation instructions are added, including the causes of pseudo myopia, the development trend and the prevention and control measures. A pseudo myopia screening report is formed.

[0033] In a specific embodiment, the process of step S101 can specifically include the following steps:

[0034] The eye movement original signal sequence of the subject is collected by the eye movement tracking device, and the pupil response original signal sequence of the subject is collected at the same time, to obtain an eye movement vector containing horizontal direction coordinates, vertical direction coordinates and time stamps, and a pupil signal containing pupil diameter size changes and corresponding time stamps;

[0035] The eye movement vector and the pupil signal are time-synchronized, and the two signals are resampled to obtain segmented signals;

[0036] The segmented signals are filtered by a Butterworth band-pass filter to obtain preliminary filtered signals;

[0037] The preliminary filtered signal is normalized to obtain normalized signals of the eye movement signal and the pupil response signal respectively;

[0038] Threshold method is used to detect and remove the blink artifacts from the normalized signal, the upper limit of the threshold band is set to 95% of the signal amplitude, and the lower limit of the threshold band is set to 5% of the signal amplitude, to obtain the de-artifact signal;

[0039] The missing data points in the de-artifact signal are filled by using the cubic spline interpolation method to obtain the preprocessed eye movement signal and pupil response signal;

[0040] The eye muscle adjustment amplitude is identified from the preprocessed eye movement signal to obtain the eye muscle adjustment amplitude, which is used to reflect the eye muscle tension and the eye congestion;

[0041] The pupil constriction range is analyzed from the preprocessed pupil response signal to obtain the pupil constriction range;

[0042] The iris change amplitude and the lens thickness change amplitude are measured;

[0043] The ciliary muscle torque change range is obtained, and the eye muscle adjustment amplitude, the pupil constriction range, the iris change amplitude, the lens thickness change amplitude and the ciliary muscle torque change range are combined into the eye movement feedback index set.

[0044] Specifically, the eye movement tracking device uses an infrared light source to irradiate the eyeball, a high-speed camera captures the corneal reflection light point change at a frequency of 250 frames per second, records the horizontal direction coordinate x-axis value and the vertical direction coordinate y-axis value of the eyeball, each coordinate point is attached with a time stamp accurate to milliseconds to form an eye movement vector. At the same time, the device monitors the pupil edge contour change at a frequency of 100 frames per second, calculates the pupil diameter size and records the corresponding time stamp to form a pupil signal. Since the eye movement signal and the pupil signal have different collection frequencies, linear interpolation algorithm is used for time synchronization processing, the data value at the unknown time point is calculated according to the data value at the known time point, and the two signals are resampled to a unified sampling rate of 200 Hz to ensure that each time point has corresponding eye movement coordinates and pupil diameter data. Then a 2-second length sliding window is set, the window overlap rate is 50% to segment the long time sequence signal, and each window contains 400 data points as an independent unit to obtain a segmented signal.

[0045] Butterworth band-pass filter is a kind of digital filter with maximum flat amplitude-frequency characteristic in passband, the upper boundary frequency of 4-order filter is set to 45 Hz and the lower boundary frequency is set to 0.5 Hz, the transfer function of the filter preserves the effective signal components in the range of 0.5 Hz to 45 Hz by calculating the amplitude and phase response of each frequency component of the input signal, and filters out high-frequency electronic noise and low-frequency baseline drift to obtain a preliminary filtered signal. The maximum and minimum value normalization method is used for eye movement signal normalization processing, the maximum and minimum values on each coordinate axis are calculated, the original coordinate value is subtracted from the minimum value and then divided by the difference between the maximum value and the minimum value, so that the horizontal coordinate and the vertical coordinate are mapped to the standard interval of 0 to 1. The Z-score standardization method is used for pupil signal normalization, the mean and standard deviation of the pupil diameter sequence are calculated, each pupil diameter value is subtracted from the mean and then divided by the standard deviation, and the normalized signal is obtained by converting to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0046] Threshold method detects blinking artifacts by determining the boundary of abnormal values through statistical analysis, calculating the 95% of the normalized signal amplitude as the upper limit of the threshold band and the 5% as the lower limit of the threshold band, and determining blinking artifacts when the signal amplitude exceeds this range and marking it as a missing value. The cubic spline interpolation method is a data completion method that maintains the smoothness of the curve, takes 5 valid data points before and after each missing area as reference points, constructs a cubic polynomial function so that the function value at the reference point is equal to the actual data value, and ensures that the first and second derivatives of the interpolation function are continuous at the connection point. The coefficients of the cubic polynomial are obtained by solving the linear equation system, the interpolation result of the missing time point is calculated to complete the data completion and obtain the de-artifact signal.

[0047] Eye muscle adjustment amplitude recognition is achieved by analyzing the curvature change and angular velocity characteristics of the eye movement trajectory, calculating the instantaneous speed and acceleration of the eye movement trajectory in the horizontal and vertical directions, and quantifying the adjustment activity intensity of the eye muscles in different visual tasks based on the physiological mechanism of eye muscle contraction and relaxation. The eye muscle adjustment amplitude value directly reflects the tension of the ciliary muscle and extraocular muscle, and when the value is high, it indicates that the eye muscles need greater contraction strength to maintain visual adjustment, and the degree of blood congestion in the eye will also increase accordingly. This physiological state is particularly evident in patients with pseudo myopia. Pupil contraction range analysis is achieved by measuring the maximum change in pupil diameter under standard light stimulation, recording the maximum diameter of the pupil before light change and the minimum diameter after light stimulation, and the difference between the two is the pupil contraction range. This indicator reflects the coordination function of the pupil sphincter and pupil dilator and the response sensitivity of the autonomic nervous system to light stimulation.

[0048] The iris change range is measured by high-resolution image analysis technology to measure the micro changes of iris texture in different adjustment states. The iris, as an important structure of the anterior segment of the eye, is closely related to the accommodation of the lens. When the eye is adjusted for near distance, the iris will change in shape and displacement accordingly. The lens thickness change range is obtained by ultrasonic measurement technology or optical coherence tomography technology, which measures the thickness difference of the lens in the far and near distance states. The change of the lens thickness directly reflects the contraction degree of the ciliary muscle and the strength of the accommodation ability. The ciliary muscle torque change range needs to set multiple distance test points. When the near distance test point of 30 cm is set, the ciliary muscle needs to contract to produce a 3.00D adjustment force. When the distance test points of 3 meters, 5 meters, 10 meters and 100 meters are set, the ciliary muscle needs to expand to produce different adjustment forces of 1.00D, 1.50D, 2.00D and 3.00D respectively. By measuring the actual accommodation ability of the ciliary muscle at different distances, the total torque change range from 10 meters to 30 cm complete adjustment range is calculated.

[0049] In a specific embodiment, the ciliary muscle torque change range is obtained, comprising:

[0050] Setting multiple target distance test points, including a 30 cm near distance test point and 3 meters, 5 meters, 10 meters and 100 meters far distance test points;

[0051] Measuring the contraction degree of the ciliary muscle of the measured person at a 30 cm near distance, and recording the adjustment state of the ciliary muscle contraction of 3.00D;

[0052] Measuring the expansion degree of the ciliary muscle of the measured person at a 3 meters, 5 meters, 10 meters and 100 meters far distance, and recording the adjustment state of the ciliary muscle expansion of 1.00D, 1.50D, 2.00D and 3.00D respectively;

[0053] Calculating the total torque change range of the ciliary muscle from the 10 meters far distance to the 30 cm near distance range, which should reach a 5D torque change in a normal state;

[0054] According to the difference between the actual measured ciliary muscle torque change range and the theoretical normal 5D torque change, the ciliary muscle fatigue degree evaluation result is obtained.

[0055] Specifically, the setting of multiple target distance test points is based on the physiological characteristics of the ciliary accommodation mechanism. The ciliary muscle is a ring-shaped smooth muscle that controls the shape change of the lens, and adjusts the lens thickness by contraction and expansion to achieve clear imaging at different distances. The 30 cm near distance test point is set in front of the subject, and the visual acuity test target is placed 30 cm away from the eye. This distance corresponds to the normal reading and writing distance, and the ciliary muscle needs to produce the maximum contraction force to maintain clear vision at close range. The 3 m, 5 m, 10 m, and 100 m distance test points are set at the corresponding distances, and each test point is equipped with a standard visual acuity test chart. The size of the test target is adjusted in proportion to the distance to ensure consistent viewing angles at each distance. During the test, the subject's head is fixed on a chin rest, and the eyes are aligned horizontally with each test point to avoid head movement affecting the measurement results.

[0056] The ciliary muscle contraction degree measurement during the 30 cm near distance test uses an accommodation response analyzer or an ocular refractive power measurement device. When the subject gazes at the test target at a distance of 30 cm, the ciliary muscle ring fibers contract to thicken the lens and increase the refractive power. Under normal circumstances, a 3.00 D accommodation force is needed to form a clear image point at a distance of 30 cm. The ciliary muscle contraction degree is calculated by measuring the change in the curvature radius of the anterior and posterior surfaces of the lens. The anterior surface curvature radius decreases from 10 mm in the resting state to 8 mm, and the posterior surface curvature radius decreases from 6 mm to 5.5 mm. According to the lens refractive power calculation formula, the total accommodation force is 3.00 D. At the same time, infrared ciliary muscle imaging technology is used to directly observe the contraction state of the ciliary muscle fibers, measure the radial contraction amplitude and circumferential tension of the ciliary muscle ring structure, and quantify the working intensity of the ciliary muscle during near distance accommodation.

[0057] The ciliary muscle expansion degree measurement during the distance test is based on the ciliary muscle relaxation mechanism. When the gaze distance increases, the ciliary muscle ring fibers gradually relax, and the lens returns to a flat shape under its own elasticity to reduce the refractive power. At a distance of 3 meters, the ciliary muscle needs to expand to produce a 1.00 D accommodation force, and the ciliary muscle fibers relax by about 25% relative to the maximum contraction state, and the anterior surface curvature radius of the lens returns to 9 mm. At a distance of 5 meters, the ciliary muscle expands to produce a 1.50 D accommodation force, and the ciliary muscle relaxation degree reaches 50%, and the anterior surface curvature radius of the lens returns to 9.5 mm. At a distance of 10 meters, the ciliary muscle expands to produce a 2.00 D accommodation force, and the ciliary muscle relaxation degree reaches 75%, and the anterior surface curvature radius of the lens approaches the resting state of 10 mm. At a distance of 100 meters, the ciliary muscle expands to produce a 3.00 D accommodation force, and the ciliary muscle completely relaxes to the resting state, and the lens returns to the most flat shape, with the anterior surface curvature radius reaching a maximum of 10 mm. At this time, the eye is in a far-seeing state and does not require additional accommodation force.

[0058] The ciliary muscle total moment change range calculation adopts the moment superposition principle, and the whole process of the ciliary muscle from a 10-meter distance adjustment state to a 30cm distance adjustment state is regarded as a continuous moment change process. The ciliary muscle is in a 2.00D expansion state at a 10-meter distance, and in a 3.00D contraction state at a 30cm distance. The total moment change range between the two states is 3.00D of the contraction state plus 2.00D of the expansion state, equaling 5.00D. The calculation of the moment change range needs to consider that the contraction direction and the expansion direction of the ciliary muscle fiber are opposite. The contraction generates a positive moment to increase the refractive power of the lens, and the expansion generates a negative moment to reduce the refractive power of the lens. The total moment change range reflects the working capacity of the ciliary muscle in the whole adjustment range. The ciliary muscle of a normal teenager should have a complete moment change range of 5D, which can fully contract to achieve near distance clear vision and fully expand to achieve far distance easy viewing.

[0059] The ciliary muscle fatigue degree evaluation is realized by analyzing the deviation of the measured moment change range from the theoretical normal value. When the measured ciliary muscle moment change range is less than 5D, it indicates that the ciliary muscle adjustment capacity is limited and there is fatigue phenomenon. The fatigue degree quantization calculation adopts the difference ratio method, taking the theoretical normal 5D moment change as the reference value. The difference between the measured moment change range and the reference value is divided by the reference value to obtain the fatigue degree coefficient. The fatigue degree coefficient changes between 0 and 1. The larger the value, the more serious the ciliary muscle fatigue. When the coefficient exceeds 0.3, it indicates moderate fatigue, and when it exceeds 0.6, it indicates severe fatigue. The ciliary muscle fatigue state is closely related to the occurrence of pseudo myopia. Long-term fatigue leads to ciliary muscle spasm and cannot fully relax, so that it remains in a certain degree of contraction state even when looking far, causing the lens to over-refract and forming pseudo myopia phenomenon.

[0060] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0061] The eye muscle adjustment amplitude calculation calculates the change between adjacent sampling points to obtain an eye muscle adjustment speed sequence;

[0062] The eye muscle adjustment speed sequence calculation calculates statistical features including maximum value, minimum value, average value and standard deviation, and simultaneously performs frequency domain transformation to extract energy distribution features in the 0.1-10Hz frequency band to obtain an eye muscle feature set;

[0063] The pupil contraction range calculation calculates the pupil diameter change rate and timing characteristics, measures the ratio of the maximum diameter to the minimum diameter of the pupil before and after the light change, and divides it into the pupil contraction period and the pupil recovery period to calculate the time length and change rate of the two stages respectively to obtain a pupil feature set;

[0064] The wavelet transform is performed on the pupil feature set, a Daubechies wavelet basis function is selected to perform multi-scale decomposition on the pupil signal, and statistical characteristics of wavelet coefficients at different decomposition levels are extracted to obtain a pupil wavelet feature set;

[0065] The eye muscle feature set, the pupil wavelet feature set, the iris change amplitude feature, and the lens thickness change feature are spliced in order of feature dimension to form a feature vector.

[0066] A principal component analysis algorithm is applied to the feature vector, and Z-score standardization processing is performed to convert each dimension feature into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminate dimension and numerical range differences, and obtain a multi-modal fusion feature vector.

[0067] Specifically, the adjacent sampling point change amount of the eye muscle adjustment amplitude is calculated using the difference method for data processing. The eye muscle adjustment amplitude is time series data obtained by measuring the contraction and relaxation changes of the extraocular muscle through an eye movement tracking device. The adjustment amplitude values recorded at a sampling rate of 200 Hz form a continuous numerical sequence on the time axis. The difference method calculation process is to subtract the adjacent two adjustment amplitude values on the time sequence to obtain the change amount per unit time, and then divide by the time interval of 5 milliseconds to obtain the instantaneous adjustment speed. This operation is repeated until the entire time sequence is processed to form an eye muscle adjustment speed sequence. The adjustment speed sequence reflects the speed of the dynamic change of the eye muscle, and a positive value indicates that the muscle contraction is accelerated, a negative value indicates that the muscle relaxation is accelerated, and the numerical value directly corresponds to the severity of the muscle adjustment.

[0068] The statistical feature calculation of the eye muscle adjustment speed sequence includes the extraction of four basic statistics. The maximum value reflects the fastest speed of the eye muscle contraction, representing the peak performance of the adjustment system. The minimum value is usually negative, indicating the fastest speed of the eye muscle relaxation, reflecting the recovery ability of the adjustment system. The average value represents the overall adjustment activity level, and the standard deviation reflects the adjustment stability and consistency. The frequency domain transformation uses the fast Fourier transform algorithm to convert the time domain adjustment speed sequence to the frequency domain for analysis. The Fourier transform decomposes the complex time domain signal into sine and cosine components of different frequencies. The frequency spectrum distribution is obtained by calculating the amplitude of each frequency component. The 0.1-10Hz frequency band is subdivided into three sub-bands. 0.1-1Hz corresponds to slow adjustment, reflecting the basic adjustment ability, 1-5Hz corresponds to medium speed adjustment, reflecting the adjustment response in normal visual tasks, and 5-10Hz corresponds to fast adjustment, reflecting the sensitivity of the adjustment system. The proportion of the energy of each sub-band to the total energy is calculated to form the eye muscle feature set.

[0069] The diameter change rate of pupil constriction range is calculated based on the quantitative analysis of the dynamic response process of the pupil, and the pupil diameter change rate is defined as the average change amount of the pupil diameter per unit time, which is obtained by dividing the maximum change amount of the pupil diameter by the corresponding time interval. The measurement of the ratio of the maximum diameter to the minimum diameter of the pupil before and after the light change eliminates the influence of the difference in the basic size of the pupil between individuals. The larger the ratio, the stronger the pupil constriction ability. The time sequence characteristic analysis divides the complete response process of the pupil to the light stimulus into two different physiological stages. The pupil constriction period starts from the light stimulus to the minimum diameter of the pupil. In this stage, the pupil sphincter muscle contracts rapidly, and the time length is usually between 200-500 milliseconds, and the change rate is fast. The pupil recovery period starts from the minimum diameter of the pupil to the recovery to the near baseline level. In this stage, the pupil dilator muscle gradually contracts to slowly dilate the pupil, and the time length is usually between 800-1500 milliseconds, and the change rate is relatively slow. The time length and change rate data of the two stages constitute the pupil feature set.

[0070] The wavelet transform of the pupil feature set adopts the Daubechies wavelet basis function for multi-scale decomposition. The Daubechies wavelet is a kind of wavelet basis function with tight support and orthogonality, which is suitable for analyzing the transient changes and local features in the pupil signal. Unlike Fourier transform, wavelet transform has good time-frequency localization characteristics, which can provide information in both time and frequency dimensions. Multi-scale decomposition decomposes the pupil signal into wavelet coefficients of different frequency levels, usually 5 layers of decomposition to obtain detail coefficients D1-D5 and approximation coefficients A5. Each layer of decomposition corresponds to a different frequency range and time resolution. D1 layer corresponds to the highest frequency detail information, D5 layer corresponds to the lowest frequency detail information, and A5 layer contains the low-frequency approximation information of the signal. Statistical property extraction includes energy calculation of each layer of wavelet coefficient obtained by coefficient square sum, entropy calculation reflecting information amount distribution, mean and variance calculation reflecting coefficient distribution characteristics, and extreme value ratio calculation obtained by the ratio of maximum absolute value to root mean square. These statistical properties constitute the pupil wavelet feature set.

[0071] The feature vector splicing merges the feature data from different sources in the order of feature dimensions. The eye muscle feature set contains 7-8 feature dimensions, the pupil wavelet feature set contains 10-12 feature dimensions, the iris change amplitude feature and the lens thickness change feature each occupy one dimension, and after splicing, a total of 20-22-dimensional original feature vectors are formed. The principal component analysis algorithm is a dimension reduction technique that maps the original feature space to a low-dimensional principal component space through linear transformation. The principal component is a linear combination of the original features, and the first few principal components contain the main variation information of the data according to the size of the variance contribution. The algorithm first calculates the covariance matrix of the feature vector, then solves the eigenvalues and eigenvectors of the covariance matrix, and the size of the eigenvalue represents the importance of the corresponding principal component. The first few principal components with a cumulative contribution rate of more than 95% are selected as the features after dimension reduction. The Z-score standardization process converts each feature dimension into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation process is to subtract the mean value of the feature dimension and divide by the standard deviation. Standardization eliminates the dimensional differences and numerical range differences between different features, so that each feature has equal weight in subsequent analysis. Finally, a multi-modal fusion feature vector is obtained.

[0072] In a specific embodiment, the step S103 process can specifically include the following steps:

[0073] The structure of the cascade forest classifier is set to three layers, each layer contains 4 forests, of which 2 are random forests and 2 are completely random forests. The number of decision trees in each forest is set to 100, and the initial classification structure is obtained;

[0074] The feature selection strategy is performed on the random forest. A plurality of features are randomly selected from the input features of the multi-modal fusion feature vector as candidate features. The Gini coefficient of each candidate feature is calculated. The feature with the optimal Gini coefficient is selected as the split point to obtain the node splitting scheme of the random forest;

[0075] The feature selection strategy is performed on the completely random forest. A single feature is directly randomly selected from the input features of the multi-modal fusion feature vector as a split feature to obtain the node splitting scheme of the completely random forest;

[0076] The multi-modal fusion feature vector is input into the first cascade forest, and the class probability feature is generated by cross-validation. The training data is divided into 5 parts. Each time, 4 parts are used to train the forest, and the remaining 1 part is used to predict the class probability to obtain the output feature of the first layer;

[0077] The output feature of the first layer is spliced with the original multi-modal fusion feature vector and input into the second cascade forest. The cross-validation prediction process is repeated to obtain the output feature of the second layer;

[0078] The output features of the second layer are spliced with the original multi-modal fusion feature vector and input to a third layer of cascaded forests. Through feature transmission and enhancement between layers, a weighted average is performed on the forest output of the final layer to obtain sample classification results containing normal vision, false myopia and true myopia.

[0079] Specifically, the three-layer structure design of the cascaded forest classifier is based on the idea of hierarchical feature learning in deep learning. Each layer contains 4 forest configurations to ensure model complexity and avoid overfitting risk. The first, second and third layers are responsible for primary feature learning, intermediate feature fusion and high-level decision output, respectively. The 2 random forests and 2 completely random forests in each layer form a complementary learning mechanism. Random forests achieve higher accuracy through feature selection, and completely random forests enhance model generalization ability through high randomization. The setting of 100 decision trees in each forest is based on the balance between experience and computational efficiency, and a total of 1200 decision trees form a powerful ensemble learning ability. After the initial classification structure is established, the forests in each layer are independently initialized parameters but connected between layers through data flow.

[0080] The feature selection strategy of random forests follows the core idea of the classic random forest algorithm. At each node of each decision tree, about 4 features are randomly selected from the 15-dimensional multi-modal fusion feature vector as candidate features. The number of selections is usually the square root of the total number of features to ensure both sufficient selection space and randomness. The Gini coefficient is an important indicator of measuring the purity of a dataset, and the formula is 1 minus the sum of the squares of the proportions of each class. The smaller the Gini coefficient, the purer the dataset and the better the classification effect. For the false myopia screening task, the Gini coefficient calculation of each candidate feature at the current node needs to count the distribution of normal vision, false myopia and true myopia in the left and right child nodes. By comparing the Gini coefficients of different features, the feature that minimizes the weighted average Gini coefficient after splitting is selected as the optimal splitting feature, and the threshold point that minimizes the Gini coefficient is determined as the node splitting scheme.

[0081] The feature selection strategy of completely random forests adopts an extreme randomization method, directly selecting 1 feature from the 15-dimensional multi-modal fusion feature vector as the splitting feature without Gini coefficient calculation and comparison. The splitting threshold is usually the median or a random threshold of the feature value. Although the accuracy of a single tree in completely random forests is low, the ensemble of a large number of random trees can effectively reduce the risk of overfitting and enhance the robustness of the model to noisy data. In the false myopia screening task, individual differences and measurement noise often exist in the data, and the addition of completely random forests can improve the stability of the overall classifier.

[0082] The first layer of cascaded forests uses a 5-fold cross-validation strategy. The training dataset is randomly divided into 5 equal subsets, each of which maintains the same proportion of samples in the normal vision, pseudo-myopia, and true myopia categories. The training process is repeated 5 times, with 4 subsets used as training data for the 4 forests in the first layer and the remaining subset used as validation data for predicting class probabilities. Each forest outputs probability values for 3 categories for each sample in the validation data, and the 4 forests generate a total of 12 probability features. After 5 cycles, each training sample has a set of probability predictions from forests that did not participate in its training, avoiding overfitting and obtaining reliable probability features. These 12 probability features form the output features of the first layer.

[0083] The input data for the second layer of cascaded forests is formed by feature concatenation. The 12 probability features from the first layer are concatenated with the original 15-dimensional multi-modal fusion feature vector in order to form a 27-dimensional enhanced feature vector. The second layer of forests can utilize both the original features and the high-level probability features learned by the first layer to learn more complex feature combinations and interactions. The training process of the second layer repeats the 5-fold cross-validation strategy of the first layer, using the 27-dimensional enhanced feature vector as input. Each forest performs the corresponding feature selection strategy based on its type. Random forests randomly select about 5 candidate features from the 27 features to calculate the Gini coefficient, and completely random forests directly select 1 feature for splitting. After training, the second layer also outputs 12 probability features.

[0084] The input of the third layer of cascaded forests is obtained by feature concatenation again. The 12 probability features from the second layer are concatenated with the original 15-dimensional multi-modal fusion feature vector to form a 27-dimensional enhanced feature vector. Note that the third layer does not use the output features of the first layer to avoid information redundancy. The training and prediction processes of the third layer of forests are the same as those of the previous two layers, but as the final decision layer, it needs to output the final classification result. The feature transfer between layers is reflected in the fact that each layer can access the original feature information and obtain the abstract features of the previous layer. Feature enhancement is reflected in the gradual improvement of feature representation ability as the level deepens. The weighted average output by the final layer of forests uses simple average or a weighting strategy based on validation performance to fuse the class probabilities of the 4 forests and output the final probability distribution of the normal vision, pseudo-myopia, and true myopia categories.

[0085] In a specific embodiment, the process of executing the step of concatenating the output features of the first layer with the original multi-modal fusion feature vector and inputting them into the second layer of cascaded forests, and repeating the cross-validation prediction process to obtain the output features of the second layer can specifically include the following steps:

[0086] The class probability calculation is performed on the four forests of the first cascade forest, and each forest generates a probability feature equal to the number of classes. The 12 probability feature values are combined to obtain a first layer probability feature vector;

[0087] The first layer probability feature vector is directly spliced with the original multi-modal fusion feature vector according to the feature dimension to obtain an enhanced feature vector;

[0088] The feature importance of the enhanced feature vector is sorted, and the average impurity reduction method is used to calculate the contribution of each feature to the classification result to obtain a feature weight distribution;

[0089] According to the feature weight distribution, the features in the enhanced feature vector are weighted, the features with high contribution are highlighted, and the features with low contribution are reduced, to obtain a weighted enhanced feature vector;

[0090] The weighted enhanced feature vector is input into the two random forests and two completely random forests in the second cascade forest, and the training and prediction are performed according to the same cross-validation method as the first layer to obtain the class probability output of the second layer;

[0091] The class probability output of the second layer is normalized, the probability value is normalized to the interval [0, 1] using the softmax function, and the sum of the class probabilities is 1, to obtain the output feature of the second layer.

[0092] Specifically, the class probability calculation of the four forests of the first cascade forest is based on the voting mechanism of ensemble learning, each forest contains 100 decision trees, and each decision tree classifies and predicts the input multi-modal fusion feature vector to output a judgment result in one of the three categories of normal vision, pseudomyopia and true myopia. The class probability of the forest is calculated by counting the voting results of the internal decision trees. Assuming that 65 of the 100 decision trees in a certain forest predict pseudomyopia, 25 predict normal vision, and 10 predict true myopia, the forest outputs a pseudomyopia probability of 0.65, a normal vision probability of 0.25, and a true myopia probability of 0.10. The first layer contains two random forests and two completely random forests, each of which independently outputs three class probabilities, totaling 12 probability feature values. These probability values reflect the classification tendency of different forests from different angles for the same sample. The probability feature merging process arranges the outputs of the four forests according to the forest number and class order to form a first layer probability feature vector containing 12 elements. Each element in the vector is between 0 and 1, and the sum of the three probability values of each forest is equal to 1.

[0093] The first layer probability feature vector is directly spliced with the original multi-modal fusion feature vector in the order of feature dimension. The original multi-modal fusion feature vector contains 15 dimensions corresponding to different physiological indicators such as ocular muscle features, pupil features, iris features, and lens features. The first layer probability feature vector contains 12 dimensions corresponding to the probability outputs of 4 forests respectively. The splicing operation combines the 15-dimensional original feature vector and the 12-dimensional probability feature vector in the column direction to form a 27-dimensional enhanced feature vector. The first 15 elements of the spliced vector maintain the original feature values, and the last 12 elements correspond to the probability outputs of the first layer forest. The enhanced feature vector contains both the bottom physiological feature information and the high-level abstract features learned by the first layer.

[0094] The feature importance ranking calculates the contribution of each feature to the classification result using the average impurity reduction method, which is based on the information gain evaluation in the decision tree splitting process. Impurity is usually measured by Gini index or information entropy. The Gini index calculation formula is 1 minus the sum of the squares of the proportions of each class. When a feature is used as the splitting condition for a decision tree node, the change in impurity before and after splitting is the contribution of the feature to this node. The average impurity reduction calculation needs to traverse all nodes of all decision trees in the first layer cascade forest, and calculate the impurity reduction of each feature as a splitting condition in each node. Then, the importance score of the feature is obtained by averaging the contribution of the same feature in all nodes. Each dimension in the 27-dimensional enhanced feature vector calculates its corresponding importance score. The higher the score, the stronger the feature's ability to distinguish between normal vision, pseudomyopia, and true myopia. The size order of the importance score constitutes the feature weight distribution.

[0095] The feature weighting process adjusts the weights of each dimension in the enhanced feature vector according to the feature weight distribution. The purpose of weighting is to amplify the influence of important features while suppressing the interference of unimportant features. The weighting process uses normalized weight coefficient multiplication. First, the importance scores of the 27 features are normalized so that the sum of all weight coefficients is equal to 1. Then, the original value of each feature is multiplied by the corresponding normalized weight coefficient. Assuming that a feature related to ciliary muscle regulation has a high importance score, the normalized weight coefficient is 0.08. Then, the original value of the feature is multiplied by 0.08 to get the weighted value. The weight coefficient of a feature with low importance is only 0.02, and its original value is multiplied by 0.02, which significantly reduces its influence in the vector. The weighted enhanced feature vector maintains the 27-dimensional dimension structure, but the values of each dimension are adjusted according to their contribution to the classification of pseudomyopia.

[0096] The training and prediction process of the second layer of cascaded forests receives the weighted enhanced feature vector as input data. The second layer also contains 2 random forests and 2 completely random forests, each containing 100 decision trees. The cross-validation method is exactly the same as the first layer, and the training data is divided into 5 subsets. Each time, 4 subsets are used to train the second layer forest, and the remaining 1 subset is used to predict the category probability. The random forest randomly selects about 5 features from the 27-dimensional weighted enhanced features as candidate features, calculates the Gini coefficient of each candidate feature to select the optimal split feature, and the completely random forest directly selects 1 feature for node splitting. In the training process of the second layer forest, features with higher weights in the weighted enhanced feature vector are more likely to be selected as candidate features by the random forest, and are more likely to become the optimal split feature in the Gini coefficient calculation. Features with lower weights are less likely to be selected and their splitting effect is relatively poor. After 5 rounds of cross-validation training, the 4 forests of the second layer each output 3 category probabilities, and a total of 12 probability output values are generated to form the original category probability output of the second layer.

[0097] The softmax function normalization process converts the category probability output of the second layer into a standard probability distribution. The softmax function is a mathematical function that converts an arbitrary real number vector into a probability distribution. For the 3 original scores of each forest output in the second layer, the softmax function first calculates the exponential value of each score, and then divides each exponential value by the sum of all exponential values to obtain the normalized probability. Assuming that the three original scores of the normal vision, false myopia and true myopia output by a forest are 1.2, 2.1 and 0.8 respectively, the corresponding exponential values are 3.32, 8.17 and 2.23, and the sum of the exponential values is 13.72. After softmax normalization, the probabilities of the three categories are 0.24, 0.60 and 0.16 respectively, and the sum of the probabilities is equal to 1 and all probability values are within the interval of 0 to 1. The 4 forests of the second layer all perform the same softmax normalization process, and finally obtain the normalized second layer output features.

[0098] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0099] Conducting contribution degree evaluation on the features in the sample classification result, calculating the influence degree of each feature on the classification accuracy, and obtaining a feature importance score;

[0100] According to the feature importance score, the features are arranged in descending order to establish a feature importance ranking list and obtain a feature priority sequence;

[0101] Based on the feature priority sequence, recursive feature elimination is performed to remove the least important features from the feature set one by one, and record the classification effect after each removal to obtain a feature number and classification effect relationship curve;

[0102] An inflection point position is identified from the feature quantity and classification effect relationship curve, the minimum feature quantity required under the premise of maintaining the classification effect is determined, and an initial feature set is obtained;

[0103] Correlation analysis is performed on the initial feature set, the correlation coefficient between features is calculated, and highly correlated feature pairs are identified. From each pair of highly correlated features, the one with higher importance is retained, and a low-redundancy feature set is obtained.

[0104] The low-redundancy feature set is combined as a key indicator for screening pseudo-myopia, forming a feature standard for pseudo-myopia screening, and obtaining a target feature combination.

[0105] Specifically, the contribution degree evaluation of features in the sample classification result is based on the information gain analysis of decision tree node splitting in the cascade forest classifier. Each feature will produce a change in impurity when used as a splitting condition in the decision tree construction process. The impurity change directly reflects the contribution degree of the feature to the classification accuracy. Contribution degree calculation requires traversing all 1200 internal nodes of the cascade forest, and calculating the Gini impurity reduction of each feature as a splitting condition in each node. The Gini impurity is calculated by subtracting the sum of the squares of the proportions of each category from 1. When a specific feature is used to split a node, the difference between the Gini value before splitting and the weighted average Gini value of the child nodes after splitting is the contribution degree of the feature at this node. The importance score of a feature is obtained by averaging the contribution degrees of the same feature at all nodes of all decision trees. The higher the importance score, the stronger the discriminant ability of the feature in distinguishing between normal vision, pseudo-myopia and true myopia. The score calculation process takes into account the frequency of feature use in the entire forest and the classification effect of each use.

[0106] The feature importance score is arranged in descending order to establish a feature importance ranking list. The sorting process arranges all features in descending order of their importance scores to form a feature priority sequence. The features at the front of the sorting list have a greater contribution to pseudo-myopia classification and should be retained first in the subsequent feature selection. The features at the back have relatively small contribution and should be removed first in the feature reduction process. The feature priority sequence provides a clear operation sequence for the recursive feature elimination algorithm, ensuring that the most important feature dimensions are retained first in the feature reduction process, and providing a quantitative basis for feature quantity optimization.

[0107] Recursive feature elimination (RFE) is a backward feature selection method that starts from the full feature set and removes the least important features one by one. Each removal operation re-trains the cascade forest classifier and evaluates the classification performance. The algorithm execution process first trains the classifier using all features and records the baseline performance indicator, then removes the least important features from the end of the feature priority sequence, re-trains the classifier with the remaining features and evaluates the performance change. After each removal, the current number of features and the corresponding classification accuracy or F1 score are recorded, forming data points of feature quantity and classification effect, and repeating this process until only one feature is left. The classification effect evaluation uses the cross-validation method to ensure the reliability of the results, and the same validation set is used for each performance evaluation to avoid data bias. The feature quantity and classification effect relationship curve is formed by connecting each data point, and the curve shape reflects the trend of feature reduction on classification performance, and the change of curve slope shows the distribution characteristics of feature importance.

[0108] The inflection point position is identified by analyzing the slope change of the feature quantity and classification effect relationship curve. The inflection point is the position where the slope of the curve changes significantly, indicating that further reducing the number of features will cause a significant decline in classification performance. The inflection point recognition algorithm calculates the slope change rate of each segment of the curve, and when the slope change rate exceeds the preset threshold, it is determined as the inflection point position. The number of features corresponding to the inflection point is the minimum number of features required to maintain the classification effect. After determining the inflection point, the highest importance corresponding number of features from the feature priority sequence is selected to form the preliminary feature set, which significantly reduces the feature dimension and computational complexity while maintaining high classification performance.

[0109] The correlation analysis of the preliminary feature set uses the Pearson correlation coefficient to calculate the linear correlation between features. The Pearson correlation coefficient has a value range of -1 to 1, and the closer the absolute value is to 1, the stronger the linear correlation between the two features. Correlation calculation requires calculating the correlation coefficient for each pair of features in the preliminary feature set, forming a correlation coefficient matrix. The diagonal elements of the matrix are 1, indicating that the feature is completely correlated with itself, and the non-diagonal elements reflect the correlation between different features. The identification of highly correlated feature pairs is achieved by setting a correlation coefficient threshold. When the absolute value of the correlation coefficient of two features exceeds 0.8, it is considered to have a high correlation, indicating that the two features contain a large amount of redundant information. From each pair of highly correlated features, the feature with higher importance is retained and the feature with lower importance is removed. The importance comparison is based on the previously calculated feature importance scores, and features with higher importance scores are preferentially retained. The redundant feature removal process processes all highly correlated feature pairs one by one until the correlation coefficients between the remaining features are below the set threshold, forming a low-redundancy feature set.

[0110] The low-redundancy feature set is converted into a key indicator combination for pseudomyopia screening, and each feature in the set represents a physiological indicator that has a unique contribution to the judgment of pseudomyopia and a low correlation with other features. The key indicator combination includes core features such as the ciliary muscle torque change range, pupil contraction speed, and eye movement regulation stability, which reflect the state of eye regulation function from different angles and can effectively distinguish the physiological differences between pseudomyopia and true myopia. The formation of the feature standard converts the key indicator combination into an operational standard for clinical screening, providing clear judgment basis for pseudomyopia identification in practical applications. The target feature combination not only maintains the accuracy of classification but also has practicality and interpretability.

[0111] In a specific embodiment, the process of performing step S105 can specifically include the following steps:

[0112] Based on the target feature combination, calculate the pseudomyopia degree index of the measured person to obtain a pseudomyopia degree quantitative score;

[0113] Calculate the theoretical myopia degree according to the ciliary muscle torque change range, and measure the actual myopia degree of the measured person;

[0114] Compare and analyze the actual myopia degree and the theoretical myopia degree: when the actual myopia degree is in the range of -0.25D to -1.25D and the ciliary muscle torque change range is less than the normal 5D change, it is determined as pseudomyopia; when the actual myopia degree is in the range of -0.25D to -1.50D and the theoretical myopia degree is calculated as 0.00D and above, it is determined as pseudomyopia; when the actual myopia degree is less than the theoretical myopia degree, it is determined as ciliary muscle fatigue accompanied by rapid development trend of myopia;

[0115] According to the pseudomyopia degree quantitative score, the pseudomyopia is subdivided into three grades of mild, moderate and severe, the mild range is greater than or equal to 0 and less than 0.3, the moderate range is greater than or equal to 0.3 and less than 0.7, and the severe range is greater than or equal to 0.7 and less than or equal to 1.0, to obtain the pseudomyopia grading result;

[0116] Calculate the classification confidence for the pseudomyopia grading result, calculate the difference between the highest class probability and the second highest class probability, mark the samples with a confidence lower than 0.2 as to be further checked, and obtain the classification result with confidence marking;

[0117] Generate individualized eye use suggestions for the classification result with confidence marking, develop corresponding eye use time arrangement, environmental light adjustment and vision care plan according to the different degrees of pseudomyopia, and obtain the vision health guidance plan;

[0118] The pseudo-myopia grading result, confidence label, ciliary muscle fatigue state evaluation and vision health guidance scheme are integrated into a comprehensive screening report, including screening time, basic information, classification result, confidence level and intervention suggestion, to obtain a pseudo-myopia screening preliminary report;

[0119] The pseudo-myopia screening preliminary report is added with screening result interpretation instructions, including pseudo-myopia formation causes, development trend and prevention and control measures, to obtain a pseudo-myopia screening report.

[0120] Specifically, the calculation of the pseudo-myopia degree index is based on weighted fusion analysis of the target feature combination, which includes key physiological indicators such as ciliary muscle torque change range, pupil accommodation flexibility, eye movement fixation stability and visual fatigue degree. The index calculation uses a weighted summation method. Each feature is first normalized to map the original value to the 0 to 1 interval, and then multiplied by the corresponding weight coefficient for weighted summation. The ciliary muscle torque change range contributes the most in pseudo-myopia judgment, so the weight coefficient is set to 0.4, the pupil accommodation flexibility weight coefficient is 0.3, the eye movement fixation stability weight coefficient is 0.2, and the visual fatigue degree weight coefficient is 0.1. The sum of each weight coefficient is equal to 1 to ensure that the index value is within the range of 0 to 1. The higher the pseudo-myopia degree index value, the more serious the pseudo-myopia degree. An index of 0 indicates that the eye accommodation function is completely normal, and an index of 1 indicates the presence of severe pseudo-myopia symptoms.

[0121] The theoretical myopia degree calculation of the ciliary muscle torque change range is based on the physiology of eye accommodation. Normal teenagers should have a complete torque change range of 5D. The difference between the measured torque change range and the normal value directly reflects the degree of decline in eye accommodation ability. The theoretical myopia degree calculation formula is to subtract the measured torque change range from the normal 5D torque. The difference is the equivalent myopia degree caused by insufficient ciliary muscle accommodation ability. The actual myopia degree is measured by standard optometry procedures, including subjective optometry and objective optometry. Subjective optometry determines the lens degree corresponding to the best corrected visual acuity through visual acuity testing. Objective optometry uses a computer optometry instrument or a lensometer to measure the eye refraction state.

[0122] The comparison between the actual myopia degree and the theoretical myopia degree uses a segmented judgment logic. The first judgment condition is that the actual myopia degree is in the range of -0.25D to -1.25D and the ciliary muscle torque change range is less than the normal 5D change. This condition indicates that the subject has mild to moderate myopia performance, and the ciliary muscle accommodation ability is indeed insufficient. When both conditions are met, it is determined to be pseudo myopia. The second judgment condition is that the actual myopia degree is in the range of -0.25D to -1.50D and the theoretical myopia degree is calculated to be 0.00D or above. This condition indicates that the subject's myopia degree is relatively light, and according to the actual accommodation ability of the ciliary muscle, the eye should not have myopia or even be emmetropia. The difference between the actual and theoretical values indicates that the myopia is mainly caused by accommodation factors and is pseudo myopia. The third judgment condition is that the actual myopia degree is smaller than the theoretical myopia degree, i.e., the negative value is larger. This condition indicates that the subject's myopia degree exceeds the range that can be explained by the decrease in ciliary muscle accommodation ability, indicating that the ciliary muscle is fatigued and has a rapid development trend of myopia, which needs to be focused on and intervened.

[0123] The grading process of the pseudo myopia degree quantitative score converts continuous index values into discrete level labels. Mild pseudo myopia corresponds to an index range of greater than or equal to 0 and less than 0.3. In this range, the subject's eye accommodation function is slightly abnormal, and the symptoms are relatively mild, mainly manifested as mild visual fatigue and temporary visual decline. Moderate pseudo myopia corresponds to an index range of greater than or equal to 0.3 and less than 0.7. In this range, the subject's eye accommodation function is significantly abnormal, and the symptoms are more obvious, manifested as persistent visual fatigue and visual decline. Severe pseudo myopia corresponds to an index range of greater than or equal to 0.7 and less than or equal to 1.0. In this range, the subject's eye accommodation function is severely abnormal, and the symptoms are severe, with a high risk of developing into true myopia. The grading results provide a basis for subsequent intervention program development, and different levels correspond to different intensity of intervention measures.

[0124] The calculation of classification confidence is based on the class probability distribution output by the cascade forest classifier. The cascade forest outputs probability values for three classes: normal vision, pseudo myopia, and true myopia for each sample. The sum of the probability values is equal to 1. The confidence calculation uses the difference between the highest probability and the second highest probability. The larger the difference, the more confident the classifier is in the judgment result. The smaller the difference, the more uncertain the judgment due to the fuzzy classification boundary. The confidence threshold is set to 0.2. When the difference between the highest probability and the second highest probability is less than 0.2, the sample is marked for further examination, indicating that the current screening result is not reliable enough and needs to be confirmed by more professional ophthalmic examination. The classification result with confidence label contains two-dimensional information of classification category and confidence level, providing quality evaluation basis for subsequent processing.

[0125] The generation of personalized eye use recommendations is based on the pseudo-myopia degree classification and confidence evaluation results. For mild pseudo-myopia, it is recommended to rest for 10 minutes after using eyes for 30-40 minutes, maintain the ambient light at 300-500 lux, perform eye care exercises 2 times a day, and increase outdoor activity time to 1 hour per day. For moderate pseudo-myopia, it is recommended to shorten the continuous eye use time to 20-30 minutes, extend the rest time to 15 minutes, increase the ambient light to 500-800 lux, increase the frequency of eye care exercises to 3-4 times a day, and increase the outdoor activity time to 1.5 hours per day. For severe pseudo-myopia, it is recommended to strictly control the eye use time, rest for 15 minutes every 15-20 minutes, adjust the ambient light to 800-1000 lux, increase the eye care exercises to 4-5 times a day, and increase the outdoor activity time to no less than 2 hours per day. At the same time, it is recommended to use eye protection instruments to assist in adjustment. For samples with low confidence, it is additionally recommended to go to a regular eye hospital for professional examinations such as mydriasis refraction as soon as possible.

[0126] The integration of the comprehensive screening report combines the pseudo-myopia classification results, confidence level, ciliary muscle fatigue state evaluation, and vision health guidance scheme into a unified document format. The report structure includes a basic information section that records screening time, location, equipment model, and basic information of the subject, a classification result section that records the pseudo-myopia degree index classification level and confidence level, a ciliary muscle evaluation section that records the range of moment change, fatigue degree, and comparison of theoretical and actual myopia degrees, and an intervention recommendation section that records personalized eye use time arrangement, environmental adjustment scheme, and vision care measures. The preliminary screening report uses a structured format for easy reading and subsequent processing, with key data displayed in table form and trends and contrast relationships displayed visually in charts.

[0127] The addition of the screening result interpretation explanation includes three aspects of popular science content: the causes of pseudo-myopia formation, development trends, and prevention and control measures. The interpretation explanation of the causes of pseudo-myopia formation explains the mechanism of accommodation spasm caused by long-term tension of the ciliary muscle, analyzes the possible development path and outcome of different degrees of pseudo-myopia, and introduces specific methods of scientific eye use and practical recommendations for vision protection. The interpretation explanation uses simple and easy-to-understand language, avoids excessive professional terms, and explains the handling methods for different situations with specific cases, ultimately forming a pseudo-myopia screening report.

[0128] The above describes the AI-based pseudo-myopia rapid screening method in the embodiments of the present application, and the AI-based pseudo-myopia rapid screening system in the embodiments of the present application is described below. Please refer to Figure 2 An embodiment of the AI-based pseudo-myopia rapid screening system in the embodiments of the present application includes:

[0129] The processing module 201 is configured to collect the eye movement original signal sequence and the pupil reaction original signal sequence of the subject and process them to obtain the eye movement feedback index set.

[0130] The splicing module 202 is configured to perform multi-modal feature extraction on the eye movement feedback index set and splice the multi-modal feature vector.

[0131] The input module 203 is configured to construct a cascade forest classifier based on random forest and complete random forest two forest algorithms, input the multi-modal feature vector into the cascade forest classifier to calculate a category probability distribution, and obtain a sample classification result.

[0132] The screening module 204 is configured to perform sample feature screening on the sample classification result to obtain a target feature combination.

[0133] The classification module 205 is configured to perform pseudomyopia classification on the subject according to the target feature combination to obtain a pseudomyopia screening report.

[0134] The application further provides a computer readable storage medium, which can be a non-volatile computer readable storage medium or a volatile computer readable storage medium, and the computer readable storage medium stores instructions, and when the instructions are run on a computer, the computer executes the steps of the AI-based pseudomyopia rapid screening method.

[0135] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, system and unit can refer to the corresponding process in the foregoing method embodiments, and will not be described here.

[0136] The integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing an AI-based pseudomyopia rapid screening device (which can be a personal computer, a server, or a network device) to execute all or part of the steps of the method described in various embodiments of the application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0137] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An AI-based method for rapid screening of pseudo-myopia, characterized in that, The method comprises: collecting and processing eye movement original signal sequences and pupil response original signal sequences of a testee to obtain an eye movement feedback index set; performing multi-modal feature extraction and splicing on the eye movement feedback index set to obtain a multi-modal fusion feature vector; constructing a cascade forest classifier based on two forest algorithms of random forest and complete random forest, inputting the multi-modal fusion feature vector into the cascade forest classifier to calculate a category probability distribution, and obtaining a sample classification result; performing sample feature screening on the sample classification result to obtain a target feature combination; performing pseudomyopia classification on the testee according to the target feature combination to obtain a pseudomyopia screening report. 2.The AI-based rapid screening method for pseudo-myopia according to claim 1, characterized in that, The eye movement feedback index set comprises: eye muscle adjustment amplitude, pupil contraction range, iris change amplitude, lens thickness change amplitude, and ciliary muscle torque change range. The collecting and processing of the eye movement original signal sequences and the pupil response original signal sequences of the testee to obtain the eye movement feedback index set comprises: collecting the eye movement original signal sequences of the testee through an eye movement tracking device, and simultaneously collecting the pupil response original signal sequences of the testee to obtain eye movement vectors containing horizontal direction coordinates, vertical direction coordinates and time stamps, and pupil signals containing pupil diameter size changes and corresponding time stamps; performing time synchronization on the eye movement vectors and the pupil signals, and resampling the two kinds of signals to obtain segmented signals; filtering the segmented signals through a Butterworth band-pass filter to obtain preliminary filtered signals; performing normalization processing on the preliminary filtered signals to process the eye movement signals and the pupil response signals respectively to obtain normalized signals; detecting and removing blink artifacts from the normalized signals by threshold method, setting the upper limit value of the threshold band to 95% of the signal amplitude, and setting the lower limit value of the threshold band to 5% of the signal amplitude to obtain de-artifact signals; completing data of missing data points in the de-artifact signals by using a cubic spline interpolation method to obtain preprocessed eye movement signals and pupil response signals; performing eye muscle adjustment amplitude identification on the preprocessed eye movement signals to obtain the eye muscle adjustment amplitude, which is used to reflect the eye muscle tension degree and the eye congestion degree; performing pupil contraction range analysis on the preprocessed pupil response signals to obtain the pupil contraction range; measuring the iris change amplitude and the lens thickness change amplitude; obtaining the ciliary muscle torque change range, and combining the eye muscle adjustment amplitude, the pupil contraction range, the iris change amplitude, the lens thickness change amplitude, and the ciliary muscle torque change range into the eye movement feedback index set. 3.The AI-based rapid screening method for pseudo-myopia according to claim 2, characterized in that, The obtaining of the ciliary muscle torque change range comprises: setting a plurality of target distance test points, including a 30cm near distance test point and 3m, 5m, 10m, and 100m far distance test points; measuring the ciliary muscle contraction degree of the testee at a 30cm near distance, and recording the ciliary muscle contraction 3.00D adjustment state; Measuring the ciliary muscle expansion degree of the measured person at a distance of 3 meters, 5 meters, 10 meters, and 100 meters, and recording the accommodation state of the ciliary muscle expansion of 1.00D, 1.50D, 2.00D, and 3.00D respectively; Calculating the total moment range of the ciliary muscle in the range from 10 meters to 30 cm, which should reach 5D in normal state; According to the difference between the actual ciliary muscle moment range and the theoretical normal 5D moment range, the ciliary muscle fatigue degree evaluation result is obtained. 4.The AI-based rapid screening method for pseudo-myopia according to claim 2, characterized in that, The multi-modal feature extraction and splicing are performed on the eye movement feedback indicator set to obtain a multi-modal fusion feature vector, including: The change amount between adjacent sampling points is calculated for the eye muscle adjustment amplitude to obtain an eye muscle adjustment speed sequence; Statistical features including maximum value, minimum value, average value and standard deviation are calculated for the eye muscle adjustment speed sequence, and a frequency domain transformation is performed to extract the energy distribution features of the 0.1-10Hz frequency band to obtain an eye muscle feature set; The pupil contraction range is calculated to obtain a pupil feature set, including the pupil diameter change rate and timing characteristics, the ratio of the maximum diameter to the minimum diameter of the pupil before and after the light change, and the division into the pupil contraction period and the pupil recovery period, and the calculation of the time length and the change rate of the two stages respectively; Wavelet transformation is performed on the pupil feature set, a Daubechies wavelet basis function is selected for multi-scale decomposition of the pupil signal, and the statistical characteristics of the wavelet coefficients at different decomposition levels are extracted to obtain a pupil wavelet feature set; The eye muscle feature set, the pupil wavelet feature set, the iris change amplitude feature and the lens thickness change feature are spliced in order of feature dimension to form a feature vector; The principal component analysis algorithm is applied to the feature vector and Z-score standardization processing is performed to convert each dimension feature into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminate the dimension and numerical range difference, and obtain a multi-modal fusion feature vector. 5.The AI-based rapid screening method for pseudo-myopia according to claim 1, characterized in that, The cascade forest classifier is constructed based on the random forest and complete random forest two forest algorithms, the multi-modal fusion feature vector is input into the cascade forest classifier to calculate the category probability distribution, and the sample classification result is obtained, including: The structure of the cascade forest classifier is set to three layers, each layer contains 4 forests, 2 of which are random forests and 2 of which are complete random forests, the number of decision trees in each forest is set to 100, and the initial classification structure is obtained; The feature selection strategy is performed on the random forest, a plurality of features are randomly selected from the input features of the multi-modal fusion feature vector as candidate features, the Gini coefficient of each candidate feature is calculated, the feature with the optimal Gini coefficient is selected as the split point, and the node split scheme of the random forest is obtained; The feature selection strategy is performed on the complete random forest, a single feature is directly randomly selected from the input features of the multi-modal fusion feature vector as a split feature, and the node split scheme of the complete random forest is obtained; inputting the multi-modal fusion feature vector into a first cascade forest, generating class probability features through cross-validation, dividing the training data into 5 parts, training the forest using 4 parts each time, and predicting the class probability on the remaining 1 part to obtain the output features of the first layer; concatenating the output features of the first layer with the original multi-modal fusion feature vector, inputting into a second cascade forest, and repeating the cross-validation prediction process to obtain the output features of the second layer; concatenating the output features of the second layer with the original multi-modal fusion feature vector, inputting into a third cascade forest, performing weighted averaging on the forest output of the final layer through feature transmission and enhancement between layers, and obtaining the sample classification results of three categories including normal vision, false myopia and true myopia. 6.The AI-based method for rapid screening of pseudo-myopia according to claim 5, characterized in that, The process of concatenating the output features of the first layer with the original multi-modal fusion feature vector, inputting into a second cascade forest, and repeating the cross-validation prediction process to obtain the output features of the second layer includes: performing class probability calculation on the 4 forests of the first cascade forest respectively, generating probability features equal to the number of categories for each forest, merging 12 probability feature values to obtain a first layer probability feature vector; directly concatenating the first layer probability feature vector with the original multi-modal fusion feature vector according to the feature dimension to obtain an enhanced feature vector; performing feature importance sorting on the enhanced feature vector, calculating the contribution of each feature to the classification result using the average impurity reduction method to obtain a feature weight distribution; performing weighted processing on the features in the enhanced feature vector according to the feature weight distribution, highlighting features with high contribution and reducing the influence of features with low contribution, to obtain a weighted enhanced feature vector; inputting the weighted enhanced feature vector into 2 random forests and 2 completely random forests in the second cascade forest, training and predicting according to the same cross-validation method as the first layer to obtain the class probability output of the second layer; normalizing the class probability output of the second layer, using a softmax function to normalize the probability value to the [0,1] interval, and making the sum of the class probabilities equal to 1 to obtain the output features of the second layer. 7.The AI-based rapid screening method for pseudo-myopia according to claim 1, characterized in that, The process of performing sample feature screening on the sample classification result to obtain a target feature combination includes: evaluating the contribution of features in the sample classification result, calculating the influence of each feature on the classification accuracy, and obtaining a feature importance score; arranging the features in descending order according to the feature importance score to establish a feature importance sorting list and obtain a feature priority sequence; based on the feature priority sequence, performing recursive feature elimination to remove the least important features from the feature set one by one and record the classification effect after each removal to obtain a feature number and classification effect relationship curve; identifying the inflection point position from the feature number and classification effect relationship curve to determine the minimum number of features required while maintaining the classification effect, and obtaining a preliminary feature set; Performing correlation analysis on the initial feature set, calculating the correlation coefficient between features, identifying highly correlated feature pairs, retaining the one with higher importance from each pair, and obtaining a low-redundancy feature set; Combining the low-redundancy feature set as a key indicator for screening pseudo-myopia, forming a feature standard for pseudo-myopia screening, and obtaining a target feature combination. 8.The AI-based rapid screening method for pseudo-myopia according to claim 1, characterized in that, The pseudo-myopia classification of the subject according to the target feature combination includes: Calculating the pseudo-myopia degree index of the subject based on the target feature combination, and obtaining a pseudo-myopia degree quantitative score; Calculating the theoretical myopia degree according to the ciliary muscle torque change range, and measuring the actual myopia degree of the subject; Comparing the actual myopia degree with the theoretical myopia degree: when the actual myopia degree is in the range of-0.25D to-1.25D and the ciliary muscle torque change range is less than the normal 5D change, it is determined as pseudo-myopia; when the actual myopia degree is in the range of-0.25D to-1.50D and the theoretical myopia degree is calculated as 0.00D and above, it is determined as pseudo-myopia; when the actual myopia degree is less than the theoretical myopia degree, it is determined as ciliary muscle fatigue with rapid development trend of myopia; Subdividing the pseudo-myopia into three grades of mild, moderate and severe according to the pseudo-myopia degree quantitative score, the mild range is greater than or equal to 0 and less than 0.3, the moderate range is greater than or equal to 0.3 and less than 0.7, and the severe range is greater than or equal to 0.7 and less than or equal to 1.0, and obtaining a pseudo-myopia grading result; Calculating the classification confidence of the pseudo-myopia grading result, calculating the difference between the highest class probability and the second highest class probability, marking the samples with confidence less than 0.2 as to be further checked, and obtaining a classification result with confidence marking; Generating individualized eye use suggestions for the classification result with confidence marking, formulating corresponding eye use time arrangement, environmental light adjustment and vision care scheme according to the degree of pseudo-myopia, and obtaining a vision health guidance scheme; Integrating the pseudo-myopia grading result, confidence marking, ciliary muscle fatigue state evaluation and vision health guidance scheme into a comprehensive screening report, including screening time, basic information, classification result, confidence level and intervention suggestion, and obtaining a pseudo-myopia screening preliminary report; Adding screening result interpretation instructions to the pseudo-myopia screening preliminary report, including the causes, development trend and prevention measures of pseudo-myopia, and obtaining a pseudo-myopia screening report.

9. An AI-based rapid screening system for pseudo-myopia, characterized in that, The AI-based pseudo-myopia rapid screening system for implementing the AI-based pseudo-myopia rapid screening method of any one of claims 1-8 includes: A processing module for collecting and processing the eye movement original signal sequence and the pupil reaction original signal sequence of the subject to obtain an eye movement feedback indicator set; A splicing module for performing multi-modal feature extraction on the eye movement feedback indicator set and splicing to obtain a multi-modal fusion feature vector; The input module is configured to construct a cascade forest classifier based on two forest algorithms of random forest and complete random forest, input the multi-modal fusion feature vector into the cascade forest classifier to calculate a category probability distribution, and obtain a sample classification result. The screening module is configured to perform sample feature screening on the sample classification result to obtain a target feature combination. The classification module is configured to perform pseudo-myopia classification on the subject according to the target feature combination to obtain a pseudo-myopia screening report.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, causes the processor to perform the AI-based pseudo-myopia rapid screening method according to any one of claims 1 to 8.

Images