AI-based pseudomyopia rapid screening method and system, and medium

Abstract

The invention relates to the technical field of data processing, and discloses an AI-based pseudomyopia rapid screening method and system, and a medium. The method comprises the following steps: acquiring 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 pseudomyopia grading report. According to the invention, the accuracy and efficiency of pseudomyopia screening are improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an AI-based rapid screening method, system, and medium for pseudomyopia. Background Art

[0002] Myopia is a common vision problem, and its incidence rate is increasing year by year, especially among adolescents. According to the clinical classification of ophthalmology, myopia can be divided into two main types: pseudomyopia and true myopia. Pseudomyopia refers to temporary myopia caused by excessive adjustment of the lens due to spasm of the ciliary muscle caused by long-term close-range eye use. In theory, it can be improved or reversed through proper eye use and adjustment and rest; while true myopia is irreversible myopia caused by anatomical changes such as axial length growth and changes in corneal curvature. Traditional myopia screening methods mainly rely on eye chart examinations and optometry instruments, including mydriasis refraction, A-ultrasound examination of axial length and other technical means. These methods can evaluate the refractive state and structural parameters of the eye to a certain extent.

[0003] There are many myopia screening methods in the existing technology, but most of them have obvious shortcomings. Although the traditional eye chart test is simple, it is difficult to distinguish different types of myopia, and it relies on subjective feedback from the subjects, which limits its accuracy. Although mydriatic refraction can accurately distinguish between pseudomyopia and true myopia, it requires the use of ciliary muscle paralytics, which is complicated to operate and time-consuming, making it unsuitable for large-scale screening. Although A-ultrasound examination of axial length is objective and reliable, the equipment is expensive and the operation is professional, which is also not suitable for widespread application. In addition, some existing computer vision-based screening technologies often only focus on a single physiological indicator, such as fundus photo analysis or corneal topography detection, and lack a comprehensive evaluation of multiple visual function indicators, making it difficult to accurately identify the early manifestations of pseudomyopia. In particular, there is a lack of effective monitoring and analysis of the two key indicators of pseudomyopia: pupil accommodation function and eye movement characteristics.

[0004] Since the cause of pseudomyopia mainly involves abnormal adjustment function of the ciliary muscle, and the existing screening methods cannot effectively quantify the actual adjustment ability of the ciliary muscle at different distances, it is impossible to accurately assess the difference between the range of change of the ciliary muscle torque and the theoretical normal value. At the same time, traditional methods lack the fusion analysis of multimodal physiological signals such as eye tracking and pupil response, and cannot build a comprehensive evaluation system that can reflect the multi-dimensional physiological characteristics of pseudomyopia. Furthermore, the existing technology lacks an artificial intelligence-based cascade classifier to process complex multimodal feature data, and cannot achieve accurate grading of pseudomyopia and generation of personalized intervention recommendations. This technical limitation directly affects the accuracy of early identification of pseudomyopia and the screening efficiency, and it is difficult to meet the actual needs of large-scale youth vision health screening. Summary of the Invention

[0005] The present application provides an AI-based rapid screening method, system and medium for pseudomyopia, which solves the problem in the existing technology that it is impossible to effectively integrate multimodal physiological signals such as eye tracking, pupil reaction, ciliary muscle adjustment, etc. to accurately identify and grade pseudomyopia, thereby improving the accuracy and efficiency of pseudomyopia screening.

[0006] In the first aspect, the present application provides an AI-based rapid screening method for pseudomyopia, which includes: collecting and processing the subject's original eye movement signal sequence and pupil reaction original signal sequence to obtain an eye movement feedback index set; performing multimodal feature extraction and splicing on the eye movement feedback index set to obtain a multimodal fusion feature vector; constructing a cascade forest classifier based on two forest algorithms, random forest and completely random forest, and inputting the multimodal fusion feature vector into the cascade forest classifier to calculate the category probability distribution to obtain a sample classification result; performing sample feature screening on the sample classification result to obtain a target feature combination; classifying the subject for pseudomyopia according to the target feature combination to obtain a pseudomyopia screening report.

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

[0008] A processing module is used to collect and process the original eye movement signal sequence and pupil response signal sequence of the subject to obtain an eye movement feedback indicator set;

[0009] A splicing module, configured to extract and splice multimodal features of the eye movement feedback indicator set to obtain a multimodal fusion feature vector;

[0010] An input module is used to construct a cascade forest classifier based on two forest algorithms: random forest and completely random forest, and input the multimodal fusion feature vector into the cascade forest classifier to calculate the category probability distribution and obtain the sample classification result;

[0011] A screening module is used to screen the sample characteristics of the sample classification results to obtain a target feature combination;

[0012] The classification module is used to classify the pseudomyopia of the subject according to the target feature combination to obtain a pseudomyopia screening report.

[0013] In a third aspect, an AI-based rapid screening device for pseudomyopia is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the AI-based rapid screening device for pseudomyopia performs the above-mentioned AI-based rapid screening method for pseudomyopia.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned AI-based rapid screening method for pseudomyopia.

[0015] The technical solution provided in this application collects and professionally processes the subject's original eye movement signal sequences and pupil response signal sequences to obtain a set of eye movement feedback indicators, including the amplitude of eye muscle adjustment, the range of pupil contraction, the amplitude of iris change, the amplitude of lens thickness change, and the range of ciliary muscle torque change. This comprehensive collection of multi-dimensional physiological characteristics overcomes the limitations of traditional methods that rely solely on a single indicator, providing a rich physiological basis for the accurate identification of pseudomyopia. The technical solution of extracting multimodal features from the eye movement feedback indicator set and splicing them to obtain a multimodal fusion feature vector effectively integrates physiological information from different sources, such as eye muscle features, pupil features, iris features, and lens features. Principal component analysis and Z-score normalization eliminate dimensional differences and numerical range differences, ensuring a balanced weight distribution of various features in subsequent analysis, significantly improving the integrity and reliability of feature expression. The cascade forest classifier, constructed based on two forest algorithms—random forest and completely random forest—adopts a three-layer design. Each layer contains four forests, ensuring model complexity while avoiding the risk of overfitting. Through inter-layer feature transfer and enhancement mechanisms, it is able to learn more abstract feature representations layer by layer. Compared with traditional single classifiers, it has stronger ability to process nonlinear feature relationships and higher classification accuracy. Sample classification results are screened for sample features to obtain the technical characteristics of the target feature combination. Through a strategy combining recursive feature elimination and correlation analysis, the key features that contribute most to pseudomyopia classification are retained while eliminating redundant information between features. This significantly reduces computational complexity while ensuring classification performance, making the method more suitable for rapid screening needs in practical application scenarios.

[0016] The quantitative assessment algorithm for the range of ciliary muscle torque variation accurately reflects the true state of ocular accommodation. By setting multiple target distance test points and calculating the difference between theoretical and actual myopia, it provides an objective physiological criterion for distinguishing pseudomyopia from true myopia. The random forest in the cascade forest classifier optimizes feature selection through the Gini coefficient, while the completely random forest enhances model generalization through a high degree of randomization. The collaborative work of these two algorithms is particularly suitable for addressing practical challenges such as large individual variability, high feature noise, and sample imbalance in pseudomyopia screening. Compared with traditional machine learning methods, it offers significant advantages in the reliability and stability of medical diagnosis. The multimodal feature fusion algorithm is specifically designed to address the diversity and complexity of ocular physiological signals. It extracts the time-frequency characteristics of pupil signals through wavelet transform and captures the dynamic characteristics of eye muscle accommodation through frequency domain analysis. These targeted signal processing methods fully consider the manifestations of pseudomyopia in different physiological dimensions, enabling the algorithm to more accurately capture early signs and subtle changes of pseudomyopia, providing strong technical support for early clinical intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 This is a schematic diagram of an embodiment of the AI-based rapid screening method for pseudomyopia in an embodiment of the present application;

[0019] Figure 2 This is a schematic diagram of an embodiment of an AI-based rapid pseudomyopia screening system in an embodiment of the present application. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide an AI-based rapid screening method, system and medium for pseudomyopia. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may 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. Figure 1 In the embodiments of the present application, an embodiment of the AI-based rapid pseudomyopia screening method includes:

[0022] Step S101: collecting and processing the original eye movement signal sequence and pupil response signal sequence of the subject to obtain an eye movement feedback indicator set;

[0023] Step S102: extracting and concatenating multimodal features of the eye movement feedback indicator set to obtain a multimodal fusion feature vector;

[0024] Step S103: constructing a cascade forest classifier based on the random forest and completely random forest algorithms, inputting the multimodal fusion feature vector into the cascade forest classifier to calculate the category probability distribution and obtain a sample classification result;

[0025] Step S104: screening the sample classification results by sample features to obtain a target feature combination;

[0026] Step S105: classify the subject for pseudomyopia according to the target feature combination to obtain a pseudomyopia screening report.

[0027] It is understandable that the execution subject of this application can be an AI-based pseudomyopia rapid screening system, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0028] Specifically, the original eye movement signal sequence and pupil response signal sequence of the subject are collected and processed to obtain the eye movement feedback indicator set. The horizontal coordinate x-axis and vertical coordinate y-axis of the subject's eyeball are recorded at a frequency of 250 frames per second through the eye tracking device, and the change in pupil diameter is collected at a frequency of 100 frames per second. Due to the different acquisition frequencies of the two signals, the linear interpolation algorithm is used to resample the two signals to a unified sampling rate of 200Hz to ensure that the data points correspond one to one in time. A Butterworth bandpass filter is used to set the upper boundary frequency to 45Hz and the lower boundary frequency to 0.5Hz to filter out high-frequency electronic noise and low-frequency baseline drift, and then normalization is performed. The eye movement signal uses the maximum and minimum value normalization method to map the coordinates to the range of 0 to 1, and the pupil signal uses the Z-score normalization method to convert it into a distribution with a mean of 0 and a standard deviation of 1. Blink artifacts were detected using a thresholding method. Signal values ​​exceeding 95% or falling below 5% were marked as missing. Data were then incomplete using cubic spline interpolation to construct a cubic polynomial with smooth transitions between adjacent data points. The amplitude of ocular muscle accommodation was calculated by calculating the amplitude of extraocular muscle contraction and relaxation during eye movement, and the range of pupil constriction was determined by analyzing the maximum change in pupil diameter under varying standard lighting conditions. The range of ciliary muscle torque variation was measured using a 30 cm near-distance test point and 3 m, 5 m, 10 m, and 100 m distance test points. The accommodative state of the ciliary muscle contracting at 30 cm for near vision and expanding at 1.00 D, 1.50 D, 2.00 D, and 3.00 D for distant vision was recorded. The total ciliary muscle torque variation was calculated from 10 m to 30 cm. Under normal conditions, a torque variation of 5 D should be achieved. The difference between the measured and theoretical values ​​was used to assess ciliary muscle fatigue.

[0029] Multimodal feature extraction and concatenation were performed on the eye movement feedback indicator set to generate a multimodal fusion feature vector. The change in eye muscle adjustment amplitude was calculated using the difference method between adjacent sampling points. The difference between two adjacent adjustment amplitude values ​​in the time series was calculated and then divided by the 5-millisecond interval to obtain the instantaneous adjustment speed, forming an eye muscle adjustment speed sequence. Statistical features of this sequence, including maximum, minimum, mean, and standard deviation, were calculated. A fast Fourier transform was performed to convert the time domain signal into the frequency domain. Energy distribution features were extracted from three sub-bands: 0.1-1 Hz slow adjustment, 1-5 Hz medium adjustment, and 5-10 Hz fast adjustment. The percentage of energy in each sub-band relative to the total energy constituted the eye muscle feature set. Pupil constriction range analysis was performed to calculate the pupil diameter change rate, defined as the average change in pupil diameter per unit time. The ratio of maximum to minimum pupil diameter was measured before and after the illumination change to eliminate inter-individual differences in baseline pupil size. The process was divided into pupil constriction and recovery phases based on temporal characteristics. The duration and change rate of each phase were calculated to form the pupil feature set. The pupil signal was decomposed into five layers using the Daubechies wavelet basis function, yielding detail coefficients D1-D5 and approximation coefficient A5 corresponding to different frequency ranges. The energy entropy, mean variance, and extreme value ratio of the wavelet coefficients at each decomposition layer were calculated, with particular attention paid to the energy proportion of the 0.3-0.5 Hz frequency band, which is associated with visual accommodation, forming a pupil wavelet feature set. The eye muscle feature set, pupil wavelet feature set, iris amplitude variation features, and lens thickness variation features were concatenated in order of feature dimensions to form a 17-20 dimensional feature vector. Principal component analysis was applied to identify the main directions of variation in the data. Correlated variables were converted to linearly uncorrelated principal components. Principal components with cumulative contributions exceeding 95% were selected and reduced to 12-15 dimensions. Z-score normalization was then performed to transform each feature component into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0030] A cascade forest classifier was constructed based on two forest algorithms: random forest and completely random forest. A three-layer architecture was established, with each layer containing four forests: two random forests and two completely random forests. Each forest contained 100 decision trees, forming a powerful ensemble of 1,200 decision trees. When constructing each node in each decision tree, the random forest randomly selected approximately five features from the 25-dimensional multimodal fusion feature vector as candidate features and calculated the Gini coefficient for each candidate feature. The Gini coefficient is calculated by evaluating the class purity of the child nodes after the feature splits the sample. The Gini coefficient is calculated by subtracting the sum of the squares of the proportions of each class from 1. The feature with the smallest Gini coefficient is selected as the splitting feature and the optimal splitting point is determined. The completely random forest directly randomly selects one feature from the 25-dimensional feature vector as the splitting feature, and the median of this feature is selected as the splitting point to increase model diversity and prevent overfitting. The multimodal fusion feature vector is input into the first-level cascade forest. A five-fold cross-validation approach is used to divide the training data into five equal parts, maintaining a consistent ratio of sample types. Four of the data sets are used in turn to train the forests, which predict class probabilities on the remaining set. Each forest outputs probabilities for three categories: normal vision, pseudomyopia, and true myopia. The four forests generate a total of 12 probability features, forming the first-level output features. The 12 probability features output from the first layer are concatenated with the original 25-dimensional multimodal fusion feature vector to form a 37-dimensional enhanced feature vector. This is then input into the second-level cascade forest, where the cross-validation prediction process is repeated to obtain the second-level output features. This concatenation is then fed into the third-level cascade forest. Through inter-level feature transfer and enhancement, the final forest output is weighted averaged to obtain the three-category classification results.

[0031] The sample classification results were screened for sample features to obtain the target feature combination. The average impurity reduction method was used to calculate feature contribution. Each decision tree in the cascade forest classifier was traversed to count the nodes that used a specific feature as a splitting condition. The impurity change before and after the node split was calculated and measured using the Gini index. The impurity reduction of the same feature in all trees was averaged to obtain the feature importance score. All 25 features were sorted in descending order according to the feature importance score to establish a feature importance ranking list. A recursive feature elimination algorithm was performed, starting from the complete feature set and removing the least important feature in each iteration. The cascade forest classifier was retrained with the remaining features and performance was evaluated. The number of features and the corresponding classification performance were recorded, and a curve showing the relationship between the number of features and classification effect was plotted. The inflection point on the curve where the slope changed significantly was identified, indicating that further feature reduction would lead to a significant decline in performance. The number of features corresponding to the inflection point was determined, and the corresponding number of features with the highest importance were selected from the feature priority sequence to form the preliminary feature set. A correlation analysis is performed on the preliminary feature set to calculate the Pearson correlation coefficient between the features. When a feature pair with an absolute value of the correlation coefficient exceeding 0.8 is found, it is considered to contain a large amount of redundant information. From each pair of highly correlated features, the more important one is retained and the other is removed to obtain a low-redundancy feature set as the target feature combination.

[0032] The subjects were classified for pseudomyopia based on the target feature combination, and a pseudomyopia screening report was obtained. The pseudomyopia severity index was calculated based on the target feature combination by weighted fusion of pupil accommodation flexibility, eye movement stability, and visual fatigue to quantify the severity of pseudomyopia. Each feature was first normalized to a value between 0 and 1, then multiplied by the corresponding weight and summed to obtain the final index value. The theoretical myopia degree was calculated based on the range of ciliary muscle torque, and the expected visual state was assessed based on the actual accommodation ability of the ciliary muscle. The actual myopia degree of the subjects was measured for comparative analysis. When the actual myopia degree is within the range of -0.25D to -1.25D and the ciliary muscle torque variation range is less than the normal 5D variation, it is judged as pseudomyopia. When the actual myopia degree is within the range of -0.25D to -1.50D and the theoretical myopia degree is calculated to be 0.00D or above, it is also judged as pseudomyopia. When the actual myopia degree is smaller than the theoretical myopia degree, it is judged as ciliary muscle fatigue with a rapid development of myopia. According to the pseudomyopia degree index, pseudomyopia is subdivided into three levels: mild 0 to 0.3, moderate 0.3 to 0.7, and severe 0.7 to 1.0. The difference between the highest category probability and the second highest category probability output by the cascade forest classifier is calculated as the classification confidence. Samples with confidence levels below 0.2 are marked for further inspection. Personalized eye care recommendations are generated for varying degrees of pseudomyopia. Mild pseudomyopia is recommended with a 10-minute break after every 30-40 minutes of eye use, maintaining ambient light levels between 300 and 500 lux. Moderate pseudomyopia is recommended with shorter continuous eye use time to 20-30 minutes and increased breaks to 15 minutes. Severe pseudomyopia requires stricter control of eye use time, requiring a 15-minute break every 15-20 minutes. The report integrates the pseudomyopia grading result confidence marker, ciliary muscle fatigue status assessment, and vision health guidance plan into a comprehensive screening report. Interpretation of the screening results, including the causes, development trends, and prevention and control measures for pseudomyopia, is also included to form a pseudomyopia screening report.

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

[0034] The eye tracking device collects the subject's original eye movement signal sequence and the subject's pupil response signal sequence at the same time, and obtains an eye movement vector including horizontal coordinates, vertical coordinates and timestamps, and a pupil signal including pupil diameter changes and corresponding timestamps;

[0035] Time-synchronize the eye movement vector and pupil signal, resample the two signals to obtain segmented signals;

[0036] The segmented signal is filtered through a Butterworth bandpass filter to obtain a preliminary filtered signal;

[0037] Normalize the preliminary filtered signal, process the eye movement signal and pupil response signal separately, and obtain normalized signals;

[0038] The threshold method was used to detect and remove blink artifacts from the normalized signal. The upper threshold value was set at 95% of the signal amplitude, and the lower threshold value was set at 5% of the signal amplitude to obtain the artifact-free signal.

[0039] The missing data points in the artifact-free signal are supplemented by cubic spline interpolation to obtain the preprocessed eye movement signal and pupil response signal.

[0040] The eye muscle adjustment amplitude is identified on the pre-processed eye movement signal to obtain the eye muscle adjustment amplitude, which is used to reflect the degree of eye muscle tension and eye congestion;

[0041] Performing pupil contraction range analysis on the preprocessed pupil response signal to obtain the pupil contraction range;

[0042] Measure the amplitude of iris changes and lens thickness changes;

[0043] The range of ciliary muscle torque variation is obtained, and the eye muscle adjustment amplitude, pupil contraction range, iris variation amplitude, lens thickness variation amplitude, and ciliary muscle torque variation range are combined into an eye movement feedback indicator set.

[0044] Specifically, the eye tracking device illuminates the eyeball with an infrared light source. A high-speed camera captures changes in the corneal reflected light at a frequency of 250 frames per second, recording the horizontal x-axis value and the vertical y-axis value of the eyeball. Each coordinate point is accompanied by a millisecond-accurate timestamp to form the eye movement vector. Simultaneously, the device monitors changes in the pupil edge contour at a frequency of 100 frames per second, calculates the pupil diameter, and records the corresponding timestamp to form the pupil signal. Because the eye movement signal and the pupil signal are collected at different frequencies, a linear interpolation algorithm is used for time synchronization. The values ​​at unknown time points are calculated based on the data values ​​at known time points. Both signals are resampled to a unified 200Hz sampling rate to ensure that each time point has corresponding eye movement coordinates and pupil diameter data. The long time series signal is then segmented using a 2-second sliding window with a 50% window overlap. Each window contains 400 data points, which are processed as independent units to produce the segmented signal.

[0045] The Butterworth bandpass filter is a digital filter with maximally flat amplitude-frequency characteristics within its passband. The upper boundary frequency of the fourth-order filter is set at 45 Hz and the lower boundary frequency is set at 0.5 Hz. The filter transfer function calculates the amplitude and phase response of each frequency component of the input signal, retaining the effective signal components in the range of 0.5 Hz to 45 Hz while filtering out high-frequency electronic noise and low-frequency baseline drift to obtain a preliminary filtered signal. Eye movement signal normalization uses the maximum-minimum normalization method. 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 and minimum values, so that both horizontal and vertical coordinates are mapped to the standard range of 0 to 1. Pupil signal normalization uses the Z-score normalization method. The mean and standard deviation of the pupil diameter series are calculated. Each pupil diameter value is subtracted from the mean and divided by the standard deviation to convert it to a standard normal distribution with a mean of 0 and a standard deviation of 1 to obtain the normalized signal.

[0046] The threshold method for detecting blink artifacts uses statistical analysis to determine outlier boundaries. The 95th percentile of the normalized signal amplitude is calculated as the upper threshold band value, and the 5th percentile is calculated as the lower threshold band value. Signal amplitudes outside this range are identified as blink artifacts and marked as missing. Cubic spline interpolation is a data completion method that maintains curve smoothness. Five valid data points are taken before and after each missing region as reference points. A cubic polynomial function is constructed so that the function value at the reference point is equal to the actual data value. The first and second derivatives of the interpolation function at the connection points are also continuous. The coefficients of the cubic polynomial are obtained by solving a system of linear equations. The interpolation results for the missing time points are then calculated to complete the data completion and obtain the artifact-free signal.

[0047] Identifying the amplitude of eye muscle accommodation is achieved by analyzing the curvature changes and angular velocity characteristics of the eye movement trajectory. The instantaneous velocity and acceleration of the eye movement trajectory in the horizontal and vertical directions are calculated. Combined with the physiological mechanisms of extraocular muscle contraction and relaxation, the intensity of eye muscle accommodation activity during different visual tasks is quantified. The amplitude of eye muscle accommodation directly reflects the tension of the ciliary and extraocular muscles. A higher value indicates that the eye muscles require greater contraction force to maintain visual accommodation, and the degree of ocular vascular congestion will also increase accordingly. This physiological state is particularly evident in patients with pseudomyopia. Pupil constriction range analysis is achieved by measuring the maximum change in pupil diameter under standard light stimulation. The maximum pupil diameter before the light change and the minimum pupil diameter after the light stimulation are recorded. The difference between the two is the pupil constriction range. This indicator reflects the coordinated function of the pupillary sphincter and pupillary dilator muscles and the sensitivity of the autonomic nervous system to light stimulation.

[0048] The iris amplitude variation measures microscopic changes in iris texture under different accommodation states using high-resolution image analysis. As a crucial structure in the anterior segment of the eye, the iris' morphological changes are closely linked to lens accommodation. When the eye accommodates near objects, the iris deforms and moves accordingly. The lens thickness variation is measured using ultrasonic measurement or optical coherence tomography. This difference in lens thickness between near and far vision is directly reflected in the degree of ciliary muscle contraction and accommodative capacity. The range of ciliary muscle torque variation requires multiple test distances. At a close distance of 30 cm, the ciliary muscle contracts to produce 3.00 D of accommodation force. At distances of 3 meters, 5 meters, 10 meters, and 100 meters, the ciliary muscle expands to produce varying degrees of accommodation force, such as 1.00 D, 1.50 D, 2.00 D, or 3.00 D. By measuring the actual accommodative capacity of the ciliary muscle at different distances, the total torque variation over the full accommodative range from 10 meters to 30 cm is calculated.

[0049] In a specific embodiment, obtaining the ciliary muscle torque variation range includes:

[0050] Set up multiple target distance test points, including a 30cm close-range test point and 3m, 5m, 10m, and 100m long-range test points;

[0051] Measure the degree of ciliary muscle contraction of the subject at a close distance of 30 cm, and record the accommodation state of ciliary muscle contraction of 3.00D;

[0052] Measure the degree of ciliary muscle dilation of the subject at distances of 3 meters, 5 meters, 10 meters, and 100 meters, and record the accommodation states of ciliary muscle dilation of 1.00D, 1.50D, 2.00D, and 3.00D respectively;

[0053] Calculate the total torque variation of the ciliary muscle from a distance of 10 meters to a close distance of 30 cm. Under normal conditions, the torque variation should reach 5D.

[0054] The ciliary muscle fatigue degree assessment result was obtained based on the difference between the measured ciliary muscle torque variation range and the theoretical normal 5D torque variation.

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

[0056] The degree of ciliary muscle contraction during the 30cm close-up test is measured using an accommodation response analyzer or ocular diopter measurement device. When the subject focuses on a test target at a distance of 30cm, the circular fibers of the ciliary muscle contract, thickening the lens and increasing its refractive power. Under normal circumstances, 3.00 diopters of accommodation force are required to form a sharp image at a distance of 30cm. The degree of ciliary muscle contraction is calculated by measuring the change in the radius of curvature of the anterior and posterior surfaces of the lens. The anterior radius of curvature decreases from a resting state of 10mm to 8mm, and the posterior radius of curvature decreases from 6mm to 5.5mm. Based on the lens refractive power calculation formula, the total accommodation force is 3.00 diopters. Infrared ciliary muscle imaging is also used to directly observe the contraction of the ciliary muscle fibers, measure the radial contraction amplitude and circumferential tension of the ciliary muscle annular structure, and quantify the intensity of the ciliary muscle during close-up accommodation.

[0057] The measurement of ciliary muscle dilation during distance testing is based on the mechanism of ciliary muscle relaxation. As the viewing distance increases, the circular fibers of the ciliary muscle gradually relax, and the lens, under its own elasticity, returns to a flattened shape, reducing refractive power. At a distance of 3 meters, the ciliary muscle needs to dilate to produce 1.00 D of accommodation. The ciliary muscle fibers relax approximately 25% relative to their maximum contraction, and the radius of curvature of the lens's anterior surface returns to 9 mm. At a distance of 5 meters, the ciliary muscle dilates to produce 1.50 D of accommodation, reaching 50% relaxation, and the radius of curvature of the lens's anterior surface returns to 9.5 mm. At a distance of 10 meters, the ciliary muscle dilates to produce 2.00 D of accommodation, reaching 75% relaxation, and the radius of curvature of the lens's anterior surface approaches its resting state of 10 mm. At a distance of 100 meters, the ciliary muscle dilates to produce 3.00 D of accommodation. The ciliary muscle fully relaxes to its resting state, the lens returns to its flattest shape, and the radius of curvature reaches its maximum value of 10 mm. At this point, the eye no longer needs additional accommodation for distance viewing.

[0058] The total ciliary muscle torque range is calculated using the principle of torque superposition, treating the entire process from 10-meter accommodation to 30-cm accommodation as a continuous torque change. At 10 meters, the ciliary muscle is in a 2.00D dilated state, while at 30 cm, it is in a 3.00D contracted state. The total torque change between the two states is 3.00D in contraction plus 2.00D in dilation, which equals 5.00D. The calculation of the torque range takes into account that ciliary muscle fibers contract in opposite directions to dilation. Contraction generates a positive torque that increases the lens's refractive power, while dilation generates a negative torque that decreases it. The total torque range reflects the ciliary muscle's ability to function throughout the entire accommodative range. A normal adolescent's ciliary muscle should have a full 5D torque range, allowing it to contract sufficiently for clear near vision and dilate sufficiently for comfortable distance viewing.

[0059] The degree of ciliary muscle fatigue is assessed by analyzing the deviation between the measured torque variation range and the theoretical normal value. When the measured ciliary muscle torque variation range is less than 5D, it indicates that the ciliary muscle's ability to adjust is limited and fatigue is present. The degree of fatigue is quantified using the difference ratio method. The theoretical normal 5D torque variation is used as the baseline value, and the difference between the measured torque variation range and the baseline value is divided by the baseline value to obtain the fatigue coefficient. The fatigue coefficient varies between 0 and 1. A larger value indicates more severe ciliary muscle fatigue. When the coefficient exceeds 0.3, it indicates moderate fatigue, and when it exceeds 0.6, it indicates severe fatigue. The state of ciliary muscle fatigue is closely related to the occurrence of pseudomyopia. Long-term fatigue causes ciliary muscle spasm and inability to fully relax. Even when looking at distance, it still maintains a certain degree of contraction, causing excessive refraction of the lens and forming pseudomyopia.

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

[0061] The change between adjacent sampling points of the eye muscle adjustment amplitude is calculated to obtain the eye muscle adjustment speed sequence;

[0062] Calculate statistical features of the eye muscle adjustment speed sequence, including maximum, minimum, mean, and standard deviation. Perform frequency domain transformation and extract energy distribution features in the 0.1-10 Hz frequency band to obtain an eye muscle feature set.

[0063] The pupil diameter change rate and timing characteristics of the pupil contraction range are calculated. The ratio of the maximum to minimum pupil diameter before and after the illumination change is measured. The pupil is divided into the pupil contraction period and the pupil recovery period. The time length and change rate of the two stages are calculated respectively to obtain the pupil feature set.

[0064] Perform wavelet transform on pupil feature set, select Daubechies wavelet basis function to perform multi-scale decomposition on pupil signal, extract statistical characteristics of wavelet coefficients at different decomposition levels, and obtain pupil wavelet feature set;

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

[0066] The principal component analysis algorithm is applied to the feature vector and Z-score normalization is performed to convert each dimension feature into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the dimension and value range differences, and obtaining a multimodal fusion feature vector.

[0067] Specifically, the difference method is used to process the data for calculating the change in the eye muscle adjustment amplitude between adjacent sampling points. The eye muscle adjustment amplitude is a time series data obtained by measuring the contraction and relaxation changes of the extraocular muscles using an eye tracking device. The adjustment amplitude values ​​recorded at a sampling rate of 200Hz form a continuous numerical sequence on the time axis. The difference method calculation process is to subtract the two adjacent adjustment amplitude values ​​in the time series to obtain the change per unit time, and then divide it by the time interval of 5 milliseconds to obtain the instantaneous adjustment speed. This operation is repeated until the entire time series is processed to form the eye muscle adjustment speed sequence. The adjustment speed sequence reflects the speed of the dynamic changes of the eye muscles. Positive values ​​indicate faster muscle contraction, and negative values ​​indicate faster muscle relaxation. The numerical value directly corresponds to the intensity of muscle adjustment.

[0068] The statistical feature calculation of the eye muscle adjustment speed sequence includes the extraction of four basic statistical quantities. The maximum value reflects the fastest speed of eye muscle contraction and represents the peak performance of the adjustment system. The minimum value, which is usually a negative value, indicates the fastest speed of eye muscle relaxation and reflects the recovery ability of the adjustment system. The average value represents the overall adjustment activity level. 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 wave components of different frequencies. The 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; 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 calculation of the rate of change in pupil diameter within the range of pupil constriction is based on a quantitative analysis of the pupil's dynamic response. The pupil diameter change rate is defined as the average change in pupil diameter per unit time and is calculated by dividing the maximum change in pupil diameter by the corresponding time interval. Measuring the ratio of maximum to minimum pupil diameter before and after the illumination change eliminates the influence of individual differences in baseline pupil size. A larger ratio indicates greater pupil constriction. Temporal characteristic analysis divides the complete pupillary response to light stimulation into two distinct physiological phases. The pupil constriction phase, which lasts from the onset of light stimulation until the pupil reaches its minimum diameter, involves rapid contraction of the pupillary sphincter, typically lasting between 200 and 500 milliseconds and exhibiting a relatively rapid rate of change. The pupil recovery phase, which lasts from the onset of pupil diameter until it returns to near baseline, involves gradual contraction of the pupillary dilator muscles, causing the pupil to dilate slowly. This phase typically lasts between 800 and 1500 milliseconds and exhibits a relatively slow rate of change. The duration and rate of change data for these two phases constitute the pupil feature set.

[0070] The wavelet transform of the pupil feature set uses Daubechies wavelet basis functions for multi-scale decomposition. Daubechies wavelets are a class of wavelet basis functions with compact support and orthogonality, making them suitable for analyzing transient changes and local features in pupil signals. Unlike the Fourier transform, the wavelet transform exhibits excellent time-frequency localization, providing information about the signal in both time and frequency dimensions. Multi-scale decomposition decomposes the pupil signal into wavelet coefficients at different frequency levels, typically using five layers of decomposition to produce detail coefficients D1-D5 and approximation coefficients A5. Each layer corresponds to a different frequency range and time resolution. Layer D1 corresponds to the highest-frequency detail information, layer D5 to the lowest-frequency detail information, and layer A5 to contain low-frequency approximation information of the signal. Statistical feature extraction involves calculating the energy of each wavelet coefficient layer (by summing the squares of the coefficients), entropy to reflect information distribution, mean and variance to reflect coefficient distribution characteristics, and extreme value ratio (by the ratio of the maximum absolute value to the root mean square). These statistical features constitute the pupil wavelet feature set.

[0071] Feature vector concatenation combines 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, and the iris amplitude and lens thickness feature sets each have one dimension. After concatenation, a total of 20-22 dimensions are formed in the original feature vector. Principal component analysis (PCA) is a dimensionality reduction technique that uses a linear transformation to map the original feature space to a lower-dimensional principal component space. Principal components are linear combinations of the original features and are sorted by their variance contribution. The top principal components contain the main information about the data variation. The algorithm first calculates the covariance matrix of the eigenvectors and then solves for the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​indicate the importance of the corresponding principal component. The top principal components with a cumulative contribution exceeding 95% are selected as the features after dimensionality reduction. Z-score normalization 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 of the dimension from the feature value and divide it by the standard deviation. Normalization eliminates the dimensional differences and numerical range differences between different features, so that each feature has equal weight in subsequent analysis, and finally obtains a multimodal fusion feature vector.

[0072] In a specific embodiment, the process of executing step S103 may 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 to obtain the initial classification structure;

[0074] Executing a feature selection strategy on the random forest, randomly selecting multiple features from the input features of the multimodal fusion feature vector as candidate features, calculating the Gini coefficient of each candidate feature, selecting the feature with the best Gini coefficient as the splitting point, and obtaining a node splitting scheme of the random forest;

[0075] Performing a feature selection strategy on the completely random forest, directly and randomly selecting a single feature from the input features of the multimodal fusion feature vector as a split feature, and obtaining a node splitting scheme of the completely random forest;

[0076] Input the multimodal fusion feature vector into the first layer of cascade forest, generate class probability features through cross-validation, divide the training data into 5 parts, use 4 parts to train the forest each time, and predict the class probability on the remaining part to obtain the output features of the first layer;

[0077] The output features of the first layer are concatenated with the original multimodal fusion feature vector and input into the second layer cascade forest. The cross-validation prediction process is repeated to obtain the output features of the second layer.

[0078] The output features of the second layer are spliced ​​with the original multimodal fusion feature vector and input into the third-layer cascade forest. Through feature transfer and enhancement between layers, the forest output of the final layer is weighted averaged to obtain sample classification results including normal vision, pseudomyopia and true myopia.

[0079] Specifically, the three-layer structure of the cascade forest classifier is based on the idea of ​​hierarchical feature learning in deep learning. The configuration of four forests per layer ensures model complexity while avoiding the risk of overfitting. The first, second, and third layers are responsible for primary feature learning, intermediate feature fusion, and high-level decision output, respectively. The two random forests and two fully random forests within each layer form a complementary learning mechanism. Random forests achieve high accuracy through feature selection, while fully random forests enhance model generalization through a high degree of randomization. The configuration of 100 decision trees per forest is based on a balance between experience and computational efficiency. The total of 1,200 decision trees forms a powerful ensemble learning capability. After the initial classification structure is established, each forest layer has its parameters initialized independently but is connected between layers via data flow.

[0080] The feature selection strategy of random forest follows the core idea of ​​the classic random forest algorithm. When constructing each node of each decision tree, approximately 4 features are randomly selected from the 15-dimensional multimodal fusion feature vector as candidate features. The number of features selected is usually the square root of the total number of features to ensure sufficient selection space and maintain randomness. The Gini coefficient is an important indicator for measuring the purity of a dataset. The calculation formula is 1 minus the sum of the squares of the sample proportions of each category. The smaller the Gini coefficient, the purer the dataset and the better the classification effect. For the pseudomyopia screening task, the calculation of the Gini coefficient of each candidate feature at the current node requires statistical analysis of the distribution of the three categories of normal vision, pseudomyopia, and true myopia in the left and right child nodes. By comparing the Gini coefficients of different features, the feature with the smallest weighted average Gini coefficient after splitting is selected as the optimal split feature. At the same time, the split threshold point that minimizes the Gini coefficient is determined as the node splitting scheme.

[0081] The feature selection strategy of the fully randomized forest uses an extreme randomization method, directly selecting a single feature from the 15-dimensional multimodal fusion feature vector as the splitting feature without calculating or comparing the Gini coefficient. The splitting threshold is typically the median of the feature value or a random threshold. Although the high degree of randomization in fully randomized forests results in lower accuracy for individual trees, the ensemble of a large number of randomized trees can effectively reduce the risk of overfitting and enhance the model's robustness to noisy data. In medical diagnostic tasks such as pseudomyopia screening, where data often exhibit individual variability and measurement noise, the addition of fully randomized forests can improve the stability of the overall classifier.

[0082] Cross-validation training for the first-level cascade forests uses a 5-fold cross-validation strategy, randomly dividing the training dataset into five equal subsets. Each subset maintains a consistent ratio of samples from the three categories of normal vision, pseudomyopia, and true myopia. The training process repeats five times, each time selecting four subsets as training data for the four forests in the first layer, and the remaining subset as validation data for predicting class probabilities. Each forest outputs probability values ​​for three categories for each sample in the validation data, generating a total of 12 probabilistic features across the four forests. After five cycles, each training sample receives a set of probability predictions from forests that did not participate in its training, avoiding overfitting while obtaining reliable probabilistic features. These 12 probabilistic features constitute the output features of the first layer.

[0083] The input data for the second-layer cascade forest is formed through feature concatenation. The 12 probabilistic features output by the first layer are concatenated with the original 15-dimensional multimodal fusion feature vector in dimensional order to form a 27-dimensional enhanced feature vector. The second-layer forest can simultaneously utilize the original features and the high-level probabilistic features learned in 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 a feature selection strategy based on its type. The random forest randomly selects approximately 5 candidate features from the 27 features to calculate the Gini coefficient, while the fully random forest directly randomly selects 1 feature for splitting. After training, the second layer also outputs 12 probabilistic features.

[0084] The input to the third-layer cascade forest is obtained through feature concatenation. The 12 probabilistic features output by the second layer are concatenated with the original 15-dimensional multimodal 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 process of the third-layer forest is the same as the first 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 while obtaining the abstract features of the previous layer. Feature enhancement is manifested as the feature representation capability gradually improves with the deepening of the layer. The weighted average of the output of the final layer forest adopts a simple average or a weighting strategy based on verification performance to fuse the category probabilities of the four forests and output the final probability distribution of the three categories of normal vision, pseudomyopia, and true myopia.

[0085] In a specific embodiment, the step of concatenating the output features of the first layer with the original multimodal fusion feature vector, inputting the concatenated features into the second layer of the cascade forest, and repeatedly performing the cross-validation prediction process to obtain the output features of the second layer may specifically include the following steps:

[0086] Calculate the class probabilities for each of the four forests in the first-level cascade forest. Each forest generates probability features equal to the number of classes. Combine the 12 probability feature values ​​to obtain the first-level probability feature vector.

[0087] The first layer of probability feature vectors are directly concatenated with the original multimodal fusion feature vectors according to the feature dimensions to obtain an enhanced feature vector;

[0088] Sorting the enhanced feature vectors by feature importance, calculating the contribution of each feature to the classification result using the average impurity reduction method, and obtaining a feature weight distribution;

[0089] performing weighted processing on the features in the enhanced feature vector according to the feature weight distribution, highlighting the features with high contribution and reducing the influence of the features with low contribution, to obtain a weighted enhanced feature vector;

[0090] Input the weighted enhanced feature vector into two random forests and two completely random forests in the second layer of cascade forests, and perform training and prediction in the same cross-validation manner as the first layer to obtain the class probability output of the second layer;

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

[0092] Specifically, the probability calculation of the four forest categories in the first-level 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 multimodal fusion feature vector and outputs a judgment result of one of the three categories of normal vision, pseudomyopia, and true myopia. The category probability of the forest is calculated by counting the voting results of its internal decision trees. Assuming that 65 of the 100 decision trees in a forest predict pseudomyopia, 25 predict normal vision, and 10 predict true myopia, the probability of pseudomyopia output by the forest is 0.65, the probability of normal vision is 0.25, and the probability of true myopia is 0.10. The first layer contains 2 random forests and 2 completely random forests. Each forest independently outputs 3 category probabilities, resulting in a total of 12 probability feature values. These probability values ​​reflect the classification tendencies of different forests for the same sample from different angles. The probabilistic feature merging process arranges the outputs of the four forests in order of forest number and category to form a first-level probabilistic 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 probabilistic feature vector is directly concatenated with the original multimodal fusion feature vector, combining data in the order of feature dimensions. The original multimodal fusion feature vector contains 15 dimensions corresponding to different physiological indicators such as eye muscle features, pupil features, iris features, and lens features. The first-layer probabilistic feature vector contains 12 dimensions corresponding to the three-category probabilities output by each of the four forests. The concatenation operation merges the 15-dimensional original feature vector and the 12-dimensional probabilistic feature vector column-wise to form a 27-dimensional enhanced feature vector. The first 15 elements of the concatenated vector retain the values ​​of the original features, while the last 12 elements correspond to the probabilistic output of the first-layer forest. The enhanced feature vector contains both the underlying physiological feature information and the high-level abstract features learned in the first layer.

[0094] Feature importance ranking uses the average impurity reduction method to calculate each feature's contribution to the classification result. This method is based on the information gain assessment during the decision tree splitting process. Impurity is typically measured using the Gini index or information entropy. The Gini index is calculated as 1 minus the sum of the squares of the sample proportions in each category. When a feature is used as a splitting condition for a decision tree node, the change in impurity before and after the split is the contribution of the feature to that node. Calculating the average impurity reduction requires traversing all nodes of all decision trees in the first-level cascade forest. The impurity reduction of each feature when used as a splitting condition at each node is calculated. The contribution of the same feature across all nodes is then averaged to obtain the feature's importance score. An importance score is calculated for each dimension in the 27-dimensional enhanced feature vector. A higher score indicates a greater ability of the feature to distinguish between normal vision, pseudomyopia, and true myopia. The order of importance scores constitutes the feature weight distribution.

[0095] Feature weighting adjusts the weights of each dimension in the enhanced feature vector based on 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, and 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 and a normalized weight coefficient of 0.08, the original value of the feature is multiplied by 0.08 to obtain the weighted value. The weight coefficient of a feature with lower importance is only 0.02. After its original value is multiplied by 0.02, its influence in the vector is significantly reduced. The weighted enhanced feature vector maintains a 27-dimensional structure, but the values ​​of each dimension are readjusted according to their contribution to the classification of pseudomyopia.

[0096] The training and prediction process for the second-layer cascade forest takes the weighted augmented feature vector as input. The second layer also consists of two random forests and two completely random forests, each containing 100 decision trees. The cross-validation method is identical to the first layer: the training data is divided into five subsets. Four subsets are used to train the second-layer forest at a time, and class probabilities are predicted on the remaining subset. The random forest randomly selects approximately five features from the 27-dimensional weighted augmented features as candidate features, calculates the Gini coefficient for each candidate feature, and selects the optimal splitting feature. The completely random forest directly randomly selects one feature for node splitting. During the training of the second-layer forest, features with higher weights in the weighted augmented feature vector are more likely to be selected as candidate features by the random forest and are more likely to become the optimal splitting feature in the Gini coefficient calculation. Even if selected, features with lower weights have relatively poor splitting effects. After five rounds of cross-validation training, each of the four forests in the second layer outputs three class probabilities, totaling 12 probability outputs, which constitute the raw class 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 any real number vector into a probability distribution. For the three category raw scores output by a forest 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. Suppose that the three raw scores of normal vision, pseudomyopia, 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. The sum of the probabilities is equal to 1, and all probability values ​​are in the range of 0 to 1. The four forests in the second layer all perform the same softmax normalization process, and finally obtain the standardized second-layer output features.

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

[0099] Evaluate the contribution of the features in the sample classification results, calculate the impact of each feature on the classification accuracy, and obtain the feature importance score;

[0100] Arrange the features in descending order according to the feature importance scores, establish a feature importance ranking list, and obtain a feature priority sequence;

[0101] Recursive feature elimination is performed based on the feature priority sequence to remove the least important features from the feature set one by one, and the classification effect after each removal is recorded to obtain a curve showing the relationship between the number of features and the classification effect;

[0102] Identify the inflection point position from the relationship curve between the number of features and the classification effect, determine the minimum number of features required to maintain the classification effect, and obtain a preliminary feature set;

[0103] Performing correlation analysis on the pre-selected feature set, calculating correlation coefficients between features, identifying highly correlated feature pairs, and retaining the more important feature from each pair of highly correlated features to obtain a low-redundancy feature set;

[0104] The low-redundancy feature set is used as a key indicator combination for screening pseudomyopia to form a feature standard for pseudomyopia screening and obtain a target feature combination.

[0105] Specifically, the contribution of features in sample classification results is assessed based on information gain analysis of splitting decision tree nodes in a cascade forest classifier. Each feature, when used as a splitting condition during decision tree construction, experiences a change in impurity, and this change in impurity directly reflects the feature's contribution to classification accuracy. Contribution calculation involves traversing each internal node of all 1,200 decision trees in the cascade forest and calculating the reduction in Gini impurity for each feature when used as a splitting condition at each node. Gini impurity is calculated by subtracting the sum of squared sample proportions from 1. When a node is split using a specific feature, the contribution of that feature at that node is calculated as the Gini value before the split minus the weighted average Gini value of the child nodes after the split. The importance score of the feature is calculated by averaging the contribution of the same feature across all nodes in all decision trees. A higher importance score indicates a stronger discriminative ability of the feature in distinguishing true from normal vision from pseudomyopia. The score calculation takes into account the frequency of use of the feature throughout the forest and the classification performance of each use.

[0106] A feature importance ranking list is established by arranging the feature importance scores in descending order. The ranking process arranges all features from high to low according to their importance scores, forming a feature priority sequence. Features that are ranked higher in the ranking list indicate that they contribute more to the classification of pseudomyopia and should be retained in subsequent feature selection. Features ranked lower in the ranking list contribute less and should be removed first during feature reduction. The feature priority sequence provides a clear order of operations for the recursive feature elimination algorithm, ensuring that the most important feature dimensions are retained first during feature reduction, while also providing a quantitative basis for optimizing the number of features.

[0107] The recursive feature elimination algorithm is a reverse feature selection method that starts with the complete feature set and gradually removes the least important features. Each time a feature is removed, the cascade forest classifier is retrained and the classification performance is evaluated. The algorithm first trains the classifier using all features and records baseline performance metrics. Then, based on the feature priority sequence, the least important features are removed from the end. The classifier is retrained using the remaining features and performance is evaluated. After each feature removal, the current number of features and the corresponding classification accuracy or F1 score are recorded to form a data point comparing the number of features and classification performance. This process is repeated until only one feature remains. Classification performance is evaluated using cross-validation to ensure the reliability of the results. The same validation set is used for each performance evaluation to avoid data bias. A curve is formed by connecting the data points to show the relationship between the number of features and classification performance. The shape of the curve reflects the trend in the impact of feature reduction on classification performance, and the change in the slope of the curve indicates the distribution of feature importance.

[0108] Inflection point location identification is achieved by analyzing the slope changes of the curve that relates the number of features to classification performance. The inflection point is the location on the curve where the slope changes significantly, indicating that further reducing the number of features at this point will lead to a significant decrease in classification performance. The inflection point identification algorithm calculates the rate of change of the slope of each segment of the curve. When the slope change rate exceeds a preset threshold, it is determined to be an inflection point. The number of features corresponding to the inflection point is the minimum number of features required to maintain the classification effect. After the inflection point is determined, the corresponding number of features with the highest importance are selected from the feature priority sequence to form a preliminary feature set. The preliminary feature set significantly reduces the feature dimension and computational complexity while maintaining high classification performance.

[0109] Correlation analysis of the preselected feature set uses the Pearson correlation coefficient to calculate the degree of linear correlation between features. The Pearson correlation coefficient ranges from -1 to +1, with absolute values ​​closer to 1 indicating a stronger linear correlation between the two features. Correlation calculation involves calculating the correlation coefficient for each pair of features in the preselected feature set, forming a correlation coefficient matrix. A diagonal element of 1 in this matrix indicates a perfect correlation between the features themselves, while off-diagonal elements reflect the degree of correlation between different features. Highly correlated feature pairs are identified by setting a correlation coefficient threshold. When the absolute value of the correlation coefficient between two features exceeds 0.8, a high correlation is considered, indicating that the two features contain a significant amount of duplicate information. From each highly correlated feature pair, the more important feature is retained, while the less important feature is removed. Importance comparison is based on the previously calculated feature importance scores, with features with higher importance scores being retained first. The redundant feature removal process processes all highly correlated feature pairs one by one until the correlation coefficients between the remaining features fall 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. Each feature in the set represents a physiological indicator that has a unique contribution to the judgment of pseudomyopia and has a low correlation with other features. The key indicator combination includes core features such as the range of change of ciliary muscle torque, pupil contraction speed, and eye movement accommodation stability. These features reflect the state of eye accommodation function from different angles and can effectively distinguish the physiological differences between pseudomyopia and true myopia. The formation of feature standards transforms the key indicator combination into an operational standard for clinical screening, providing a clear judgment basis for the identification of pseudomyopia in practical applications. The target feature combination maintains the accuracy of classification while being practical and interpretable.

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

[0112] Calculating the pseudomyopia degree index of the subject based on the target feature combination to obtain a quantitative score of the pseudomyopia degree;

[0113] The theoretical myopia degree is calculated according to the variation range of the ciliary muscle torque, and the actual myopia degree of the subject is measured;

[0114] Compare and analyze the actual myopia degree with the theoretical myopia degree: when the actual myopia degree is within 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 to be pseudomyopia; when the actual myopia degree is within the range of -0.25D to -1.50D and the theoretical myopia degree is calculated to be 0.00D or above, it is determined to be pseudomyopia; when the actual myopia degree is smaller than the theoretical myopia degree, it is determined to be ciliary muscle fatigue with a rapid development of myopia;

[0115] According to the quantitative scoring of the pseudomyopia degree, pseudomyopia is subdivided into three levels: mild, moderate and severe, where 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, thereby obtaining a pseudomyopia grading result;

[0116] Calculating the classification confidence for the pseudomyopia grading results, calculating the difference between the highest category probability and the second highest category probability, marking samples with a confidence level lower than 0.2 for further inspection, and obtaining a classification result with a confidence level mark;

[0117] Generate personalized eye care recommendations based on the confidence-marked classification results, formulate corresponding eye care time arrangements, environmental lighting adjustments, and vision care plans based on the degree of pseudomyopia, and obtain a vision health guidance plan;

[0118] Integrate the pseudomyopia grading results, confidence marks, ciliary muscle fatigue status assessment and vision health guidance plan into a comprehensive screening report, including screening time, basic information, classification results, confidence level and intervention recommendations, to obtain a preliminary pseudomyopia screening report;

[0119] An explanation of the screening results is added to the preliminary pseudomyopia screening report, including the causes, development trends and prevention and control measures of pseudomyopia, to obtain a pseudomyopia screening report.

[0120] Specifically, the calculation of the pseudomyopia degree index is based on the weighted fusion analysis of the target feature combination. The target feature combination includes key physiological indicators such as the range of change of ciliary muscle torque, pupil accommodation flexibility, eye movement and gaze stability, and visual fatigue. The index calculation adopts the weighted summation method. Each feature is first normalized to map the original value to the range of 0 to 1, and then multiplied by the corresponding weight coefficient for weighted summation. Since the range of change of ciliary muscle torque contributes the most to the judgment of pseudomyopia, the weight coefficient is set to 0.4, the weight coefficient of pupil accommodation flexibility is 0.3, the weight coefficient of eye movement and gaze stability is 0.2, and the weight coefficient of visual fatigue 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 value of the pseudomyopia degree index, the more severe the pseudomyopia. An index of 0 indicates that the eye accommodation function is completely normal, and an index of 1 indicates the presence of severe pseudomyopia symptoms.

[0121] The calculation of theoretical myopia based on the range of ciliary muscle torque variation is based on the physiological principles of eye accommodation. The normal ciliary muscle of adolescents should have a complete torque variation range of 5D. The difference between the measured torque variation range and the normal value directly reflects the degree of decline in the eye's accommodation ability. The formula for calculating theoretical myopia is the normal 5D torque minus the measured torque variation range. The difference is the equivalent myopia caused by insufficient accommodation ability of the ciliary muscle. The actual myopia degree is measured through standard optometry procedures, including subjective optometry and objective optometry. Subjective optometry determines the lens degree corresponding to the best corrected visual acuity through eye chart examination, and objective optometry uses a computer ophthalmometer or retinoscope to measure the refractive state of the eye.

[0122] The comparative analysis of actual myopia and theoretical myopia adopts segmented judgment logic. The first judgment condition is that the actual myopia is within the range of minus 0.25D to minus 1.25D and the range of change of ciliary muscle torque is less than the normal 5D change. This condition indicates that the subject has mild to moderate myopia and the adjustment ability of the ciliary muscle is indeed insufficient. When both conditions are met, it is judged as pseudomyopia. The second judgment condition is that the actual myopia is within the range of minus 0.25D to minus 1.50D and the theoretical myopia is calculated to be 0.00D or above. This condition indicates that the subject's myopia is relatively mild. According to the actual adjustment ability of the ciliary muscle, the eye should not have myopia or should even be emmetropia. The difference between the actual measurement and the theory suggests that myopia is mainly caused by adjustment factors and is pseudomyopia. The third judgment condition is that the actual myopia degree is smaller than the theoretical myopia degree, that is, the negative value is larger. This situation indicates that the degree of myopia of the subject exceeds the range that can be explained by the decline in the adjustment ability of the ciliary muscle, suggesting that ciliary muscle fatigue is accompanied by a rapid development trend of myopia, which requires special attention and intervention.

[0123] The grading process of quantitative scoring of pseudomyopia converts continuous index values ​​into discrete grade labels. The corresponding index range for mild pseudomyopia is greater than or equal to 0 and less than 0.3. Within this range, the subjects' eye accommodation function is slightly abnormal, and the symptoms are relatively mild, mainly manifested as mild visual fatigue and temporary vision loss. The corresponding index range for moderate pseudomyopia is greater than or equal to 0.3 and less than 0.7. Within this range, the subjects' eye accommodation function is obviously abnormal, and the symptoms are more obvious, manifested as persistent visual fatigue and vision loss. The corresponding index range for severe pseudomyopia is greater than or equal to 0.7 and less than or equal to 1.0. Within this range, the subjects' eye accommodation function is severely abnormal, the symptoms are severe, and there is a high risk of developing true myopia. The grading results provide a basis for the formulation of subsequent intervention plans, and different levels correspond to intervention measures of different intensities.

[0124] The calculation of classification confidence is based on the category probability distribution output by the cascade forest classifier. The cascade forest outputs probability values ​​for three categories of normal vision, pseudomyopia, and true myopia for each sample, and 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, and the smaller the difference, the more uncertain the judgment is due to fuzzy classification boundaries. 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 requires more professional ophthalmic examination confirmation. The classification results with confidence marks contain information in two dimensions: classification category and confidence level, providing a quality assessment basis for subsequent processing.

[0125] The generation of personalized eye use recommendations is based on the pseudomyopia degree classification and confidence assessment results. For mild pseudomyopia, it is recommended to rest for 10 minutes after every 30-40 minutes of eye use, maintain the ambient light at 300-500 lux, perform eye exercises twice a day, and increase outdoor activity time to 1 hour a day. For moderate pseudomyopia, 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 exercises to 3-4 times a day, and increase outdoor activity time to 1.5 hours a day. For severe pseudomyopia, it is recommended to strictly control the eye use time and rest for 15 minutes every 15-20 minutes, adjust the ambient light to 800-1000 lux, increase eye exercises to 4-5 times a day, and increase outdoor activity time to no less than 2 hours a day. It is also recommended to use eye protection equipment to assist in adjustment. Samples with low confidence are additionally advised to go to a regular ophthalmology hospital as soon as possible for professional examinations such as mydriasis refraction.

[0126] The integrated screening report combines pseudomyopia grading results, confidence markers, ciliary muscle fatigue status assessment, and vision health guidance plans into a unified document format. The report structure includes a basic screening information section that records the screening time, location, equipment model, and basic information of the subject. The classification results section records the pseudomyopia degree index, grading level, confidence level, etc. The ciliary muscle assessment section records the torque variation range, fatigue level, comparison of theoretical and actual myopia degrees, etc. The intervention recommendation section records personalized eye time arrangements, environmental adjustment plans, and vision care measures. The preliminary screening report uses a structured format for easy reading and subsequent processing. Key data is presented in tables, and charts are used to intuitively display trends and comparative relationships.

[0127] The added interpretation of screening results includes popular science content on three aspects: the causes, development trends, and prevention and control measures for pseudomyopia. The interpretation of causes explains the mechanism of accommodation spasm caused by long-term tension in the ciliary muscle. The development trend analyzes the possible development paths and outcomes of different degrees of pseudomyopia. The prevention and control measures introduce specific methods for scientific eye use and practical suggestions for vision protection. The interpretation uses easy-to-understand language, avoids excessive technical terms, and uses specific cases to illustrate the treatment methods for different situations, ultimately forming a pseudomyopia screening report.

[0128] The above describes the AI-based pseudomyopia rapid screening method in the embodiment of the present application. The following describes the AI-based pseudomyopia rapid screening system in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the AI-based pseudomyopia rapid screening system includes:

[0129] The processing module 201 is used to collect and process the original eye movement signal sequence and pupil response signal sequence of the subject to obtain an eye movement feedback indicator set;

[0130] A splicing module 202 is used to extract and splice multimodal features of the eye movement feedback indicator set to obtain a multimodal fusion feature vector;

[0131] An input module 203 is configured to construct a cascade forest classifier based on two forest algorithms, random forest and completely random forest, and input the multimodal fusion feature vector into the cascade forest classifier to calculate the category probability distribution and obtain a sample classification result;

[0132] A screening module 204 is used to screen the sample classification results for sample features to obtain a target feature combination;

[0133] The classification module 205 is used to classify the pseudomyopia of the subject according to the target feature combination to obtain a pseudomyopia screening report.

[0134] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, cause the computer to execute the steps of the AI-based pseudomyopia rapid screening method.

[0135] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0136] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling an AI-based pseudomyopia rapid screening device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0137] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A rapid screening method for pseudomyopia based on AI, characterized in that: The method comprises: The original eye movement signal sequence and pupil response signal sequence of the subject are collected and processed to obtain an eye movement feedback index set; Performing multimodal feature extraction and splicing on the eye movement feedback indicator set to obtain a multimodal fusion feature vector; Constructing a cascade forest classifier based on two forest algorithms, random forest and completely random forest, inputting the multimodal fusion feature vector into the cascade forest classifier to calculate the category probability distribution and obtain a sample classification result; Performing sample feature screening on the sample classification results to obtain a target feature combination; The subject is classified for pseudomyopia according to the target feature combination to obtain a pseudomyopia screening report.

2. The AI-based rapid screening method for pseudomyopia according to claim 1, characterized in that: The eye movement feedback indicator set includes: eye muscle adjustment amplitude, pupil contraction range, iris change amplitude, lens thickness change amplitude, and ciliary muscle torque change range. The eye movement feedback indicator set is obtained by collecting and processing the subject's eye movement original signal sequence and pupil response original signal sequence, including: Collecting the subject's eye movement original signal sequence and pupil response original signal sequence through an eye tracking device, and obtaining an eye movement vector including horizontal coordinates, vertical coordinates and a timestamp, and a pupil signal including a pupil diameter change and a corresponding timestamp; performing time synchronization on the eye movement vector and the pupil signal, and resampling the two signals to obtain a segmented signal; filtering the segmented signal through a Butterworth bandpass filter to obtain a preliminary filtered signal; performing normalization processing on the preliminary filtered signal, processing the eye movement signal and the pupil response signal separately to obtain a normalized signal; The normalized signal is subjected to a threshold method to detect and remove blink artifacts, with the upper threshold band being set at 95% of the signal amplitude and the lower threshold band being set at 5% of the signal amplitude, to obtain an artifact-free signal; Completing the missing data points in the artifact-removed signal using a cubic spline interpolation method to obtain preprocessed eye movement signals and pupil response signals; Performing eye muscle adjustment amplitude recognition on the preprocessed eye movement signal to obtain the eye muscle adjustment amplitude, wherein the eye muscle adjustment amplitude is used to reflect the degree of eye muscle tension and eye congestion; performing pupil contraction range analysis on the preprocessed pupil response signal to obtain the pupil contraction range; Measure the amplitude of iris changes and lens thickness changes; The ciliary muscle torque variation range is obtained, and the eye muscle adjustment amplitude, the pupil contraction range, the iris variation amplitude, the lens thickness variation amplitude, and the ciliary muscle torque variation range are combined into the eye movement feedback indicator set.

3. The AI-based rapid screening method for pseudomyopia according to claim 2, characterized in that: The obtaining of the ciliary muscle torque variation range includes: Set up multiple target distance test points, including a 30cm close-range test point and 3m, 5m, 10m, and 100m long-range test points; Measure the degree of ciliary muscle contraction of the subject at a close distance of 30 cm, and record the accommodation state of ciliary muscle contraction of 3.00D; Measure the degree of ciliary muscle dilation of the subject at distances of 3 meters, 5 meters, 10 meters, and 100 meters, and record the accommodation states of ciliary muscle dilation of 1.00D, 1.50D, 2.00D, and 3.00D respectively; Calculate the total torque variation of the ciliary muscle from a distance of 10 meters to a close distance of 30 cm. Under normal conditions, the torque variation should reach 5D. The ciliary muscle fatigue degree assessment result was obtained based on the difference between the measured ciliary muscle torque variation range and the theoretical normal 5D torque variation.

4. The AI-based rapid screening method for pseudomyopia according to claim 2, characterized in that: The multimodal feature extraction and splicing of the eye movement feedback indicator set to obtain a multimodal fusion feature vector includes: Calculating the variation between adjacent sampling points of the eye muscle adjustment amplitude to obtain an eye muscle adjustment speed sequence; Calculating statistical features of the eye muscle adjustment speed sequence, including maximum value, minimum value, average value and standard deviation, and performing frequency domain transformation to extract energy distribution features in the 0.1-10 Hz frequency band to obtain an eye muscle feature set; Calculating the pupil diameter change rate and timing characteristics for the pupil contraction range, measuring the ratio of the maximum pupil diameter to the minimum pupil diameter before and after the illumination change, and dividing the pupil into a pupil contraction period and a pupil recovery period, calculating the time length and change rate of the two stages respectively, to obtain a pupil feature set; Performing wavelet transform on the pupil feature set, selecting Daubechies wavelet basis function to perform multi-scale decomposition on pupil signal, extracting statistical characteristics of wavelet coefficients at different decomposition levels, and obtaining 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 concatenated in order of feature dimensions to form a feature vector; The principal component analysis algorithm is applied to the feature vector and Z-score normalization is performed to convert each dimension feature into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the dimension and value range differences, and obtaining a multimodal fusion feature vector.

5. The AI-based rapid screening method for pseudomyopia according to claim 1, characterized in that: The method comprises: constructing a cascade forest classifier based on two forest algorithms, random forest and completely random forest, inputting the multimodal fusion feature vector into the cascade forest classifier to calculate the category probability distribution, and obtaining a sample classification result, including: 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 to obtain the initial classification structure; Executing a feature selection strategy on the random forest, randomly selecting multiple features from the input features of the multimodal fusion feature vector as candidate features, calculating the Gini coefficient of each candidate feature, selecting the feature with the best Gini coefficient as the splitting point, and obtaining a node splitting scheme of the random forest; Performing a feature selection strategy on the completely random forest, directly and randomly selecting a single feature from the input features of the multimodal fusion feature vector as a split feature, and obtaining a node splitting scheme of the completely random forest; Input the multimodal fusion feature vector into the first layer of cascade forest, generate class probability features through cross-validation, divide the training data into 5 parts, use 4 parts to train the forest each time, and predict the class probability on the remaining part to obtain the output features of the first layer; The output features of the first layer are concatenated with the original multimodal fusion feature vector and input into the second layer cascade forest. The cross-validation prediction process is repeated to obtain the output features of the second layer. The output features of the second layer are spliced ​​with the original multimodal fusion feature vector and input into the third-layer cascade forest. Through feature transfer and enhancement between layers, the forest output of the final layer is weighted averaged to obtain sample classification results including normal vision, pseudomyopia and true myopia.

6. The AI-based rapid screening method for pseudomyopia according to claim 5, characterized in that: The output features of the first layer are concatenated with the original multimodal fusion feature vector, inputted into the second layer cascade forest, and the cross-validation prediction process is repeated to obtain the output features of the second layer, including: Calculate the class probabilities for each of the four forests in the first-level cascade forest. Each forest generates probability features equal to the number of classes. Combine the 12 probability feature values ​​to obtain the first-level probability feature vector. The first layer of probability feature vectors are directly concatenated with the original multimodal fusion feature vectors according to the feature dimensions to obtain an enhanced feature vector; Sorting the enhanced feature vectors by feature importance, calculating the contribution of each feature to the classification result using the average impurity reduction method, and obtaining a feature weight distribution; performing weighted processing on the features in the enhanced feature vector according to the feature weight distribution, highlighting the features with high contribution and reducing the influence of the features with low contribution, to obtain a weighted enhanced feature vector; Input the weighted enhanced feature vector into two random forests and two completely random forests in the second layer of cascade forests, and perform training and prediction in the same cross-validation manner as the first layer to obtain the class probability output of the second layer; The category probability output of the second layer is normalized, and the probability value is normalized to the interval [0, 1] using the softmax function, and the sum of the probabilities of each category is made 1 to obtain the output features of the second layer.

7. The AI-based rapid screening method for pseudomyopia according to claim 1, characterized in that: The sample feature screening of the sample classification results to obtain a target feature combination includes: Evaluate the contribution of the features in the sample classification results, calculate the impact of each feature on the classification accuracy, and obtain the feature importance score; Arrange the features in descending order according to the feature importance scores, establish a feature importance ranking list, and obtain a feature priority sequence; Recursive feature elimination is performed based on the feature priority sequence to remove the least important features from the feature set one by one, and the classification effect after each removal is recorded to obtain a curve showing the relationship between the number of features and the classification effect; Identify the inflection point position from the relationship curve between the number of features and the classification effect, determine the minimum number of features required to maintain the classification effect, and obtain a preliminary feature set; Performing correlation analysis on the pre-selected feature set, calculating correlation coefficients between features, identifying highly correlated feature pairs, and retaining the more important feature from each pair of highly correlated features to obtain a low-redundancy feature set; The low-redundancy feature set is used as a key indicator combination for screening pseudomyopia to form a feature standard for pseudomyopia screening and obtain a target feature combination.

8. The AI-based rapid screening method for pseudomyopia according to claim 1, characterized in that: The pseudomyopia classification of the subject according to the target feature combination to obtain a pseudomyopia screening report includes: Calculating the pseudomyopia degree index of the subject based on the target feature combination to obtain a quantitative score of the pseudomyopia degree; The theoretical myopia degree is calculated according to the variation range of the ciliary muscle torque, and the actual myopia degree of the subject is measured; Compare and analyze the actual myopia degree with the theoretical myopia degree: when the actual myopia degree is within 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 to be pseudomyopia; when the actual myopia degree is within the range of -0.25D to -1.50D and the theoretical myopia degree is calculated to be 0.00D or above, it is determined to be pseudomyopia; when the actual myopia degree is smaller than the theoretical myopia degree, it is determined to be ciliary muscle fatigue with a rapid development of myopia; According to the quantitative scoring of the pseudomyopia degree, pseudomyopia is subdivided into three levels: mild, moderate and severe, where 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, thereby obtaining a pseudomyopia grading result; Calculating the classification confidence for the pseudomyopia grading results, calculating the difference between the highest category probability and the second highest category probability, marking samples with a confidence level lower than 0.2 for further inspection, and obtaining a classification result with a confidence level mark; Generate personalized eye care recommendations based on the confidence-marked classification results, formulate corresponding eye care time arrangements, environmental lighting adjustments, and vision care plans based on the degree of pseudomyopia, and obtain a vision health guidance plan; Integrate the pseudomyopia grading results, confidence marks, ciliary muscle fatigue status assessment and vision health guidance plan into a comprehensive screening report, including screening time, basic information, classification results, confidence level and intervention recommendations, to obtain a preliminary pseudomyopia screening report; An explanation of the screening results is added to the preliminary pseudomyopia screening report, including the causes, development trends and prevention and control measures of pseudomyopia, to obtain a pseudomyopia screening report.

9. An AI-based rapid screening system for pseudomyopia, characterized in that: Used to implement the AI-based pseudomyopia rapid screening method according to any one of claims 1 to 8, the AI-based pseudomyopia rapid screening system comprises: A processing module is used to collect and process the original eye movement signal sequence and pupil response signal sequence of the subject to obtain an eye movement feedback indicator set; A splicing module, configured to extract and splice multimodal features of the eye movement feedback indicator set to obtain a multimodal fusion feature vector; An input module is used to construct a cascade forest classifier based on two forest algorithms: random forest and completely random forest, and input the multimodal fusion feature vector into the cascade forest classifier to calculate the category probability distribution and obtain the sample classification result; A screening module is used to screen the sample characteristics of the sample classification results to obtain a target feature combination; The classification module is used to classify the pseudomyopia of the subject according to the target feature combination to obtain a pseudomyopia screening report.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to perform the AI-based rapid pseudomyopia screening method according to any one of claims 1 to 8.

Images