A corneal biomechanics-based refractive surgery type intelligent recommendation method

Abstract

This invention discloses an intelligent recommendation and optimization method for refractive surgery types based on corneal biomechanics, relating to the fields of intelligent medical decision-making and ophthalmic refractive surgery planning. After collecting preoperative feature sets of patients, this invention outputs a range of potential surgical procedures through a preliminary screening model that integrates a rule engine and LightGBM. By dividing the historical feature sets of four surgical procedures—ICL, SMILE, LASIK, and SURFACE—a two-stage classification and prediction framework is constructed. First, it distinguishes between ICL and corneal refractive surgery, and then further subdivides the corneal refractive surgery types. During training, optimization models are employed using five-fold cross-validation and SMOTE to output recommended surgical procedures. An interpretability force graph is drawn based on SHAP values ​​to improve the accuracy and safety of surgical procedure recommendations. This invention is the first to systematically incorporate corneal biomechanical parameters into a refractive surgery prediction model, achieving intelligent recommendation within a unified framework for multiple surgical procedures. It provides a reliable auxiliary decision-making basis for refractive surgery planning and has broad application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical decision-making technology, specifically to an intelligent recommendation and optimization method for refractive surgery types based on corneal biomechanics, which combines CorvisST corneal biomechanical parameters, Pentacam corneal topography, clinical demographic data, and machine learning technology. Background Technology

[0002] With the global population suffering from myopia continuing to grow, refractive surgery has become one of the preferred vision correction options in clinical practice due to its ability to quickly and effectively correct refractive errors such as myopia, hyperopia, and astigmatism. Currently, a variety of surgical procedures coexist in clinical settings, including ICL, SMILE, LASIK, and SURFACE. These different procedures differ significantly in their applicable populations, surgical principles, postoperative recovery, and long-term safety. The core difference lies essentially in the way they intervene in and the extent of their impact on the corneal biomechanical microenvironment. For example, ICL does not ablate corneal tissue and therefore has no significant impact on corneal biomechanical stability, making it suitable for patients with high myopia, thin corneas, and low corneal elastic modulus. SMILE, on the other hand, ablates the corneal stroma through a minimally invasive incision, causing less alteration to the distribution of corneal shear stress. Postoperative corneal hysteresis (CH) and corneal resistance factor (CRF) decrease less than with LASIK, making it more popular among patients with low to moderate myopia due to its superior biomechanical preservation.

[0003] However, in clinical practice, surgical recommendations traditionally rely on the doctor's experience and judgment, which are greatly influenced by subjective factors. Inadequate assessment of individual corneal biomechanical phenotypes can lead to inappropriate surgical procedures. Such mismatches can directly disrupt the normal mechanical balance of the cornea, causing a series of complications: such as excessive corneal ablation leading to decreased corneal stiffness, making it unable to withstand the continuous compressive stress generated by intraocular pressure, thereby inducing corneal ectasia; or the surgery altering the mechanical properties of the corneal tear film distribution area, leading to decreased tear film stability and causing postoperative dry eye; in addition, non-uniform changes in the mechanical properties of the corneal optical zone can also lead to increased wavefront aberration, resulting in decreased visual quality and seriously affecting surgical outcomes and patient prognosis.

[0004] Crucially, existing technologies for utilizing corneal biomechanical characteristics are largely limited to preoperative contraindication screening, failing to fully consider the varying impacts of different surgical procedures on corneal stress transmission pathways and mechanical stability reserves. This results in a technological gap in identifying surgical procedure risk boundaries and providing precise recommendations. While some technological advancements have emerged in the field regarding biomechanical interventions during surgery, there is still a lack of relevant technological achievements on how to integrate multi-dimensional biomechanical parameters to achieve intelligent and efficient surgical procedure recommendations for different patients.

[0005] Against this backdrop, there is an urgent need to develop an intelligent recommendation and optimization method for refractive surgery types that can integrate corneal biomechanical parameters, corneal morphological parameters, refractive status, and systemic biomechanical factors, while taking into account safety, accuracy, and interpretability, so as to provide reliable technical support for the formulation of personalized refractive surgery plans. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses an intelligent recommendation and optimization method for refractive surgery types based on corneal biomechanics, specifically including:

[0007] Clinical patient data for refractive surgery were collected to form a preoperative characteristic set for the patients;

[0008] Based on historical preoperative feature sets, a targeted preliminary screening model is constructed. The patient's preoperative feature set is input into the preliminary screening model, and the potential range of surgical procedures is output.

[0009] Based on the historical feature sets of four surgical procedures—ICL, SMILE, LASIK, and SURFACE—divided into potential surgical procedure ranges, a two-stage classification and prediction framework is constructed.

[0010] Based on the potential range of surgical procedures and the patient's preoperative feature set, a two-stage classification prediction framework is used to efficiently predict surgical procedures, output recommended surgical types, and simultaneously generate interpretability force maps.

[0011] Preferably, the targeted preliminary screening model is specifically: based on historical clinical patient data screened by clinical standards and expert consensus, the model is trained through a hybrid architecture of rule engine and lightweight machine learning to accurately remove samples that do not meet clinical safety standards and to initially delineate surgical procedure adaptation boundaries.

[0012] Preferably, the historical clinical patient data screened based on clinical standards and expert consensus specifically includes features covering three dimensions: CorvisST corneal biomechanical characteristics, Pentacam corneal topography data, and demographic characteristics.

[0013] The data criteria for screening based on clinical standards and expert consensus include at least age, recent increase in myopia refractive error, thickness of the central residual stromal bed after expected lens removal in SMILE surgery, intraocular pressure, estimated thickness of the thinnest point of the entire cornea after surgery, anterior chamber depth, central corneal thickness, and severity of dry eye in ICL implantation.

[0014] Preferably, the model training using a hybrid architecture combining a rule engine and a lightweight machine learning model specifically involves: cleaning historical data to remove invalid data caused by equipment failure or recording errors; imputing a small amount of missing data with clinical reference ranges using the clinical mean imputation method; and labeling data based on the data conditions screened according to clinical standards and expert consensus.

[0015] The rule engine transforms the screening criteria into Boolean logic expressions, builds a rule base, and converts the data annotation results into Boolean values. Based on the Boolean values, it initially filters the range of fuzzy potential applicable techniques.

[0016] A lightweight machine learning model is trained using the gradient boosting decision tree LightGBM algorithm, and the corrected range of potential applicable techniques is output.

[0017] Preferably, the historical feature sets of the four surgical procedures ICL, SMILE, LASIK and SURFACE are specifically: four independent dataset containers are created, and each dataset has a preset sample identification field, preoperative feature field, screening annotation field and surgical procedure label field;

[0018] The sample identification field includes the patient's unique ID, data collection time, and screening pass indicator; the preoperative feature field includes at least 48 preoperative features; the screening annotation field includes the potential range of suitable surgical procedures, the achievement status of key screening features, and lightweight model correction records; and the surgical procedure label field includes the target surgical procedure label.

[0019] By traversing the historical preoperative feature set and preprocessing it, the sample attribution is determined one by one, and historical feature sets for the four surgical procedures are constructed.

[0020] Preferably, the construction of the two-stage classification prediction framework specifically involves: inputting the potential range of surgical procedures after preliminary screening and the historical feature sets of four surgical procedures, and training the first-stage preparatory model and the second-stage preparatory model using six algorithms: decision tree, random forest, linear regression, support vector machine, multilayer perceptron and XGBoost.

[0021] The first-stage model is used to perform binary classification predictions for ICL implantation and corneal refractive surgery.

[0022] The second-stage model was used to further differentiate between SMILE, LASIK, and SURFACE within the corneal refractive surgery subset.

[0023] During training, five-fold cross-validation was used for hyperparameter tuning at each stage, and SMOTE was used for class balancing in each fold.

[0024] The performance of six preliminary models in each of the two stages was verified, and the best model was selected to form a two-stage classification prediction framework.

[0025] Preferably, the binary classification prediction of ICL implantation and corneal refractive surgery specifically involves: extracting all samples from the historical feature set of ICL procedures and constructing a positive sample set labeled as ICL; merging all samples from the historical feature sets of SMILE, LASIK, and SURFACE procedures and constructing a negative sample set uniformly labeled as corneal refractive surgery.

[0026] The positive and negative sample sets are combined into the first-stage total dataset. After extracting the independent test set from the total dataset, the remaining data is divided into the first-stage dataset, and further divided into training set and validation set.

[0027] The training of six algorithms—decision tree, random forest, linear regression, support vector machine, multilayer perceptron, and XGBoost—is based on a unified training strategy to ensure the standardization of the training process.

[0028] The training process involves hyperparameter tuning through five-fold cross-validation. In each fold, SMOTE technology is used to oversample the class with a smaller sample size to generate synthetic samples, balancing the proportion of ICL and corneal refractive surgery samples to avoid the model being biased towards the majority class. The optimized hyperparameters are then substituted into each algorithm, and the model is trained based on the first-stage training set. The loss curves and validation set performance are recorded during the training process.

[0029] Preferably, the further distinction between SMILE, LASIK, and SURFACE specifically involves: dividing the negative sample set of corneal refractive surgery into a training set and a validation set; and, based on a unified training strategy, using five-fold cross-validation combined with grid search to optimize the multi-class accuracy (Micro-ACC) and the recall rate of each class, taking into account the task characteristics of the three classifications and the internal subdivisions of corneal refractive surgery.

[0030] Preferably, the verification of the performance of each of the six models in the two stages, and the selection of the optimal model, specifically involves: inputting the independent test sets of the first and second stages into the six models trained in the corresponding stages, and calculating the overall test accuracy by weighted multiplication, using the following formula:

[0031]

[0032] in, ; Indicates the first The proportion of similar surgeries in the real sample; Indicates the category number Verification accuracy at each decision-making stage; This indicates the number of decision-making stages required for this category.

[0033] Preferably, the process of drawing the interpretability map specifically involves: selecting a first-stage SHAP interpreter based on the optimal model type of the first stage, inputting the features of the new sample into the interpreter, generating a SHAP value corresponding to each feature, where a positive SHAP value indicates that the feature drives the model to predict in the direction of ICL, and a negative SHAP value indicates that it drives the model to predict in the direction of corneal refractive surgery;

[0034] Based on the three classifications in the second stage, the feature SHAP value corresponding to each category is automatically generated. For the target procedure with the highest prediction probability, the corresponding feature SHAP values ​​are extracted to determine the feature types that positively promote and negatively inhibit the recommendation of the procedure.

[0035] Centered on the baseline value, the SHAP values ​​of each key feature are displayed horizontally, and a SHAP force-directed diagram is plotted using the following formula:

[0036]

[0037] in, The model's target prediction results for the new samples. As the baseline value, The number of key features used in model training. For the first The contribution of key features to the prediction results.

[0038] Compared with the prior art, the technical solution of this application has the following technical effects:

[0039] This invention is the first to incorporate corneal biomechanical parameters into a refractive surgery prediction model, and combines multi-source feature fusion of corneal topography and refractive status information to achieve intelligent recommendation under a unified framework for multiple procedures, providing a reliable auxiliary decision-making basis for refractive surgery planning.

[0040] This invention employs a comprehensive framework combining preliminary screening with two-stage classification prediction, covering four mainstream refractive surgery procedures: ICL, SMILE, LASIK, and SURFACE. Simultaneously, the preliminary screening model accurately defines the potential range of surgical procedures, and the two-stage classification progressively refines the procedure types, achieving a progressive recommendation from safe exclusion to precise matching. This significantly improves the comprehensiveness and accuracy of multi-procedure adaptation in complex clinical scenarios, avoiding adaptation bias caused by experience-based recommendations.

[0041] This invention uses a hybrid architecture of rule engine and lightweight machine learning to achieve rigid constraints of clinical standards and flexible fitting of data features, ensuring that the recommendation results not only conform to expert consensus but also adapt to individual physiological differences, thereby reducing the risk of postoperative complications from the source.

[0042] This invention aims to intuitively verify the rationality of recommendation results and quickly identify key influencing factors. This not only improves the credibility of the recommendation plan but also provides data support for preoperative communication and plan adjustment, promoting the upgrade of intelligent recommendation technology from auxiliary calculation to clinical collaboration and enhancing the clinical applicability of the technology.

[0043] This invention establishes a dynamic optimization mechanism with screening accuracy and surgical procedure compatibility prediction accuracy as its core, and regularly updates model parameters, solving the problem that existing solutions are difficult to adapt to the updating of clinical data and the iteration of diagnostic and treatment standards, thus ensuring the long-term applicability of the technology.

[0044] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0045] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0047] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0048] Figure 1 This is a flowchart of an intelligent recommendation and optimization method for refractive surgery types based on corneal biomechanics.

[0049] Figure 2 This is a diagram illustrating the overall architecture of an intelligent recommendation and optimization method for refractive surgery types based on corneal biomechanics.

[0050] Figure 3 A structural diagram of a preliminary screening model for refractive surgery;

[0051] Figure 4 This is an architecture diagram of the first stage of a two-stage classification prediction framework.

[0052] Figure 5An architecture diagram of the second-stage function of a two-stage classification prediction framework;

[0053] Figure 6 Experimental architecture diagram for using this method to intelligently recommend refractive surgery types in an ophthalmology hospital;

[0054] Figure 7 To enhance the interpretability of SHAP generated based on patient data Figure 1 ;

[0055] Figure 8 To enhance the interpretability of SHAP generated based on patient data Figure 2 . Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0057] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0058] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0059] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0060] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0061] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0062] Example 1 mainly describes an intelligent recommendation and optimization method for refractive surgery types based on corneal biomechanics, such as... Figure 1 , Figure 2 As shown, it specifically includes:

[0063] Clinical patient data for refractive surgery were collected to form a preoperative characteristic set for the patients;

[0064] Based on historical preoperative feature sets, a targeted preliminary screening model is constructed. The patient's preoperative feature set is input into the preliminary screening model, and the potential range of surgical procedures is output.

[0065] Based on the historical feature sets of four surgical procedures—ICL, SMILE, LASIK, and SURFACE—divided into potential surgical procedure ranges, a two-stage classification and prediction framework is constructed.

[0066] Based on the potential range of surgical procedures and the patient's preoperative feature set, a two-stage classification prediction framework is used to efficiently predict surgical procedures, output recommended surgical types, and simultaneously generate interpretability force maps.

[0067] Furthermore, the exclusion criteria for clinical patient data should include at least the following:

[0068] ① Corneal diseases (such as keratoconus and corneal ectasia);

[0069] ②Systemic or metabolic diseases (such as severe hyperthyroidism, diabetic retinopathy);

[0070] ③ Active eye diseases (such as endophthalmitis, corneal infection);

[0071] ④ Severe abnormalities of the ocular adnexa and damage to fundus function (such as retinal detachment);

[0072] ⑤ Patients with cataracts that have already affected their vision or diagnosed with glaucoma, as well as patients with mental illnesses who cannot understand or cooperate with surgery.

[0073] Furthermore, preoperative characteristics should include at least the following features covering three dimensions: CorvisST corneal biomechanical properties, Pentacam corneal topography data, and demographic characteristics:

[0074] bIOP,PRFI,SSI,SP-A1,IR,ARTh,DA-Ratio,CBI,BAD-D,TBI,FrontK1,FrontK2,FrontKmean,FrontK1Axis,BackK1,BackK2,B ackKmean,BackAxisK1,TCT,CV,ACV,ACA,ACD,FrontKmax,QValue,AnteriorSurfaceElevationattheCornealThinnestPoint ,PosteriorSurfaceElevationattheCornealThinnestPoint,Df,Db,Dp,Dt,Da,PupilCenterX,PupilCenterY,AngleKappa,CD,ISValue,AveragePachymetricProgressionRate,Gender,Age,Ni-BUT,SPH,CYL,Axis,SE,UDVA,CDVA,DarkPupilDiameter

[0075] Furthermore, the targeted preliminary screening model is specifically designed as follows: based on historical clinical patient data screened according to clinical standards and expert consensus, the model is trained using a hybrid architecture that combines a rule engine with lightweight machine learning to accurately remove samples that do not meet clinical safety standards and to initially delineate the surgical procedure adaptation boundaries.

[0076] Furthermore, screening criteria based on clinical standards and expert consensus should include at least the following:

[0077] ① The age should be greater than 18 years old;

[0078] ② In the past two years, the annual increase in myopia refractive error has not exceeded 0.5D;

[0079] ③ For SMILE surgery, the thickness of the residual stromal bed in the center of the cornea after lenticule removal should be at least 280 μm.

[0080] ④ Intraocular pressure is within the normal range of 10-21 mmHg;

[0081] ⑤ If the estimated thinnest point thickness of the entire cornea after surgery is less than 380μm, then surface surgery will not be considered;

[0082] ⑥ For ICL implantation, the anterior chamber depth should be at least greater than 2.8 mm;

[0083] ⑦ If the central corneal thickness is less than 450 μm or the expected remaining central corneal stroma thickness under the corneal flap after ablation is less than 280 μm, or the expected remaining central corneal stroma thickness after surgery is less than 50% of the preoperative corneal thickness, then corneal refractive surgery will not be considered.

[0084] ⑧ Patients with severe dry eye should avoid corneal refractive surgery.

[0085] Furthermore, based on historical clinical patient data screened according to clinical standards and expert consensus, the data specifically includes features covering three dimensions: CorvisST corneal biomechanical characteristics, Pentacam corneal topography data, and demographic characteristics.

[0086] The data criteria for screening based on clinical standards and expert consensus include at least age, recent increase in myopia refractive error, thickness of the central residual stromal bed after expected lens removal in SMILE surgery, intraocular pressure, estimated thickness of the thinnest point of the entire cornea after surgery, anterior chamber depth, central corneal thickness, and severity of dry eye in ICL implantation.

[0087] Furthermore, such as Figure 3 The architecture diagram of the refractive surgery-targeted preliminary screening model shown is used for model training through a hybrid architecture combining a rule engine and a lightweight machine learning model. Specifically, the model involves: cleaning historical data to remove invalid data caused by equipment failure or recording errors; imputing a small amount of missing data with clinical reference ranges using the clinical mean imputation method; and labeling data based on the data conditions for screening based on clinical standards and expert consensus.

[0088] The rule engine transforms the screening criteria into Boolean logic expressions, builds a rule base, and converts the data annotation results into Boolean values. Based on the Boolean values, it initially filters the range of fuzzy potential applicable techniques.

[0089] A lightweight machine learning model is trained using the gradient boosting decision tree LightGBM algorithm, and the corrected range of potential applicable techniques is output.

[0090] Furthermore, the historical feature sets for the four surgical procedures—ICL, SMILE, LASIK, and SURFACE—are specifically designed as follows: four independent dataset containers are created, with each dataset having a preset sample identification field, preoperative feature field, screening annotation field, and surgical procedure label field.

[0091] The sample identification field includes the patient's unique ID, data collection time, and screening pass indicator; the preoperative feature field includes at least 48 preoperative features; the screening annotation field includes the potential range of suitable surgical procedures, the achievement status of key screening features, and lightweight model correction records; and the surgical procedure label field includes the target surgical procedure label.

[0092] By traversing the historical preoperative feature set and preprocessing it, the sample attribution is determined one by one, and historical feature sets for the four surgical procedures are constructed.

[0093] Furthermore, the two-stage classification prediction framework is constructed as follows: inputting the potential range of surgical procedures after preliminary screening and the historical feature sets of four surgical procedures, and training the first-stage preparatory model and the second-stage preparatory model through six algorithms: decision tree, random forest, linear regression, support vector machine, multilayer perceptron and XGBoost.

[0094] The first-stage model is used to perform binary classification predictions for ICL implantation and corneal refractive surgery.

[0095] The second-stage model was used to further differentiate between SMILE, LASIK, and SURFACE within the corneal refractive surgery subset.

[0096] During training, five-fold cross-validation was used for hyperparameter tuning at each stage, and SMOTE was used for class balancing in each fold.

[0097] The performance of six preliminary models in each of the two stages was verified, and the best model was selected to form a two-stage classification prediction framework.

[0098] Furthermore, such as Figure 4 The architecture diagram of the first stage of the two-stage classification prediction framework shown is as follows: the binary classification prediction of ICL implantation and corneal refractive surgery is as follows: extract all samples from the historical feature set of ICL procedures and construct a positive sample set labeled as ICL; merge all samples from the historical feature sets of SMILE, LASIK, and SURFACE procedures and construct a negative sample set uniformly labeled as corneal refractive surgery.

[0099] The positive and negative sample sets are combined into the first-stage total dataset. After extracting the independent test set from the total dataset, the remaining data is divided into the first-stage dataset, and further divided into training set and validation set.

[0100] The training of six algorithms—decision tree, random forest, linear regression, support vector machine, multilayer perceptron, and XGBoost—is based on a unified training strategy to ensure the standardization of the training process.

[0101] The training process involves hyperparameter tuning through five-fold cross-validation. In each fold, SMOTE technology is used to oversample the class with a smaller sample size to generate synthetic samples, balancing the proportion of ICL and corneal refractive surgery samples to avoid the model being biased towards the majority class. The optimized hyperparameters are then substituted into each algorithm, and the model is trained based on the first-stage training set. The loss curves and validation set performance are recorded during the training process.

[0102] Furthermore, such as Figure 5 The architecture diagram of the second stage of the two-stage classification prediction framework shown further distinguishes between SMILE, LASIK, and SURFACE. Specifically, the negative sample set of corneal refractive surgery is divided into a training set and a validation set. Based on the task characteristics of the three classifications and the internal subdivisions of corneal refractive surgery, and on the basis of a unified training strategy, hyperparameters are tuned by combining five-fold cross-validation with grid search, with the optimization objectives being the multi-class accuracy (Micro-ACC) and the recall rate of each class.

[0103] Furthermore, to verify the performance of the six models in each of the two stages, the optimal model was selected by inputting the independent test sets of the first and second stages into the six models trained in the corresponding stages, and calculating the overall test accuracy by weighted multiplication, as shown in the formula:

[0104]

[0105] in, ; Indicates the first The proportion of similar surgeries in the real sample; Indicates the category number Verification accuracy at each decision-making stage; This indicates the number of decision-making stages required for this category.

[0106] Furthermore, the interpretationability map is drawn as follows: based on the optimal model type in the first stage, the first-stage SHAP interpreter is selected, the features of the new sample are input into the interpreter, and the SHAP value corresponding to each feature is generated. A positive SHAP value indicates that the feature pushes the model to predict in the direction of ICL, and a negative value indicates that it pushes the model to predict in the direction of corneal refractive surgery.

[0107] Based on the three classifications in the second stage, the feature SHAP value corresponding to each category is automatically generated. For the target procedure with the highest prediction probability, the corresponding feature SHAP values ​​are extracted to determine the feature types that positively promote and negatively inhibit the recommendation of the procedure.

[0108] Centered on the baseline value, the SHAP values ​​of each key feature are displayed horizontally, and a SHAP force-directed diagram is plotted using the following formula:

[0109]

[0110] in, The model's target prediction results for the new samples. As the baseline value, The number of key features used in model training. For the first The contribution of key features to the prediction results.

[0111] Furthermore, the testing process for the results of training six algorithms on independent test sets is as follows: if the result of the first-stage test is ICL surgery, the test is terminated; if the result of the first-stage test is laser surgery, the test proceeds to the second stage, and the results are used to calculate the test accuracy with the test set labels.

[0112] This embodiment details an intelligent recommendation and optimization method for refractive surgery types based on corneal biomechanics. After collecting the patient's preoperative feature set, a preliminary screening model that integrates a rule engine and LightGBM is used to output the range of potential surgical procedures. By dividing the historical feature sets of four surgical procedures—ICL, SMILE, LASIK, and SURFACE—a two-stage classification and prediction framework is constructed. First, ICL is distinguished from corneal refractive surgery, and then the corneal refractive surgery type is further subdivided. During training, the model is optimized using five-fold cross-validation, SMOTE, and other methods to output recommended surgical procedures. An interpretability force graph is drawn based on the SHAP value to improve the accuracy and safety of surgical procedure recommendations.

[0113] Example 2, based on Example 1, describes in detail an experiment using this method for intelligent recommendation in an ophthalmology hospital, such as... Figure 6 The experimental architecture diagram shown below illustrates the specific process as follows:

[0114] This study selected patient data from those who successfully underwent one of the following four surgeries at the ophthalmology hospital between October 2023 and November 2024: ICL implantation, superficial corneal refractive surgery (including tPRK, PRK, and LASEK), SMILE, or LASIK. The data of the patients were ensured to be well preserved, and the patients were aged 18 to 45 years and had undergone a complete and rigorous ophthalmological examination before the surgery.

[0115] After excluding patients with corneal diseases (such as keratoconus, corneal ectasia), systemic or metabolic diseases (such as severe hyperthyroidism, diabetic retinopathy), active eye diseases (such as endophthalmitis, corneal infection), severe abnormalities of ocular adnexa, fundus dysfunction (such as retinal detachment), or cataracts or diagnosed glaucoma that have already affected vision, as well as patients with mental illnesses who cannot understand or cooperate with surgery, the surgical categories of the image data were labeled with reference to the patient's medical records to construct a historical feature set.

[0116] Based on historical feature set data, a preliminary screening model is trained. The collected data is initially screened according to the preset clinical safety standards for refractive surgery, and abnormal samples that do not meet the requirements of surgical procedure planning are removed, while ensuring the completeness and consistency of the records.

[0117] The screened data undergoes data preprocessing and feature engineering, and the dataset is then partitioned. The specific steps are as follows:

[0118] First, the missing data is cleaned. Based on the refractive surgery examination standards and physician judgment, missing entries due to equipment errors or lack of diagnostic significance are removed, and a limited number of missing data with clear clinical scope are filled in.

[0119] Secondly, outlier screening based on Z-score is employed, and all continuous variables are standardized with a threshold set to 5. Specifically, for each numerical feature χi, its standardized score is calculated:

[0120]

[0121] in The mean of this feature. Let be the standard deviation. If the sample satisfies on any feature If the value is >5, the sample is considered to deviate too much from the overall distribution in that dimension and is therefore considered an outlier and removed.

[0122] Secondly, Z-score standardization is applied to numerical variables to eliminate differences in feature dimensions;

[0123] Subsequently, one-hot encoding is performed on the categorical variables to ensure that the classification targets are discretized without association, and to prevent the model from misinterpreting the numerical relationships between categories;

[0124] Next, the features were jointly screened using two methods: Pearson correlation coefficient with a threshold of 0.95 and variance filtering with a threshold of 0.015. A total of 10 features that were weakly correlated with the discriminant method or strongly correlated with other variables were removed, including SE, UDVA, CDVA, FrontKmean, BackKmean, FrontKmax, PupilCenterX, PupilCenterY, Kappa, and Avg.

[0125] Finally, 38 features were retained for subsequent dataset partitioning and training.

[0126] A two-stage classification prediction framework for four surgical procedures—ICL, SMILE, LASIK, and SURFACE—is constructed. The specific steps are as follows: The first stage performs binary classification prediction for ICL implantation and corneal refractive surgery; the second stage further distinguishes between SMILE, LASIK, and SURFACE within the corneal refractive surgery subset.

[0127] After completing all model construction, the data of 50 patients who underwent surgery at the hospital were used to make intelligent recommendations using this method. The results of the intelligent recommendations were compared with those of expert analysis. The intelligent recommendations were consistent with the expert analysis results in 46 cases, while they were inconsistent in 4 cases, with an overall consistency of 92%, which verified the high accuracy of the intelligent prediction of surgical procedures by this method.

[0128] In the cases of inconsistency, there were 2 cases where the intelligent recommendation was SURFACE but the expert recommended LASIK, 1 case where the intelligent recommendation was ICL but the expert recommended LASIK, and 1 case where the intelligent recommendation was SMILE but the expert recommended SURFACE. It can be seen that the main focus was on the distinction between SMILE, LASIK and SURFACE within the corneal refractive surgery.

[0129] Generate SHAP-based model interpretability maps from patient data, through methods such as... Figure 7 The first-stage SHAP interpretability plot for a patient is shown in the figure. It is clear from the figure that the patient's CBI=0.01, gender=1.0, BAD-D=1.08, and spherical power=-1.25 are the data with the greatest influence weight in the first stage, which pushes the first-stage recommendation for this patient in a positive direction (ICL surgery type).

[0130] like Figure 8 The diagram shows the second-stage SHAP interpretability plot for a patient. It is clearly visible that the data with the largest influence weight in the second-stage SHAP plot for this patient is the DA-Ratio. spherical diopter , height of the thinnest part of the back surface Cylinder power The thinnest thickness of the cornea , We obtained f(x) = 1.47, which led the second-stage recommendation for the patient to the target surgical procedure with the highest initial predicted probability, thus achieving highly interpretable visualization of the intelligent recommendation results.

[0131] This embodiment details an experiment using this method for intelligent surgical procedure recommendation in an ophthalmology hospital. The experimental results fully verify the high accuracy of the intelligent surgical procedure recommendation and the high interpretability of the generated force map, demonstrating the effectiveness and clinical applicability of this method in promoting the upgrade of intelligent recommendation technology from assisted computation to clinical collaboration.

[0132] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for intelligent recommendation and optimization of refractive surgery types based on corneal biomechanics, characterized in that, include: Clinical patient data for refractive surgery were collected to form a preoperative characteristic set for the patients; Based on historical preoperative feature sets, a targeted preliminary screening model is constructed. The patient's preoperative feature set is input into the preliminary screening model, and the potential range of surgical procedures is output. Based on the historical feature sets of four surgical procedures—ICL, SMILE, LASIK, and SURFACE—divided into potential surgical procedure ranges, a two-stage classification and prediction framework is constructed. Based on the potential range of surgical procedures and the patient's preoperative feature set, a two-stage classification prediction framework is used to efficiently predict surgical procedures, output recommended surgical types, and simultaneously generate interpretability force maps.

2. The method according to claim 1, characterized in that, The targeted preliminary screening model is specifically designed to: train a model based on historical clinical patient data screened according to clinical standards and expert consensus, using a hybrid architecture combining a rule engine and lightweight machine learning, to accurately remove samples that do not meet clinical safety standards and to preliminarily define the surgical procedure adaptation boundaries.

3. The method according to claim 2, characterized in that, The historical clinical patient data screened based on clinical standards and expert consensus specifically includes features covering three dimensions: CorvisST corneal biomechanical characteristics, Pentacam corneal topography data, and demographic characteristics. The data criteria for screening based on clinical standards and expert consensus include at least age, recent increase in myopia refractive error, thickness of the central residual stromal bed after expected lens removal in SMILE surgery, intraocular pressure, estimated thickness of the thinnest point of the entire cornea after surgery, anterior chamber depth, central corneal thickness, and severity of dry eye in ICL implantation.

4. The method according to claim 2, characterized in that, The model training using a hybrid architecture combining a rule engine and a lightweight machine learning model specifically involves: cleaning historical data to remove invalid data caused by equipment malfunctions or recording errors; imputing missing data with a small clinical reference range using the clinical mean imputation method; and labeling data based on the data conditions screened according to clinical standards and expert consensus. The rule engine transforms the screening criteria into Boolean logic expressions, builds a rule base, and converts the data annotation results into Boolean values. Based on the Boolean values, it initially filters the range of fuzzy potential applicable techniques. A lightweight machine learning model is trained using the gradient boosting decision tree LightGBM algorithm, and the corrected range of potential applicable techniques is output.

5. The method according to claim 1, characterized in that, The historical feature sets of the four surgical procedures, ICL, SMILE, LASIK, and SURFACE, are specifically as follows: four independent dataset containers are created, and each dataset has a preset sample identification field, preoperative feature field, screening annotation field, and surgical procedure label field; The sample identification field includes the patient's unique ID, data collection time, and screening pass indicator; the preoperative feature field includes at least 48 preoperative features. The screening annotation field includes the potential range of suitable surgical procedures, the achievement status of key screening features, and lightweight model correction records; the surgical procedure label field includes the target surgical procedure label. By traversing the historical preoperative feature set and preprocessing it, the sample attribution is determined one by one, and historical feature sets for the four surgical procedures are constructed.

6. The method according to claim 1, characterized in that, The construction of the two-stage classification prediction framework is as follows: input the potential range of surgical procedures after preliminary screening and the historical feature sets of four surgical procedures, and train the first-stage preparatory model and the second-stage preparatory model through six algorithms: decision tree, random forest, linear regression, support vector machine, multilayer perceptron and XGBoost. The first-stage model is used to perform binary classification predictions for ICL implantation and corneal refractive surgery. The second-stage model was used to further differentiate between SMILE, LASIK, and SURFACE within the corneal refractive surgery subset. During training, five-fold cross-validation was used for hyperparameter tuning at each stage, and SMOTE was used for class balancing in each fold. The performance of six preliminary models in each of the two stages was verified, and the best model was selected to form a two-stage classification prediction framework.

7. The method according to claim 6, characterized in that, The binary classification prediction of ICL implantation and corneal refractive surgery is specifically as follows: extract all samples from the historical feature set of ICL procedures and construct a positive sample set labeled as ICL; merge all samples from the historical feature sets of SMILE, LASIK, and SURFACE procedures and construct a negative sample set uniformly labeled as corneal refractive surgery. The positive and negative sample sets are combined into the first-stage total dataset. After extracting the independent test set from the total dataset, the remaining data is divided into the first-stage dataset, and further divided into training set and validation set. The training of six algorithms—decision tree, random forest, linear regression, support vector machine, multilayer perceptron, and XGBoost—is based on a unified training strategy to ensure the standardization of the training process. The training process involves hyperparameter tuning through five-fold cross-validation. In each fold, SMOTE technology is used to oversample the class with a smaller sample size to generate synthetic samples, balancing the proportion of ICL and corneal refractive surgery samples to avoid the model being biased towards the majority class. The optimized hyperparameters are then substituted into each algorithm, and the model is trained based on the first-stage training set. The loss curves and validation set performance are recorded during the training process.

8. The method according to claim 6, characterized in that, The further distinction between SMILE, LASIK, and SURFACE is as follows: the negative sample set of corneal refractive surgery is divided into a training set and a validation set. Based on the task characteristics of the three classifications and the internal subdivisions of corneal refractive surgery, and on the basis of a unified training strategy, hyperparameters are tuned by combining five-fold cross-validation with grid search, with the optimization objectives being the multi-class accuracy (Micro-ACC) and the recall rate of each class.

9. The method according to claim 6, characterized in that, The performance of each of the six models in the two phases is verified, and the optimal model is selected. Specifically, the independent test sets are input into the six models corresponding to the first and second phases, respectively, and the overall test accuracy is calculated by weighted multiplication, using the following formula: in, ; Indicates the first The proportion of similar surgeries in the real sample; Indicates the category number Verification accuracy at each decision-making stage; This indicates the number of decision-making stages required for this category.

10. The method according to claim 1, characterized in that, The process of drawing an interpretability map is as follows: Based on the optimal model type in the first stage, select the first-stage SHAP interpreter, input the features of the new sample into the interpreter, and generate the SHAP value corresponding to each feature. A positive SHAP value indicates that the feature drives the model to predict in the direction of ICL, and a negative SHAP value indicates that it drives the model to predict in the direction of corneal refractive surgery. Based on the three classifications in the second stage, the feature SHAP value corresponding to each category is automatically generated. For the target procedure with the highest prediction probability, the corresponding feature SHAP values ​​are extracted to determine the feature types that positively promote and negatively inhibit the recommendation of the procedure. Centered on the baseline value, the SHAP values ​​of each key feature are displayed horizontally, and a SHAP force-directed diagram is plotted using the following formula: in, The model's target prediction results for the new samples. As the baseline value, The number of key features used in model training. For the first The contribution of key features to the prediction results.

Images