Posterior chamber iol implantation postoperative information prediction method, device and medium
By comprehensively considering a feature set including demographics, ICL parameters, ocular anatomy, corneal endothelial cell function, and tear film stability, and by using machine learning models to optimize feature selection and parameters, the problem of insufficient accuracy in predicting post-ICL implantation information was solved, achieving higher prediction accuracy and post-operative results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGSHAN OPHTHALMIC CENT SUN YAT SEN UNIV
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-14
Smart Images
Figure CN122392810A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of artificial intelligence, and in particular to a method, device, and medium for predicting postoperative information in posterior chamber ICL implantation. Background Technology
[0002] Myopia has become an increasingly serious public health problem worldwide. Posterior chamber intraocular lens (ICL) implantation is a safe and effective method of refractive correction, suitable for patients with high myopia, thin corneas, and those who cannot tolerate laser surgery. One of the key factors for successful ICL implantation is achieving an ideal and stable postoperative arch, which is the vertical distance between the posterior surface of the ICL and the anterior capsule of the lens. Too low an arch may increase the risk of subcapsular anterior lens opacity, while too high an arch may cause complications such as angle closure and secondary glaucoma. Furthermore, precise control of postoperative residual refractive power directly affects the patient's uncorrected visual acuity and visual quality.
[0003] Currently, the selection of ICL models and the calculation of power are mainly based on preoperative measurements of ocular anatomical parameters (such as white-to-white distance, anterior chamber depth, and corneal curvature), and rely on empirical formulas or online calculators provided by manufacturers. With the introduction of artificial intelligence technology into the field of ophthalmic surgery prediction, some studies have attempted to use machine learning models such as random forests and gradient boosting trees to predict the arch height based on richer preoperative anatomical parameters (such as angle-to-angle distance, anterior chamber width, and ciliary sulcus size), and have made some progress.
[0004] However, existing machine learning models almost entirely rely on ocular anatomical parameters, and in clinical practice, they often result in large discrepancies between the predicted arch height and the actual arch height, or unexpected refractive errors after surgery. Summary of the Invention
[0005] This disclosure is made in view of the above-mentioned circumstances, and its purpose is to provide a method, device, and medium for predicting postoperative information in posterior chamber ICL implantation that can improve the accuracy of postoperative information prediction.
[0006] Therefore, the first aspect of this disclosure provides a method for predicting postoperative information of phakic posterior chamber ICL implantation, comprising: acquiring a first feature set related to demographics for a first time period of a target subject; acquiring a second feature set related to ICL parameters of the ICL implantation for the first time period of the target subject; acquiring a third feature set related to ocular anatomy for the first time period of the target subject; acquiring a fourth feature set related to the functional characteristics of corneal endothelial cells for the first time period of the target subject; and processing the target feature set including the first feature set, the second feature set, the third feature set, and the fourth feature set using a machine learning-based prediction model to obtain postoperative information of the ICL implantation for a second time period after the first time period, wherein the first time period is the preoperative time period of the ICL implantation, the second time period is the postoperative time period of the ICL implantation, and the postoperative information includes at least one of arch height, spherical refractive power, and cylindrical refractive power. This improves the accuracy of postoperative information prediction. Furthermore, incorporating the fourth feature set related to the functional characteristics of corneal endothelial cells into the target feature set further improves the accuracy of postoperative information prediction.
[0007] Additionally, in the prediction method according to the first aspect of this disclosure, optionally, when predicting the arch height, the first feature set includes intraocular pressure; the second feature set includes ICL size; the third feature set includes anterior chamber depth, axial length, anterior chamber volume, and corneal curvature K2; and the fourth feature set includes hexagonal cell percentage and endothelial cell variation coefficient.
[0008] Additionally, in the prediction method involved in the first aspect of this disclosure, optionally, when predicting spherical refractive power: the first feature set includes gender; the second feature set includes ICL size; the third feature set includes axial length; and the fourth feature set includes hexagonal cell percentage.
[0009] Additionally, in the prediction method involved in the first aspect of this disclosure, optionally, when predicting cylindrical refractive power: the first feature set includes gender and cylindrical refractive power; the second feature set includes ICL type, ICL cylindrical refractive power, and ICL spherical refractive power; the third feature set includes axial length, the axis of corneal curvature K2, corneal curvature K2, corneal curvature K1, the axis of corneal curvature K1, and white-to-white distance; the fourth feature set includes endothelial cell density and endothelial cell variation coefficient.
[0010] Additionally, in the prediction method according to the first aspect of this disclosure, optionally, when predicting cylindrical refractive power, the prediction method further includes: acquiring a fifth feature set related to tear film stability for the target object during the first time period, wherein the fifth feature set includes non-invasive tear river height; acquiring a sixth feature set related to visual acuity for the target object during the first time period, wherein the sixth feature set includes uncorrected visual acuity and best-corrected visual acuity; and further including the fifth and sixth feature sets in the target feature set. In this case, incorporating the fifth feature set related to tear film stability into the target feature set for predicting cylindrical refractive power can further improve the accuracy of cylindrical refractive power prediction.
[0011] Furthermore, in the prediction method according to the first aspect of this disclosure, optionally, feature filtering is performed on the initial feature set to obtain the target feature set and the model parameters of the prediction model; the feature filtering method is the Marine Predator Algorithm; and the prediction model is an SVM. In this case, MPA, as a metaheuristic optimizer with powerful global search capabilities, can improve prediction efficiency when used to filter the optimal feature subset from the initial feature set and optimize the model parameters of the SVM.
[0012] Additionally, in the prediction method according to the first aspect of this disclosure, optionally, the initial feature set includes a first initial feature set, a second initial feature set, a third initial feature set, a fourth initial feature set, a fifth initial feature set, and a sixth initial feature set; the first initial feature set includes age, gender, height, weight, intraocular pressure, spherical refractive power, cylindrical refractive power, and / or the axis of the cylindrical refractive power; the second initial feature set includes ICL type, ICL size, ICL spherical refractive power, ICL cylindrical refractive power, and / or ICL cylindrical refractive power. The initial feature set includes: axial length, anterior chamber depth, anterior chamber volume, white-to-white distance, pupil diameter, corneal diameter, central corneal thickness, corneal curvature K1, the axis of corneal curvature K1, corneal curvature K2, and / or the axis of corneal curvature K2; the fourth initial feature set includes endothelial cell density, endothelial cell coefficient of variation, and / or hexagonal cell percentage; the fifth initial feature set includes non-invasive tear river height and / or tear film breakup time; and the sixth initial feature set includes uncorrected visual acuity and / or best-corrected visual acuity. This helps to screen features relevant to postoperative information.
[0013] Additionally, in the prediction method according to the first aspect of this disclosure, the prediction method may optionally further include: removing the fourth feature set and / or the fifth feature set from the initial feature set used to screen the target feature set to obtain a validation feature set; obtaining a trained machine learning-based validation model based on the validation feature set; and comparing the performance metrics of the prediction model and the validation model on the same test set to verify the effectiveness of the fourth feature set and / or the fifth feature set in predicting the postoperative information.
[0014] A second aspect of this disclosure provides an electronic device including a processor and a memory, the memory storing a computer program that, when executed, implements the prediction method as described in the first aspect of this disclosure.
[0015] A third aspect of this disclosure provides a computer-readable storage medium storing at least one instruction that, when executed by a processor, implements the prediction method as described in the first aspect of this disclosure.
[0016] According to this disclosure, a method, device, and medium for predicting postoperative information in posterior chamber ICL implantation that can improve the accuracy of postoperative information prediction are provided. Attached Figure Description
[0017] This disclosure will now be explained in further detail by way of example only with reference to the accompanying drawings.
[0018] Figure 1 This is a schematic diagram illustrating an example of a prediction environment covered by the examples of this disclosure.
[0019] Figure 2 This is an exemplary flowchart illustrating the prediction method involved in the examples of this disclosure.
[0020] Figure 3 This is an exemplary flowchart illustrating the selection process involved in the examples of this disclosure.
[0021] Figure 4 This is an exemplary flowchart illustrating the model parameters of the target feature set and prediction model based on the marine predator algorithm involved in this disclosure.
[0022] Figure 5A This is a comparison chart showing the ranking of feature importance in predicting arch height as described in the examples of this disclosure.
[0023] Figure 5B This is a comparative diagram showing the ranking of feature importance in predicting spherical refractive power as described in the examples of this disclosure.
[0024] Figure 5CThis is a comparative diagram showing the ranking of feature importance in predicting cylindrical refractive power as described in the examples of this disclosure.
[0025] Figure 6 This is an exemplary flowchart illustrating the validation prediction model involved in the examples of this disclosure.
[0026] Figure 7 This is an exemplary block diagram illustrating an electronic device to which the present disclosure is based. Detailed Implementation
[0027] Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same components, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the components or the shapes of the components may differ from actual figures. It should be noted that the terms "comprising" and "having," and any variations thereof, in this disclosure, do not necessarily limit the process, method, system, product, or apparatus to the explicitly listed steps or units, but may include or have other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.
[0028] First, let me introduce the relevant terminology used in this disclosure.
[0029] "Phacophthalmic ICL implantation" refers to surgery performed on patients whose natural lens remains intact and is transparent. "Phacophthalmic" means the implanted artificial lens is placed in the posterior chamber of the eye. ICL implantation is a minimally invasive procedure to insert a specially designed artificial lens into the eye.
[0030] This disclosure provides an example of a method for predicting postoperative information in phakic posterior chamber ICL implantation (hereinafter referred to as ICL implantation), which can improve the accuracy of postoperative information prediction. Furthermore, the method for predicting postoperative information in phakic posterior chamber ICL implantation described in this disclosure can also be referred to as an evaluation method or an analysis method (hereinafter referred to as a prediction method).
[0031] Examples of this disclosure will now be described in detail with reference to the accompanying drawings. Figure 1 This is a schematic diagram illustrating an example of the prediction environment involved in the examples of this disclosure. It should be noted that this is not intended to limit the scope of this disclosure.
[0032] refer to Figure 1The prediction environment may include prediction system 100, which can be used to implement the prediction methods involved in the examples of this disclosure. Prediction system 100 can be any device with computing capabilities. For example, prediction system 100 can be a cloud server, personal computer, mainframe, and distributed computing system, etc. Continue to refer to Figure 1 The prediction environment may include a data acquisition device 9 (e.g., a corneal endothelial lens), which can be used to acquire preoperative information for ICL implantation and transmit it to the prediction system 100; the prediction system 100 can process the preoperative information to obtain postoperative information.
[0033] Furthermore, for ease of understanding, the prediction objective of the prediction method will be introduced before describing the prediction method itself. As mentioned above, the prediction method described in this disclosure can predict postoperative information. Postoperative information can be used to indicate the impact of ICL implantation on ocular anatomy. In some examples, postoperative information can be related to fundus morphology. In some examples, postoperative information related to fundus morphology may include at least one of arch height, spherical refractive power, and cylindrical refractive power. In some examples, postoperative information may also include the amount of correction, percentage of correction, deviation from the target value (i.e., the postoperative desired refractive power), or equivalent spherical power related to at least one of spherical and cylindrical refractive power. In some examples, postoperative information may also include the axial direction or axial variation of the cylindrical refractive power.
[0034] Figure 2 This is an exemplary flowchart illustrating a prediction method according to an example of this disclosure. Furthermore, at least one step in the prediction method may be implemented by at least a portion of the components in the prediction system 100. Additionally, the prediction model for processing the target feature set in the prediction method according to an example of this disclosure may be pre-trained.
[0035] In some examples, reference Figure 2 The prediction method described in this disclosure can improve the accuracy of postoperative information prediction. Furthermore, incorporating a fourth feature set related to the functional characteristics of corneal endothelial cells into the target feature set can further improve the accuracy of postoperative information prediction.
[0036] In some examples, reference Figure 2 The prediction method may include step S101. In step S101, a first set of demographic features related to the target object for a first time period can be obtained. The first time period may be the preoperative period of ICL implantation. In addition, the target object may be any biological entity (e.g., a human).
[0037] In some examples, the first feature set may include at least one of sex, cylindrical refractive power, and intraocular pressure (IOP). The specific features selected are those that are relevant to the prediction target (i.e., relevant to the postoperative information to be predicted). In some examples, the first feature set may include IOP when predicting vault height. In some examples, the first feature set may include sex when predicting spherical refractive power. In some examples, the first feature set may include both sex and cylindrical refractive power when predicting cylindrical refractive power.
[0038] In some examples, reference Figure 2 The prediction method may include step S102. In step S102, a second feature set related to the ICL parameters of the ICL implantation surgery for a first time period may be obtained. In some examples, the second feature set can be obtained using a formula provided by the ICL supplier.
[0039] In some examples, the second feature set may include at least one of ICL size, ICL type, ICL cylinder power, and ICL spherical power. The specific features selected may be relevant to the prediction target. In some examples, when predicting the arch height, the second feature set may include intraocular pressure. In some examples, when predicting spherical power, the second feature set may include ICL size. In some examples, when predicting cylinder power, the second feature set may include ICL type, ICL cylinder power, and ICL spherical power.
[0040] In some examples, reference Figure 2 The prediction method may include step S103. In step S103, a third feature set related to the ocular anatomy of the target object for a first time period may be obtained. In some examples, the third feature set may be obtained using a non-contact ocular biometric instrument.
[0041] In some examples, the third feature set may include at least one of anterior chamber depth (ACD), axial length (AL, sometimes simply referred to as axial length), anterior chamber volume (ACV), corneal curvature K2, the axis of corneal curvature K2, corneal curvature K1, the axis of corneal curvature K1, and white-to-white distance (WTW). The specific features selected may be relevant to the prediction target. In some examples, when predicting vault height, the third feature set may include anterior chamber depth, axial length, anterior chamber volume, and corneal curvature K2. In some examples, when predicting spherical refractive power, the third feature set may include axial length. In some examples, when predicting cylindrical refractive power, the third feature set may include axial length, the axis of corneal curvature K2, corneal curvature K2, corneal curvature K1, the axis of corneal curvature K1, and white-to-white distance.
[0042] The inventors accidentally discovered that incorporating functional features (such as anterior segment functional features, and more specifically, the fourth and / or fifth feature sets described later) involved in the prediction of postoperative information can further improve the accuracy of predicting postoperative information. That is, the fourth and / or fifth feature sets can play a crucial role. In the prior art, because the prediction focuses on information related to ocular anatomy (e.g., fundus morphology), researchers generally believe that functional features such as the fourth and / or fifth feature sets should not affect fundus morphology, and therefore pay little attention to these functional features. For example, postoperative information is generally predicted through ocular anatomy features such as white-to-white distance (WTW). The inventors have overcome this technical bias by proposing new technical solutions involving new features for predicting postoperative information. Furthermore, the fourth feature set, related to corneal endothelial cells, is cellular-level information and is clinically considered a set of functional parameters, generally believed to have little impact on macroscopic structures. Furthermore, regarding the fifth feature set related to tear film stability, which will be discussed later, the tear film is a layer of water in front of the cornea, which has an impact on function, but it is generally not considered to have such a large impact on this structure.
[0043] The corneal endothelial cell layer is crucial for maintaining corneal transparency and the aqueous humor barrier function. Decreased endothelial cell density and abnormal morphology can affect corneal hydration and biomechanical properties, indirectly impacting postoperative anterior chamber stability. The tear film, as the optical interface of the ocular surface, directly affects the accuracy of corneal curvature measurements and visual quality. Preoperative tear film instability can lead to errors in corneal topography measurements, resulting in inaccurate ICL refractive index calculations. Currently, no publicly available technology systematically integrates these "functional parameters" related to corneal endothelial cells and / or the tear film into a predictive model for postoperative information after ICL implantation.
[0044] In some examples, reference Figure 2 The prediction method may include step S104. In step S104, a fourth feature set related to the functional characteristics of corneal endothelial cells of the target object over a first time period may be obtained. In some examples, the fourth feature set may include at least one of endothelial cell density, hexagonal cell percentage (also referred to as hexagonal cell percentage), and endothelial cell coefficient of variation. Specific selection of which features are relevant to the prediction target is made. In some examples, the fourth feature set can be obtained using corneal endothelial microscopy.
[0045] In some examples, when predicting arch height, the fourth feature set may include the percentage of hexagonal cells and the coefficient of variation of endothelial cells. In some examples, when predicting spherical refractive power, the fourth feature set may include the percentage of hexagonal cells. In some examples, when predicting cylindrical refractive power, the fourth feature set may include endothelial cell density and the coefficient of variation of endothelial cells.
[0046] In some examples, when predicting cylindrical refractive power, the prediction method may further include obtaining a fifth feature set related to tear film stability for a first time period of the target object, thus including the fifth feature set in the target feature set. That is, when predicting cylindrical refractive power, the target feature set may also include the fifth feature set. In this case, incorporating the fifth feature set related to tear film stability into the target feature set used for predicting cylindrical refractive power can further improve the accuracy of cylindrical refractive power prediction. In some examples, the fifth feature set may include non-invasive tear river height (sometimes simply referred to as tear river height). In some examples, the fifth feature set can be obtained using an ocular surface analyzer.
[0047] In some examples, when predicting cylindrical refractive power, the prediction method may further include obtaining a sixth feature set related to visual acuity for a first time period of the target object, such that the target feature set also includes the sixth feature set. That is, when predicting cylindrical refractive power, the target feature set may also include the sixth feature set. In some examples, the sixth feature set may include uncorrected visual acuity and best-corrected visual acuity.
[0048] In some examples, reference Figure 2 The prediction method may include step S105. In step S105, the target feature set can be processed by a machine learning-based prediction model to obtain postoperative information of ICL implantation in a second time period after the first time period. The target feature set includes a first feature set, a second feature set, a third feature set, and a fourth feature set. Specifically, the target feature set can be input into the prediction model to output postoperative information.
[0049] Additionally, the second time period can be the postoperative period after ICL implantation. In some examples, the second time period can be one month after ICL implantation. This better reflects real-world application scenarios. In some examples, there can be multiple second time periods, which can be increased or decreased sequentially. This facilitates the acquisition of stable postoperative information, such as the stability of the arch height.
[0050] In some examples, different postoperative information can be predicted using independent predictive models. Furthermore, "independent" can mean that each predictive model is trained separately. In some examples, the predictive models may include at least one of a first model, a second model, and a third model. The first model can be used to predict the camber in the second time period. The second model can be used to predict the spherical refractive power in the second time period. The third model can be used to predict the cylindrical refractive power in the second time period.
[0051] In some examples, the system can receive the selection of predicted postoperative information from a source user. In response to the selected postoperative information being arch height, a first model and its corresponding target feature set can be used to predict the arch height; in response to the selected postoperative information being spherical refractive power, a second model and its corresponding target feature set can be used to predict the spherical refractive power; and in response to the selected postoperative information being cylindrical refractive power, a third model and its corresponding target feature set can be used to predict the cylindrical refractive power. This facilitates the use of models and target feature sets suitable for each type of postoperative information to predict the corresponding postoperative information.
[0052] In some examples, the target feature sets and model parameters for the first, second, and third models can be obtained through independent feature selection and training. Additionally, model parameters can include model parameter terms (i.e., the names of the model parameters) and model parameter values (i.e., the values that the model parameters take).
[0053] In some examples, the prediction model can be a regression model, thus enabling the prediction of specific numerical values. In other examples, the prediction model can be an SVM (support vector machine). In some examples, the SVM model parameters can be a penalty factor and the bandwidth of the radial basis function kernel.
[0054] In some examples, the target feature set and the model parameters of the prediction model can be obtained through feature filtering. Specifically, the initial feature set can be filtered to obtain the target feature set and the model parameters (also known as hyperparameters) of the prediction model.
[0055] In some examples, the initial feature set may include at least one of a first, second, third, fourth, fifth, and sixth initial feature set. The first initial feature set may be related to demographics. The second initial feature set may be related to ICL parameters from ICL implantation. The third initial feature set may be related to ocular anatomy. The fourth initial feature set may be related to functional characteristics of corneal endothelial cells. The fifth initial feature set may be related to tear film stability. The sixth initial feature set may be related to visual acuity. This helps to screen features relevant to postoperative information.
[0056] In some examples, the first initial feature set may include age, sex, height, weight, intraocular pressure (IOP), spherical refractive power, cylindrical refractive power, and / or the axis of the cylindrical refractive power. For example, the target subject's age, sex, and weight may be collected, IOP may be measured using a non-contact tonometer, and spherical refractive power, cylindrical refractive power, and axis may be collected based on subjective refraction results. In some examples, the first initial feature set may also include equivalent spherical refractive power.
[0057] In some examples, the second initial feature set may include ICL type (also known as ICL category), ICL size, ICL spherical power, ICL cylindrical power and / or the axis of ICL cylindrical power.
[0058] In some examples, the third initial feature set may include axial length (AL), anterior chamber depth (ACD), anterior chamber volume (ACV), white-to-white distance (WTW), pupil diameter, corneal diameter, central corneal thickness (CCT), corneal curvature K1, the axis of corneal curvature K1, corneal curvature K2, and / or the axis of corneal curvature K2.
[0059] In some examples, the fourth initial feature set may include endothelial cell density, endothelial cell coefficient of variation, and / or percentage of hexagonal cells.
[0060] In some examples, the fifth initial feature set may include non-invasive tear river height and / or tear film breakup time.
[0061] In some examples, the sixth initial feature set may include uncorrected visual acuity (UDVA) and / or best-corrected visual acuity (CDVA).
[0062] Figure 3 This is an exemplary flowchart illustrating the selection process involved in the examples of this disclosure.
[0063] In some examples, feature selection methods can involve ranking and optimizing subsets of features from an initial feature set. Specifically, this can be done by evaluating the predictive performance of all feature subsets to determine the target feature set and the trained predictive model. In some examples, feature selection methods can employ swarm intelligence optimization algorithms. In some examples, swarm intelligence optimization algorithms can be used to search for candidate model parameters based on the initial feature set and the predictive model to be trained (hereinafter referred to as the candidate model) to determine the trained predictive model, the model parameters of the trained predictive model, and the target feature set. That is, the best feature terms and model parameters are selected based on swarm intelligence optimization algorithms, and a predictive model is constructed.
[0064] For example, refer to Figure 3The selection process may include: randomly selecting a subset of features from the initial feature set using a swarm intelligence optimization algorithm, and randomly selecting values within the range of model parameter values as candidate model parameters (step S201); dividing the dataset constructed based on the feature subset into a training set and a test set; inputting the training set into a candidate model with candidate model parameters; and then using the test set to obtain various performance metrics of the candidate model (e.g., mean absolute error (MAE), median absolute error (MAE)). Error (MedAE) and coefficient of determination (R-squared, R2) (step S202); select a new feature subset according to the search principle of swarm intelligence optimization algorithm (step S203); repeat step S202; comprehensively compare multiple performance indicators of multiple candidate models, and take the feature subsets of the top-ranked candidate models as at least one optimal feature subset (step S204); iterate the candidate model corresponding to each optimal feature subset a preset number of times (e.g., 100 times), take the optimal feature subset with the smallest comprehensive performance indicators as the target feature set, take the candidate model corresponding to the target feature set as the trained prediction model, and take the model parameters corresponding to the target feature set as the model parameters of the trained prediction model (step S205).
[0065] Figure 4 This is an exemplary flowchart illustrating the model parameters of the target feature set and prediction model based on the marine predator algorithm involved in this disclosure.
[0066] In some examples, the swarm intelligence optimization algorithm can be the Marine Predator Algorithm (MPA). This allows for the selection of the most relevant features from smaller datasets and more effective management of parameter correlations. To this end, this disclosure also provides an example of obtaining a target feature set and model parameters for a prediction model based on the Marine Predator Algorithm. Reference Figure 4 This example may include: initializing the prey matrix and elite matrix based on the model parameter terms and initial feature set of the candidate model (step S301); calculating the fitness of each individual in the prey matrix based on the candidate model (step S302); iteratively updating the prey matrix and elite matrix based on the fitness until the stopping condition is met to obtain the optimized prey matrix and elite matrix and simultaneously obtain the trained candidate model (i.e., the trained prediction model) (step S303); and obtaining the target feature set and the model parameter values of the trained candidate model based on the optimized prey matrix (step S304). Furthermore, the fitness can be determined by the prediction error of the candidate model corresponding to the individual. Each individual can correspond to a feature subset and a set of model parameter values.
[0067] In some examples, when using the Marine Predator Algorithm (MPA), the prediction model can be a Support Vector Machine (SVM). That is, the MPA can be combined with SVM. In this case, the MPA, as a metaheuristic optimizer with powerful global search capabilities, can improve prediction efficiency when used to select the optimal subset of features from the initial feature set and optimize the SVM model parameters. It should be noted that this is not intended to limit the scope of this disclosure, and other machine learning models can also be used.
[0068] In some examples, all features in the initial feature set can be sorted according to their importance obtained by permutation importance to determine the features used to train the prediction model. In some examples, features in the initial feature set with an importance greater than 0.50 can be used as the target feature set for the prediction model.
[0069] Figure 5A This is a comparison chart showing the ranking of feature importance in predicting arch height as described in the examples of this disclosure. Figure 5B This is a comparative diagram showing the ranking of feature importance in predicting spherical refractive power as described in the examples of this disclosure. Figure 5C This is a comparative diagram showing the ranking of feature importance in predicting cylindrical refractive power, as illustrated in the examples of this disclosure. Specifically, in... Figure 5A The middle section compares and ranks the feature importance of the predictive model and the endothelial-free validation model. Figure 5B The middle section compares and ranks the feature importance of the predictive model and the endothelial-free validation model. Figure 5C The comparison involves ranking and ranking the feature importance of the prediction model, the endothelial-free validation model, the tear film-free validation model, and the double-no validation model. Furthermore, the endothelial-free validation model can refer to a validation model trained after removing the fourth feature set. The tear film-free validation model can refer to a validation model trained after removing the fifth feature set. The double-no validation model can refer to a validation model trained after simultaneously removing both the fourth and fifth feature sets.
[0070] In some examples, reference Figure 5A , Figure 5B and Figure 5C The predictive models used to predict different postoperative information have different feature importance. Some features with lower importance (e.g., features with importance less than 0.12) are not shown in the figure.
[0071] Figure 6 This is an exemplary flowchart illustrating the validation prediction model involved in the examples of this disclosure.
[0072] In addition, this disclosure provides examples of validating predictive models. In some examples, references are made to... Figure 6 Validating the prediction model may include removing the fourth and / or fifth feature sets from the features related to the prediction model (hereinafter referred to as relevant features) to obtain a validation feature set (step S401), obtaining a trained machine learning-based validation model based on the validation feature set (step S402), and comparing the performance metrics (e.g., prediction error) of the prediction model and the validation model on the same test set to verify the effectiveness of the fourth and / or fifth feature sets in predicting postoperative information (step S403) (i.e., the impact on the accuracy of predicting postoperative information).
[0073] In some examples, in step S401, if a fourth feature set and / or a fifth feature set exist among the relevant features, the elimination can include at least one of the following methods: eliminating the fourth feature set; eliminating the fifth feature set; eliminating both the fourth and fifth feature sets. This facilitates comparing the prediction model and the validation model from different feature dimensions.
[0074] In some examples, certain features can be removed when predicting postoperative information. In some examples, the fourth feature set can be removed when predicting arch height. In some examples, the fourth feature set can be removed when predicting spherical refractive power. In some examples, the fourth feature set, the fifth feature set, and a combination of both (i.e., removing both the fourth and fifth feature sets simultaneously) can be removed when predicting cylindrical refractive power. In some examples, the relevant features can be the initial feature set. In other examples, the relevant features can be directly the target feature set of the prediction model. That is, the fourth and / or fifth feature sets in the initial or target feature set can be removed to obtain the validation feature set.
[0075] In some examples, step S402 also involves re-obtaining the model for predicting postoperative information based on the validation feature set. For example, this could involve retraining the model based on the validation feature set, or re-selecting features and retraining the model based on the validation feature set.
[0076] In some examples, if the relevant features are the initial feature set, in step S402, a swarm intelligence optimization algorithm can be used to search based on the candidate model parameters and the validation feature set of the validation model to determine the trained validation model, the model parameters of the trained validation model, and the target feature set of the trained validation model. Refer to the relevant description of the prediction model for details.
[0077] In some examples, if the relevant features are the target feature set of the prediction model, in step S402, a verification model can be trained based on the verification feature set to obtain the trained verification model and the model parameters of the trained verification model, and the verification feature set can be used as the target feature set of the verification model.
[0078] In some examples, the validation model can be the same as the predictive model. That is, the validation model can be obtained using the same machine learning algorithm as the predictive model. This improves the consistency of the comparison. For example, models for predicting octave height, spherical refractive power, and cylindrical refractive power can be trained separately using the same machine learning algorithm as the predictive model.
[0079] In some examples, in step S403, the performance metrics may include at least one of mean absolute error (MAE), median absolute error (MedAE), coefficient of determination (R²), and percentage of regression error.
[0080] In some examples, the dataset used to train the prediction model can be data from sample subjects (i.e., the target subjects) that have been excluded from having fundus abnormalities. This reduces the influence of other unknown factors on the predicted postoperative information. In some examples, fundus abnormalities can include at least one of the following: abnormal intraocular pressure, corneal endothelial cell density less than a preset value (e.g., 1800 cells / mm²), postoperative complications, and the presence of other eye diseases. For example, after excluding abnormal intraocular pressure, corneal endothelial cell density less than 1800 cells / mm², postoperative complications, and other eye diseases, data from 282 eyes of 142 sample subjects were finally included, of which 128 eyes underwent ICL surgery and 154 eyes underwent TICL surgery.
[0081] In some examples, the dataset used to train the predictive model may also include postoperative information from a second time period. This allows the predictive model to acquire postoperative information based on a target feature set during training. For example, the dataset used to train the predictive model may also include arch height, spherical refractive power, and cylindrical refractive power one month postoperatively.
[0082] Figure 7 This is an exemplary block diagram illustrating an electronic device 8 as described in this disclosure.
[0083] The examples of this disclosure also relate to an electronic device 8, see reference 8. Figure 7 The electronic device 8 may include a processor 801 (e.g., a central processing unit or a graphics processing unit) and a memory 802. The memory 802 may store a computer program that, when executed, implements one or more steps of the prediction method described above.
[0084] In some examples, memory 802 may include a computer-readable storage medium, which may include, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory, etc. For example, computer-executable instructions may be loaded from memory unit 807 (described later) into random access memory (RAM) to execute computer-executable instructions. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, USB storage, flash memory, etc.
[0085] In some examples, reference Figure 7 The processor 801 and memory 802 can be connected to each other via bus 803. In some examples, the electronic device 8 may also include an input / output interface 804, which can be connected to bus 803. In some examples, the electronic device 8 may also include an input unit 805 (e.g., touch screen, keyboard, mouse, camera, microphone, etc.), an output unit 806 (e.g., display, speaker, etc.), a storage unit 807 (e.g., magnetic tape, hard disk, flash memory, etc.), and / or a communication unit 808, which can be connected to the input / output interface 804. Additionally, the communication unit 808 can be used to enable the electronic device 8 to communicate wirelessly or wiredly with other electronic devices.
[0086] Examples of this disclosure also disclose a computer-readable storage medium that can store at least one instruction, which, when executed by a processor, implements one or more steps of the prediction method described above. The computer-readable storage medium can be, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, and magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.
[0087] In the examples disclosed herein, functional indicators such as corneal endothelial cell morphology and density, and tear film stability are integrated into the postoperative prediction of ICL implantation for the first time, which significantly improves the prediction accuracy compared to models based on structural parameters.
[0088] Additionally, this disclosure provides a comparison of the feature importance rankings for the prediction and validation models. In the prediction arch, refer back to the reference. Figure 5A Before and after removing the fourth feature set (i.e., before and after removing endothelial function), the most important features were ICL size, anterior chamber depth, anterior chamber volume, and axial length. In predicting spherical refractive power, the reference... Figure 5BBefore and after removing the fourth feature set, the top few most important features were axial length, gender, and ICL size. In predicting spherical refractive power, the reference... Figure 5C Before and after removing the fourth and / or fifth feature sets, the top few most important features are ICL type, ICL cylinder refractive power, gender, and the axis of corneal curvature K2. Additionally, from... Figure 5A , Figure 5B and Figure 5C It can also be seen that the fourth and / or fifth feature sets are also among the most important in predicting the corresponding postoperative information, indicating that they play an important role in the prediction of the corresponding postoperative information.
[0089] To verify the effectiveness of the prediction model involved in this disclosure, the inventors compared its performance with a corresponding validation model. In the comparison, both the prediction model and the validation model are SVMs, employing the Ocean Predator algorithm for feature selection. Specifically, the prediction model selects features based on an initial feature set, while the validation model selects features based on a validation feature set. The final target feature sets for both the prediction and validation models are... Figure 5A , Figure 5B and Figure 5C Features with a corresponding importance greater than 0.5.
[0090] Table 1 shows a performance comparison between the prediction model and the corresponding validation model:
[0091] Table 1. Performance comparison of the prediction model and its corresponding validation model. As shown in Table 1, in the prediction of corneal arch height, the predictive model performed better than the validation model. The MAE of the predictive model was 138.55, while that of the validation model was 142.32. The MedAE of the predictive model was 110.41, while that of the validation model was 115.84. In other words, both the MAE and MedAE of the predictive model were lower than those of the validation model. This indicates that parameters related to corneal endothelial cells significantly improved the accuracy of arch height prediction. In the prediction of spherical refractive power, the predictive model showed some improvement over the validation model, with a lower median absolute error and an increased proportion of prediction errors within ±0.25D, ±0.50D, and ±0.75D. In the prediction of cylindrical refractive power, the predictive model outperformed both the predictive and validation models, while the model without either lens or lens performed the worst. In other words, in cylinder refractive error prediction, combining tear film, endothelial parameters and other ocular parameters yields better prediction results (MAE=0.38; MedAE=0.32; R²=0.19), which is significantly better than the model that does not use tear film and endothelial parameters.
[0092] Furthermore, the study disclosed herein has been approved by the Ethics Review Committee of Zhongshan Hospital, and all research procedures have followed the principles of the Declaration of Helsinki. Each participant received a detailed explanation of the procedures and signed an informed consent form before treatment. All identifiable information of patients in this study has been concealed.
[0093] While the present disclosure has been specifically described above in conjunction with the accompanying drawings and examples, it is to be understood that the foregoing description does not limit the present disclosure in any way. Those skilled in the art can make modifications and variations to the present disclosure as needed without departing from its essential spirit and scope, and all such modifications and variations shall fall within the scope of the present disclosure.
Claims
1. A method for predicting postoperative information in phakic posterior chamber ICL implantation, characterized in that, Obtain the first set of demographic features of the target object in the first time period; Obtain the second feature set related to the ICL parameters of the ICL implantation surgery during the first time period; Obtain the third feature set related to the eye anatomy of the target object in the first time period; Obtain the fourth feature set related to the functional characteristics of corneal endothelial cells of the target object during the first time period; and The target feature set, including the first feature set, the second feature set, the third feature set, and the fourth feature set, is processed by a machine learning-based prediction model to obtain postoperative information of ICL implantation in a second time period after the first time period. The first time period is the preoperative time period of ICL implantation, and the second time period is the postoperative time period of ICL implantation. The postoperative information includes at least one of arch height, spherical refractive power, and cylindrical refractive power.
2. The prediction method according to claim 1, characterized in that, When predicting arch height: The first feature set includes intraocular pressure; The second feature set includes the ICL size; The third feature set includes anterior chamber depth, axial length, anterior chamber volume, and corneal curvature K2; The fourth feature set includes the percentage of hexagonal cells and the coefficient of variation of endothelial cells.
3. The prediction method according to claim 1, characterized in that, When predicting spherical refractive power: The first feature set includes gender; The second feature set includes the ICL size; The third feature set includes axial length; The fourth feature set includes the percentage of hexagonal cells.
4. The prediction method according to claim 1, characterized in that, When predicting cylindrical refractive power: The first feature set includes gender and cylindrical refractive power; The second feature set includes ICL type, ICL cylindrical refractive power, and ICL spherical refractive power; The third feature set includes axial length, the axis of corneal curvature K2, corneal curvature K2, corneal curvature K1, the axis of corneal curvature K1, and white-to-white distance; The fourth feature set includes endothelial cell density and endothelial cell variation coefficient.
5. The prediction method according to claim 4, characterized in that, When predicting cylindrical lens refractive power, the prediction method further includes: Obtain a fifth feature set related to tear film stability for the target object during the first time period, wherein the fifth feature set includes non-invasive tear river height; Obtain a sixth feature set related to vision for the target object during the first time period, wherein the sixth feature set includes uncorrected visual acuity and best corrected visual acuity; The target feature set may also include the fifth feature set and the sixth feature set.
6. The prediction method according to claim 1, characterized in that, The initial feature set is subjected to feature filtering to obtain the target feature set and the model parameters of the prediction model; The feature selection method is the marine predator algorithm; The prediction model is SVM.
7. The prediction method according to claim 6, characterized in that, The initial feature set includes a first initial feature set, a second initial feature set, a third initial feature set, a fourth initial feature set, a fifth initial feature set, and a sixth initial feature set; The first initial feature set includes age, gender, height, weight, intraocular pressure, spherical refractive power, cylindrical refractive power and / or the axis of cylindrical refractive power; The second initial feature set includes ICL type, ICL size, ICL spherical power, ICL cylindrical power and / or the axis of ICL cylindrical power; The third initial feature set includes axial length, anterior chamber depth, anterior chamber volume, white-to-white distance, pupil diameter, corneal diameter, central corneal thickness, corneal curvature K1, the axis of corneal curvature K1, corneal curvature K2 and / or the axis of corneal curvature K2; The fourth initial feature set includes endothelial cell density, endothelial cell coefficient of variation, and / or percentage of hexagonal cells; The fifth initial feature set includes non-invasive tear river height and / or tear film breakup time; The sixth initial feature set includes uncorrected visual acuity and / or best-corrected visual acuity.
8. The prediction method according to claim 5, characterized in that, The prediction method further includes: The fourth feature set and / or the fifth feature set in the initial feature set used to filter the target feature set are removed to obtain the verification feature set; A trained machine learning-based verification model is obtained based on the verification feature set. The performance metrics of the prediction model and the validation model are compared on the same test set to verify the effectiveness of the fourth feature set and / or the fifth feature set in predicting the postoperative information.
9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed, implements the prediction method as described in any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the prediction method as described in any one of claims 1 to 8.