Myopia surgery plan generation method, generation system and storage medium

Abstract

The application discloses a myopia surgery scheme generation method, a generation system and a storage medium, and comprises the following steps: collecting historical surgery data, wherein the historical surgery data comprises preoperative data, intraoperative data and postoperative data of a plurality of patients, the preoperative data comprises eye parameters, living habit data, previous medical history data and patient basic information; preprocessing the historical surgery data to obtain a training set; constructing a dynamic Bayesian network model based on the training set; collecting preoperative data of a patient to be diagnosed, inputting the dynamic Bayesian network model, and generating a myopia surgery scheme of the patient to be diagnosed based on a surgery scheme library. Through multi-modal data fusion to construct a dynamic Bayesian network, preoperative individualized scheme recommendation, intraoperative real-time parameter optimization and postoperative risk tracking are realized, and the surgery success rate and patient satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence medical diagnosis technology, and in particular to a method, a system, and a storage medium for generating a myopia surgery plan. Background Art

[0002] With the popularization of electronic devices and the significant increase in the intensity of eye use, myopia has become a global public health problem. Myopia surgery, as an effective means of correction, has received widespread attention. However, traditional myopia surgical plan recommendations mainly rely on the doctor's experience and limited eye parameter data, such as corneal curvature, refractive power, etc. This method has the following limitations: 1. High subjectivity: Differences in doctor's experience may lead to different surgical plans being recommended for the same patient, affecting the consistency of decision-making. 2. One-sided data: Traditional methods only focus on a few static parameters (such as refractive power, corneal thickness), ignoring dynamic and multidimensional factors such as corneal biomechanical properties, intraocular pressure, and living habits, resulting in incomplete evaluation. 3. Lack of personalization: Existing methods make it difficult to dynamically adjust surgical plans based on individual differences in patients, resulting in the implemented surgical plan may not be the best choice.

[0003] In response to the above problems, there is an urgent need for a more comprehensive, accurate and personalized myopia surgery adaptation diagnostic method to improve the success rate of surgery and patient satisfaction. Summary of the Invention

[0004] In order to overcome the defects in the prior art, the first purpose of the present invention is to provide a method for generating a myopia surgery plan, the second purpose of the present invention is to provide a system for generating a myopia surgery plan, and the third purpose of the present invention is to provide a storage medium.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] In a first aspect, a method for generating a myopia surgery plan comprises the following steps:

[0007] Collect historical surgical data, including preoperative, intraoperative, and postoperative data of the patient. The preoperative data includes eye parameters, lifestyle data, medical history data, and basic patient information.

[0008] Preprocessing the historical surgical data to obtain a training set;

[0009] Building a dynamic Bayesian network model based on the training set;

[0010] Preoperative data of the patient to be diagnosed is collected and input into the dynamic Bayesian network model. The dynamic Bayesian network model generates a myopia surgery plan for the patient to be diagnosed by calling the surgery type, parameter range and indication rules in the surgery plan library.

[0011] The method for generating myopia surgical plans described in this paper uses multi-dimensional historical surgical data and a dynamic Bayesian network (DBN) model to achieve personalized recommendations. This historical surgical data can be obtained by integrating clinical data from various medical institutions, covering the entire preoperative, intraoperative, and postoperative cycle of multiple patients. The data must include complete records for at least thousands of patients to ensure statistical significance and the reliability of model training.

[0012] The historical surgical data includes preoperative data, intraoperative data, and postoperative data, where preoperative data includes eye parameters, lifestyle data, medical history data, and basic patient information. Therefore, the historical surgical data includes data of multiple different types and sources. Eye parameters provide basic biometric support for surgical plans; lifestyle data and basic patient information can assist in assessing the causes of myopia and the impact of postoperative eye load on efficacy; and medical history data is used to avoid contraindications. The integration of intraoperative and postoperative data can reveal the dynamic correlation between surgical procedures and efficacy, enhance the causal reasoning ability of the model, and thus improve the matching accuracy between plans and individual patient characteristics.

[0013] The myopia surgery plan is a recommended plan generated based on the preoperative data of the patient to be diagnosed, which includes the types of surgery that can be performed on the patient to be diagnosed, the optimal parameter combination corresponding to each type of surgery, and the prediction of the postoperative effect.

[0014] Optionally, the preprocessing includes:

[0015] Data cleaning, used to remove outliers, missing values, and duplicate values;

[0016] Data normalization, converting data of different dimensions into a unified scale;

[0017] Data augmentation is used to expand image data and unstructured data to improve the generalization ability of the dynamic Bayesian network model.

[0018] The training efficiency and prediction accuracy of the dynamic Bayesian network model can be improved through the preprocessing process of data cleaning, data normalization and data enhancement.

[0019] Specifically, data cleaning can remove data outside a reasonable range, fill in missing patient data, and remove duplicate records to ensure the integrity and reliability of input data and prevent noise from interfering with model learning. For example, for refractive data, a reasonable range is set (e.g., -15D to +5D). Data outside this range will be considered outliers and removed.

[0020] Data normalization refers to converting data of different dimensions to the same scale. Common methods include Min-Max normalization and Z-score normalization. Min-Max normalization linearly transforms the original data using the minimum and maximum values ​​of the data to eliminate the interference of dimensional differences on model training. Min-Max normalization formula: , where Min is the minimum value of the feature in the dataset, and Max is the maximum value of the feature in the dataset. Z-score normalization (also known as standardization) is used to convert data of different dimensions and ranges into a standard normal distribution with a mean of 0 and a standard deviation of 1. Its core is to eliminate the interference of dimensional differences on model training by calculating the distance between the data point and the mean (in units of standard deviation). Z-score normalization formula: ,in, is the mean value of the feature in the dataset, The standard deviation of this feature in the dataset. For example, diopters are measured in "D" and corneal thickness is measured in "μm." Converting to a unified scale using the Min-Max or Z-score method eliminates the interference of magnitude differences on model weight updates, accelerates model convergence, and improves the stability of parameter optimization.

[0021] Data augmentation involves performing operations such as rotation, flipping, scaling, translation, and noise addition on image data. It also involves semantic expansion or synthetic generation of unstructured data to expand the size and diversity of the training set, effectively alleviating overfitting caused by insufficient data and enhancing the model's generalization capabilities for complex clinical scenarios. Image data can include corneal topography and OCT images, while unstructured data can include descriptions of lifestyle habits and medical history text.

[0022] The above preprocessing steps work together to provide high-quality, standardized input data for the dynamic Bayesian network model, thereby ensuring that the generated myopia surgery plan is both scientific and personalized.

[0023] Optionally, the network structure of the dynamic Bayesian network model includes a parent node, an intermediate node and a child node, the parent node is preoperative data, the intermediate node is the myopia surgery plan, and the child node is postoperative data. The connection relationship between the parent node, the intermediate node and the child node is based on the causal relationship defined by medical standard documents and statistical results of historical clinical databases.

[0024] The network structure of the dynamic Bayesian network model, based on medical standard documents and statistical results from historical clinical databases, clearly defines causal relationships and consists of parent nodes, intermediate nodes, and child nodes. Parent nodes represent preoperative data, including key features such as ocular parameters, lifestyle data, and medical history. Intermediate nodes correspond to surgical plans, including the type of surgery and specific parameter configurations for each type, such as the full-length femtosecond SMILE procedure or the semi-femtosecond LASIK procedure and its laser energy setting. Child nodes represent postoperative outcome indicators, covering visual recovery, complication risk, and long-term visual quality assessment. The connections between nodes strictly adhere to medical logic. For example, the preoperative feature of insufficient corneal thickness is directly linked to the LASIK exclusion rule, thereby reducing the risk of postoperative corneal ectasia. By embedding the clinical decision-making path into the model structure, the reasoning process is transparent and consistent with medical practice guidelines. This allows for a traceable mapping between model output and physician experience, thereby increasing physicians' willingness to verify and adopt the model results.

[0025] The core advantages of dynamic Bayesian networks lie in their dynamic reasoning capabilities and data robustness. Compared to traditional static models, dynamic Bayesian networks can process sudden intraoperative data in real time, such as eye movement deviations or laser parameter fluctuations. By updating the conditional probability table, they can dynamically adjust the parameter configuration in the myopia surgery plan, such as correcting the cutting center position or energy parameters, thereby improving the accuracy of the plan and intraoperative safety. In addition, dynamic Bayesian networks have strong fault tolerance for missing data. For example, some patients lack long-term follow-up records, but they can still make reasonable inferences based on the prior causal chain, ensuring practicality in complex clinical scenarios.

[0026] The method for defining causal relationships is achieved by systematically integrating medical standard documents and surgical case patterns in historical clinical databases. The medical standard documents may include ophthalmic clinical guidelines, authoritative literature, etc. The clinical guidelines may be "Expert Consensus on Transepithelial Excimer Keratectomy in China (2019)", "Expert Consensus on the Surgery Standards for Femtosecond Laser Small Incision Corneal Lens Removal in my country (2018)", and "Expert Consensus on Clinical Diagnosis and Treatment of Laser Corneal Refractive Surgery (2015)". The authoritative literature may be:

[0027] Wang Chuang. Characteristics of corneal epithelial thickness distribution and its role in the diagnosis of keratoconus[D]. Henan University, 2021.

[0028] Li Dan, Zhang Yang, Guo Lili, et al. Study on measuring corneal thickness in patients with keratoconus and myopia using Pentacam HR and Optovue iVue OCT[J]. Ophthalmology, 2021. DOI: 10.13281 / j.cnki.issn.1004-4469.2021.02.011.

[0029] Ren Yaru, Zhuang Xinyu, Kou Sasa. Measurement of corneal epithelial thickness parameters using Fourier domain optical coherence tomography for identification of early keratoconus[J]. International Journal of Ophthalmology, Vol. 22, No. 2, 2022, pp. 200-204, ISTIC PKU, 2022. DOI: 10.3980 / j.issn.1672-5123.2022.2.05.

[0030] Jiang Feng, Li Yang. Role of Fourier OCT corneal thickness parameters in preclinical keratoconus screening[J]. Ophthalmology, 2013, 22(3): 165-169.

[0031] The historical clinical database refers to a collection of standardized medical data accumulated in clinical practice related to myopia surgery.

[0032] Specifically, key logic is extracted from clinical guidelines and authoritative literature, such as the priority of SMILE surgery for patients with a history of dry eye, and converted into directed connection rules in the network. At the same time, its statistical significance is verified using historical clinical databases, such as analyzing the strength of the association between corneal thickness and keratoconus risk. When there is a deviation between the actual data and the preset causal link, the variable relationship is re-evaluated through conditional probability analysis, and the network structure is modified if necessary to make the model consistent with medical practice and adaptable to data characteristics. This process transforms abstract clinical experience into a traceable reasoning path, supporting dynamic risk prediction and real-time decision optimization while ensuring the interpretability of the model.

[0033] Optionally, constructing a dynamic Bayesian network model based on the training set includes:

[0034] Based on the training set, a conditional probability table is learned using a Bayesian estimation method;

[0035] The structured variational inference algorithm is used to optimize the parameters of the conditional probability table and obtain a dynamic Bayesian network model.

[0036] The Bayesian estimation method can learn the Conditional Probability Table (CPT) from the training set, fully leveraging the statistical laws of historical clinical databases while incorporating prior medical knowledge to avoid the overfitting problem caused by purely data-driven approaches. The structured variational inference algorithm optimizes the parameter learning process by introducing constraints (such as causal dependencies of network structures). While ensuring model interpretability, it significantly reduces the computational complexity of complex probabilistic reasoning and is particularly suitable for processing multimodal, high-dimensional clinical data. The synergy between the two enables the model to efficiently handle small sample sizes or data-missing scenarios, while dynamically adapting to real-time intraoperative data, providing doctors with real-time, reliable decision-making support and ultimately achieving accurate and personalized recommendations for surgical plans.

[0037] The conditional probability table (CPT) is a core tool for describing causal relationships between variables in dynamic Bayesian networks. It is used to quantify the conditional probability distribution of a node under different combinations of its parent node states. In the context of personalized adaptation for myopia surgery, CPT combines historical data with clinical experience to accurately depict the dynamic correlation between surgical plan selection and postoperative outcomes. The conditional probability table is a multidimensional table, with each row representing a combination of parent node states and each column representing the possible state of the current node. The values ​​in the table represent the corresponding conditional probabilities.

[0038] Structured variational inference is an efficient approximate inference method used to estimate the posterior distribution that is difficult to calculate directly in complex probabilistic models (such as dynamic Bayesian networks). Its core idea is to introduce structured constraints to design an approximate distribution with a structure similar to the true posterior distribution, thereby transforming the complex probabilistic inference problem into an optimizable parameter learning problem, thereby significantly reducing the computational complexity. Unlike traditional variational inference, which assumes that the approximate distribution is completely independent, structured variational inference allows the approximate distribution to retain some of the model's dependencies. For example, in a dynamic Bayesian network (DBN), the approximate distribution The causal relationship between nodes can be maintained (such as "corneal thickness → surgical plan selection → postoperative vision"), thus being closer to the structure of the true posterior distribution.

[0039] Optionally, the expression of the Bayesian estimation method is: ,in, Represents model parameters, such as the probability values ​​of the conditional probability table in a dynamic Bayesian network; represents observational data, i.e., preoperative, intraoperative, and postoperative data in the historical surgical dataset; Represents the prior probability, reflecting the prior knowledge of the parameters before the data is observed; is the likelihood function, which represents the probability distribution of observed data when the parameters are known; is the posterior probability, which represents the parameter update result after combining data with prior knowledge; is the marginal likelihood and serves as a normalization constant to ensure that the probabilities sum to 1.

[0040] In the present invention, the Bayesian estimation method is used to learn the conditional probability table of the dynamic Bayesian network from the historical clinical database. The specific steps include:

[0041] Prior knowledge encoding: converting clinical experience into prior probabilities ,For example, the initial probability of dry eye patients choosing PRK surgery is set to a low value;

[0042] Data-driven update: Calculating likelihood functions through historical clinical databases For example, the probability of visual recovery after SMILE surgery is calculated for patients with a corneal thickness of 480 μm.

[0043] Posterior probability calculation: integrating prior data to obtain posterior distribution , such as the probability of success of the modified SMILE procedure at a specific corneal thickness;

[0044] Dynamic optimization: During the operation, after real-time data (such as laser cutting parameters) are input, the posterior probability is dynamically updated to achieve intraoperative parameter adjustment.

[0045] The specific steps of the structured variational inference include:

[0046] Step 1: Define the structured approximate distribution , for example, maintaining the directed connection relationship between nodes in DBN.

[0047] Step 2: Optimize by gradient descent or coordinate ascent parameters, minimizing the KL divergence .

[0048] Step 3: Utilize the optimized Approximate the true posterior distribution and complete CPT parameter estimation.

[0049] Optionally, the surgical plan library includes myopia surgery types, parameter configuration ranges associated with the myopia surgery types, and indications, contraindications, and historical efficacy risk data corresponding to the myopia surgery types. The surgical types include corneal ablation, lens implantation, and other auxiliary procedures. The corneal ablation types include SMILE, LASIK, PRK, etc. The lens implantation types include ICL lens implantation, lens exchange, etc. The other auxiliary procedures include AK astigmatic keratotomy, etc. For each surgical type, the biomechanical safety parameter range defined by medical specification documents and historical clinical database statistical results is associated. For example, the laser energy range of the SMILE procedure is 0.5 mJ to 5.0 mJ, and the cutting diameter is 6.0 mm to 7.0 mm. The LASIK procedure requires the corneal flap thickness to be controlled between 90 μm and 130 μm and the residual stromal bed thickness to be no less than 250 μm. Indications and contraindications clearly define the applicable conditions for various types of surgeries, such as a specific refractive power range or corneal thickness threshold, as well as prohibited situations such as a history of keratoconus or severe dry eye. Historical efficacy risk data, through the analysis of a large number of surgical cases, records the postoperative effect indicators of different parameter combinations, including visual acuity recovery rate, dry eye incidence, and refractive regression probability. The surgical plan library integrates clinical guidelines, authoritative literature, historical clinical databases, and empirical data to provide the dynamic Bayesian network model with a basis for screening candidate plans and parameter optimization boundaries, ensuring that the generated personalized surgical plans are safe, effective, and traceable.

[0050] Optionally, the surgical plan library construction process includes first defining the indications, contraindications, and initial parameter configuration ranges for each surgical type based on ophthalmic clinical guidelines and historical clinical databases. Simultaneously, by analyzing a large amount of historical surgical data, statistical association rules between parameters and efficacy are extracted, and the parameter configuration ranges are dynamically revised. This construction method not only ensures the clinical safety of parameter configuration, but also achieves continuous improvement in personalized adaptation capabilities through empirical data optimization.

[0051] The dynamic Bayesian network model and the surgical plan library collaborate through multi-stage linkage to generate myopia surgical plans. Specifically, candidate plans are screened first. The dynamic Bayesian network model calls the clinical rules and parameter configuration ranges in the surgical plan library, and generates a list of operable surgical types in combination with the preoperative data of the patient to be diagnosed. Parameter optimization is then performed. The dynamic Bayesian network model uses the Bayesian optimization algorithm to search for the optimal combination of configuration parameters within the biomechanical safety boundaries set by the surgical plan library to maximize the probability of postoperative vision recovery and minimize the risk of complications. After obtaining the postoperative follow-up data, it will be fed back to the surgical plan library to trigger the adjustment of the parameter range and risk probability. The dynamic Bayesian network model updates the postoperative effect prediction logic in real time by relearning the parameters of the conditional probability table, forming a full-cycle collaborative mechanism of "rule screening → parameter optimization → data feedback → closed-loop iteration" to ensure that the surgical plan continues to evolve with clinical practice.

[0052] Optionally, the ocular parameters include at least one of refractive power, corneal thickness, corneal curvature, anterior chamber depth, axial length, corneal endothelial cell count, corneal biomechanical parameters, intraocular pressure, tear film break-up time, and meibomian gland function; and the lifestyle data includes at least one of eye habits, time spent outdoors, and dietary structure. Eye habits include average daily close-up eye use, screen usage patterns (such as blue light exposure and blinking frequency), and ambient light intensity. Outdoor time refers to average daily exposure to natural light. Less than one hour of outdoor time may exacerbate myopia progression. Dietary deficiencies in vitamins A / D and omega-3 may delay corneal repair and exacerbate postoperative dry eye or inflammatory reactions, necessitating targeted nutritional intervention. Patient basic information includes age, gender, and occupation. Past medical history includes a history of dry eye, family history of keratoconus, history of eye surgery, allergies, associated systemic diseases, and other ophthalmic conditions. Allergies may include allergies to anesthetics. The history of systemic diseases such as diabetes, hypertension, etc. The history of other ophthalmic diseases such as glaucoma, retinopathy, keratitis, etc.

[0053] Optionally, the method for generating a myopia surgery plan further includes:

[0054] collecting intraoperative data of the patient in real time during the operation, wherein the intraoperative data includes at least one of an eye movement trajectory, laser cutting parameters, and corneal flap production parameters;

[0055] The intraoperative data is input into the dynamic Bayesian network model to obtain parameter configuration for updating the surgical plan.

[0056] The synergy between real-time intraoperative data collection and a dynamic Bayesian network (DBN) model significantly improves the accuracy and safety of surgical plans. Specifically, during surgery, intraoperative data such as eye movement trajectory, laser cutting parameters, and corneal flap creation parameters are collected in real time. Leveraging the temporal reasoning capabilities of the dynamic Bayesian network, the model's conditional probability table is updated in real time, dynamically correcting the operational parameters in the preoperative surgical plan. For example, when a sudden eye deviation is detected, the model adjusts the laser cutting path or energy parameters based on the real-time eye movement trajectory to avoid the risk of postoperative high-order aberrations caused by deviation from the cutting center. If the corneal flap thickness deviates from the preset value during surgery, the model recalculates the safe range of the remaining cutting amount to ensure corneal biomechanical stability. This dynamic optimization mechanism overcomes the limitations of traditional static models, which are unable to adapt to real-time intraoperative changes. It also enables personalized fine-tuning of surgical plans through closed-loop feedback, ultimately reducing the incidence of complications (such as dry eye and glare) and improving postoperative visual quality.

[0057] Optionally, the method for generating a surgical plan further includes:

[0058] After the operation, obtaining follow-up data of the patient, wherein the follow-up data includes at least one of postoperative visual recovery, occurrence of complications, and change in visual quality;

[0059] The follow-up data is used to update the conditional probability table of the dynamic Bayesian network model to optimize the model's predictive ability.

[0060] After follow-up data is collected, the conditional probability table (CPT) of the dynamic Bayesian network model is updated to recalibrate the probabilistic dependencies between the nodes in the model. For example, if postoperative data show that a certain type of patient (such as high myopia with thin cornea) has a significantly higher-than-expected risk of refractive regression after SMILE surgery, the model will automatically correct the CPT link of "corneal thickness → surgical plan selection → refractive stability", lowering the recommendation priority of SMILE for this group and dynamically optimizing the risk-benefit ratio of other surgical procedures. This continuous learning mechanism based on real-world efficacy not only compensates for the discrepancy between preoperative predictions and actual postoperative results, but also provides more accurate long-term risk warnings for rare cases or complex clinical scenarios, ultimately enabling iterative optimization and full lifecycle management of surgical plans.

[0061] In a second aspect, a system for generating a myopia surgery plan includes:

[0062] Data collection module: collects historical surgical data, including preoperative data, intraoperative data, and postoperative data of multiple patients. The preoperative parameters include eye parameters, lifestyle data, medical history data, and basic patient information.

[0063] Preprocessing module: preprocessing the historical surgical data to obtain a training set;

[0064] Model building module: building a dynamic Bayesian network model based on the training set;

[0065] Plan generation module: collects preoperative data of the patient to be diagnosed, inputs the dynamic Bayesian network model, and generates a myopia surgery plan for the patient to be diagnosed.

[0066] In a third aspect, a computer-readable storage medium stores a computer program, wherein when the program is executed by a processor, the program implements the steps of the method for generating a myopia surgery plan.

[0067] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art:

[0068] 1. By integrating the patient's ocular biometrics, lifestyle, medical history, basic information, intraoperative data, and postoperative follow-up results, we construct multi-dimensional, multi-source, multimodal data. This multimodal data fusion overcomes the limitations of traditional reliance on a single parameter and provides comprehensive support for surgical decision-making.

[0069] 2. The dynamic Bayesian network model uses a network structure designed based on medical causal logic, with clear and transparent node relationships. The model uses a particle filtering algorithm to update in real time during surgery, dynamically adjusting risk predictions and providing proactive risk warnings. Furthermore, the Bayesian approach incorporates clinical prior knowledge, maintaining inference stability even in rare cases or with insufficient data, significantly improving decision reliability.

[0070] 3. The solution provides personalized adaptation throughout the entire preoperative, intraoperative, and postoperative process. Preoperatively, the model predicts the efficacy probability and risk distribution of different surgical options, generating the optimal recommendation based on the patient's condition. During surgery, real-time dynamic data is used to adjust operating parameters and reduce operational errors. Postoperatively, long-term tracking of visual acuity changes and complication data continuously optimizes the model's predictive capabilities. This closed-loop system upgrades the traditional experience-driven approach to one driven by data and models, ensuring precise adaptation at every stage.

[0071] 4. In clinical practice, this approach reduces reliance on the physician's subjective experience and improves the consistency of surgical plans through standardized data models. A dynamic intraoperative warning mechanism proactively mitigates operational risks and enhances surgical safety. Personalized postoperative care further improves patient visual quality and satisfaction, driving overall improvements in the quality of medical services.

[0072] In order to make the above and other objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0074] Figure 1 4 is a flow chart of a method for generating a myopia surgery plan in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0076] Example 1: See Figure 1 As shown, a method for generating a myopia surgery plan includes the following steps:

[0077] S1. Collect historical surgical data;

[0078] The historical surgical data includes preoperative data, intraoperative data and postoperative data of multiple patients. The preoperative data includes eye parameters, lifestyle data, medical history data and patient basic information. The eye parameters include at least one of refractive power, corneal thickness, corneal curvature, anterior chamber depth, axial length, corneal endothelial cell count, corneal biomechanical parameters, intraocular pressure, tear film rupture time, and meibomian gland function. The lifestyle data includes at least one of eye habits, outdoor activity time, and dietary structure. The patient basic information includes the patient's age, gender, and occupation. The medical history includes a history of dry eye, a family history of keratoconus, a history of eye surgery, a history of allergies, a history of related systemic diseases, and a history of other ophthalmic diseases. The allergic history includes an allergy to anesthetics. The history of related systemic diseases includes diabetes, hypertension, etc. The history of other ophthalmic diseases includes glaucoma, retinopathy, keratitis, etc. The historical surgical data is formed into a multimodal data set by integrating heterogeneous data from different sources (such as medical equipment, patient self-reports, and medical records), different forms (numerical values, images, and text), and different dimensions (physiological parameters, behavioral characteristics, and medical history information), thereby providing multi-angle information support for surgical plan recommendations and risk prediction.

[0079] The historical surgical data is obtained by integrating clinical data from various medical institutions. Specifically, the eye parameters in the preoperative data are collected by professional medical equipment, including optical biometers, corneal topographs and other equipment to obtain numerical and image data such as refractive power, corneal thickness, and corneal biomechanical parameters; lifestyle data such as eye habits and outdoor activity time are recorded as text and numerical information through patient self-report or standardized questionnaires; past medical history data such as dry eye, diabetes, glaucoma, etc. are extracted from electronic health records or patient declaration records in terms of classification and text content; basic patient information such as age, gender, and occupation are obtained through hospital information systems or registration forms to obtain structured data. Intraoperative data includes dynamic response data such as eye movement trajectory and laser cutting parameters, which are collected in real time by the built-in sensors of the surgical equipment as time series signals and numerical values. Postoperative data are obtained through follow-up examinations, such as the eye chart test to record the visual recovery value, the wavefront aberrometer to generate visual quality images, and the dry eye questionnaire to collect text feedback.

[0080] S2. Preprocessing the historical surgical data to obtain a training set;

[0081] Specifically, the preprocessing includes data cleaning, data normalization and data enhancement. The data cleaning is used to fill in the data areas that are beyond the reasonable range, fill in the missing data of patients and remove duplicate records, ensure the integrity and reliability of the input data, and avoid noise interference in model learning. Data normalization refers to converting data of different dimensions into the same scale. The methods used include Min-Max normalization and Z-score normalization. Min-Max normalization is to perform a linear transformation on the original data through the minimum and maximum values ​​of the data to eliminate the interference of dimensional differences on model training. Min-Max normalization formula: , where Min is the minimum value of the feature in the dataset, and Max is the maximum value of the feature in the dataset. Z-score normalization (also known as standardization) is used to convert data of different dimensions and ranges into a standard normal distribution with a mean of 0 and a standard deviation of 1. Its core is to eliminate the interference of dimensional differences on model training by calculating the distance between the data point and the mean (in units of standard deviation). Z-score normalization formula: ,in, is the mean value of the feature in the dataset, The standard deviation of this feature in the dataset. For example, diopters are measured in "D" and corneal thickness is measured in "μm." Converting to a unified scale using the Min-Max or Z-score method eliminates the interference of magnitude differences on model weight updates, accelerates model convergence, and improves the stability of parameter optimization.

[0082] Data augmentation involves performing operations such as rotation, flipping, scaling, translation, and noise addition on image data. It also involves semantic expansion or synthetic generation of unstructured data to expand the size and diversity of the training set, effectively alleviating overfitting caused by insufficient data and enhancing the model's generalization capabilities for complex clinical scenarios. Image data can include corneal topography and OCT images, while unstructured data can include descriptions of lifestyle habits and medical history text.

[0083] S3, constructing a dynamic Bayesian network model based on the training set;

[0084] S31. Determine the network structure;

[0085] The network structure of the dynamic Bayesian network model consists of parent nodes, intermediate nodes, and child nodes. The parent nodes represent preoperative data, representing the fundamental factors influencing surgical plan selection. The intermediate nodes represent myopia surgical plans. The child nodes represent postoperative data, reflecting surgical outcomes and real-time risks. The connections between the parent nodes, intermediate nodes, and child nodes are based on causal relationships defined by medical standard documents and statistical results from historical clinical databases.

[0086] The method for defining causal relationships is achieved through the systematic integration of surgical cases from ophthalmic clinical guidelines, authoritative literature, and historical clinical databases. Specifically, key logic is extracted from clinical guidelines, authoritative literature, and historical clinical databases and converted into directed connection rules in the network. At the same time, its statistical significance is verified using historical surgical data. When there is a deviation between the actual data and the preset causal link, the variable relationship is re-evaluated through conditional probability analysis, and the network structure is modified when necessary to make the model consistent with medical practice and adaptable to data characteristics. This process transforms abstract clinical experience into a traceable reasoning path, while ensuring the interpretability of the model and supporting dynamic risk prediction and real-time decision optimization.

[0087] S32, parameter learning in conditional probability table (CPT);

[0088] Using historical surgical data, the conditional probability distribution between nodes is calculated using the Bayesian estimation method.

[0089] formula: ,in, Represents model parameters, such as the probability values ​​of the conditional probability table in a dynamic Bayesian network; represents observational data, i.e., preoperative, intraoperative, and postoperative data in the historical surgical dataset; Represents the prior probability, reflecting the prior knowledge of the parameters before the data is observed; is the likelihood function, which represents the probability distribution of observed data when the parameters are known; is the posterior probability, which represents the parameter update result after combining data with prior knowledge; is the marginal likelihood and serves as a normalization constant to ensure that the probabilities sum to 1.

[0090] Structured variational inference is used to efficiently solve CPT parameters and address the computational complexity problem under high-dimensional data.

[0091] S4. Collect preoperative data of the patient to be diagnosed and input the data into the dynamic Bayesian network model. The dynamic Bayesian network model generates a myopia surgery plan for the patient to be diagnosed based on a surgery plan library.

[0092] Specifically, the surgical plan library includes myopia surgery types, parameter configuration ranges associated with the myopia surgery types, and indications, contraindications, and historical efficacy risk data corresponding to the myopia surgery types.

[0093] The types of surgery include corneal ablation, lens implantation and other auxiliary procedures. The corneal ablation types include SMILE, LASIK, PRK, etc. The lens implantation types include ICL lens implantation, lens exchange, etc. The other auxiliary procedures include AK astigmatic keratotomy, etc. For each type of surgery, there is an associated range of biomechanical safety parameters defined by expert knowledge. Indications and contraindications clearly define the applicable conditions for each type of surgery, as well as the circumstances in which it is prohibited. Historical efficacy risk data records postoperative effect indicators of different parameter combinations by analyzing a large number of surgical cases, including visual acuity recovery rate, incidence of dry eye, probability of refractive regression, etc.

[0094] The process of constructing the surgical plan library involves first defining the indications, contraindications, and initial parameter configuration ranges for each surgical type based on ophthalmic clinical guidelines and historical clinical databases. Simultaneously, by analyzing a large amount of historical surgical data, statistical association rules between parameters and efficacy are extracted, allowing for dynamic revision of parameter configuration ranges. This construction method ensures the clinical safety of parameter configuration while continuously improving personalized adaptation capabilities through empirical data optimization.

[0095] The dynamic Bayesian network model and the surgical plan library collaborate through multi-stage linkage to generate myopia surgical plans. Specifically, candidate plans are first screened. The dynamic Bayesian network model calls the clinical rules and parameter configuration ranges in the surgical plan library and generates a list of feasible surgical types based on the preoperative data of the patient to be diagnosed. Parameter optimization is then performed. The dynamic Bayesian network model uses the Bayesian optimization algorithm to search for the optimal combination of configuration parameters within the biomechanical safety boundaries set by the surgical plan library to maximize the probability of postoperative vision recovery and minimize the risk of complications.

[0096] Specifically, after inputting the preoperative data of the patient to be diagnosed into the dynamic Bayesian network model, after model inference, the following posterior probability distribution of each sub-node under different surgical types is obtained:

[0097] P (visual acuity ≥1.0 at 1 week postoperatively | SMILE, patient data) = 0.95;

[0098] P (visual acuity ≥1.0 at 1 week postoperatively | LASIK, patient data) = 0.90;

[0099] P (visual acuity ≥1.0 at 1 week postoperatively | PRK, patient data) = 0.85;

[0100] P(dry eye risk|SMILE, patient data) = 0.10;

[0101] P(dry eye risk|LASIK, patient data) = 0.15;

[0102] P(risk of dry eye|PRK, patient data)=0.20.

[0103] Based on the posterior probability distribution and comprehensive consideration of efficacy and risk, the dynamic Bayesian network model recommends SMILE surgery as the best type of surgery among the above-mentioned surgical types for patients to be diagnosed.

[0104] In an optional embodiment, the method for generating a myopia surgery plan further includes:

[0105] S5, real-time reasoning during surgery;

[0106] Specifically, during the operation, intraoperative data of the patient to be diagnosed is collected in real time, wherein the intraoperative data includes at least one of an eye movement trajectory, laser cutting parameters, and corneal flap production parameters;

[0107] The intraoperative data is input into the dynamic Bayesian network model to obtain parameter configuration for updating the surgical plan.

[0108] In an optional embodiment, the method for generating a myopia surgery plan further includes:

[0109] S6. Model iterative update.

[0110] After the operation, obtaining follow-up data of the patient, wherein the follow-up data includes at least one of postoperative visual recovery, occurrence of complications, and change in visual quality;

[0111] The follow-up data is used to update the conditional probability table of the dynamic Bayesian network model to optimize the model's predictive ability.

[0112] This embodiment also discloses a system for generating a myopia surgery plan, including:

[0113] Data collection module: collects historical surgical data, including preoperative data, intraoperative data, and postoperative data of multiple patients. The preoperative parameters include eye parameters, lifestyle data, medical history data, and basic patient information.

[0114] Preprocessing module: preprocessing the historical surgical data to obtain a training set;

[0115] Model building module: building a dynamic Bayesian network model based on the training set;

[0116] Plan generation module: collects preoperative data of the patient to be diagnosed, inputs the dynamic Bayesian network model, and generates a myopia surgery plan for the patient to be diagnosed.

[0117] This embodiment further discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for generating a myopia surgery plan.

[0118] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A method for generating a myopia surgery plan, characterized in that: The following steps are involved: Collect historical surgical data, including preoperative, intraoperative, and postoperative data of the patient. The preoperative data includes eye parameters, lifestyle data, medical history data, and basic patient information. Preprocessing the historical surgical data to obtain a training set; Based on the training set, a conditional probability table is learned using a Bayesian estimation method; A structured variational inference algorithm is used to optimize the parameters of the conditional probability table to obtain a dynamic Bayesian network model. The network structure of the dynamic Bayesian network model includes a parent node, an intermediate node, and a child node. The parent node is the preoperative data, the intermediate node is the myopia surgery plan, and the child node is the postoperative data. The connection relationship between the parent node, the intermediate node, and the child node is based on the causal relationship defined by medical standard documents and statistical results of historical clinical databases. Collecting preoperative data of the patient to be diagnosed and inputting it into the dynamic Bayesian network model, wherein the dynamic Bayesian network model generates a myopia surgery plan for the patient to be diagnosed based on a surgery plan library; During the operation, intraoperative data of the patient to be diagnosed is collected in real time, and the intraoperative data includes at least one of the eye movement trajectory, laser cutting parameters, and corneal flap production parameters; the intraoperative data is input into the dynamic Bayesian network model, and the conditional probability table of the dynamic Bayesian network model is updated in real time. The operating parameters in the preoperative surgical plan are dynamically corrected to obtain the parameter configuration of the updated surgical plan.

2. The method for generating a myopia surgery plan according to claim 1, wherein: The pretreatment includes: Data cleaning, used to remove outliers, missing values, and duplicate values; Data normalization, converting data of different dimensions into a unified scale; Data augmentation is used to expand image data and unstructured data to improve the generalization ability of the dynamic Bayesian network model.

3. The method for generating a myopia surgery plan according to claim 1, wherein: The surgical plan library includes myopia surgery types, parameter configuration ranges associated with the myopia surgery types, and indications, contraindications and historical efficacy risk data corresponding to the myopia surgery types.

4. The method for generating a myopia surgery plan according to claim 1, wherein: The eye parameters include at least one of refractive power, corneal thickness, corneal curvature, anterior chamber depth, axial length, corneal endothelial cell count, corneal biomechanical parameters, intraocular pressure, tear film rupture time, and meibomian gland function; the lifestyle data include at least one of eye use habits, outdoor activity time, and dietary structure; the patient basic information includes the patient's age, gender, and occupation.

5. The method for generating a myopia surgery plan according to claim 1, wherein: Also includes: After the operation, obtaining follow-up data of the patient, wherein the follow-up data includes at least one of postoperative visual recovery, occurrence of complications, and change in visual quality; The follow-up data is used to update the conditional probability table of the dynamic Bayesian network model to optimize the model's predictive ability.

6. A system for generating a myopia surgery plan, characterized in that: include: Data collection module: collects historical surgical data sets, which include preoperative data, intraoperative data, and postoperative data of multiple patients. The preoperative data includes eye parameters, lifestyle data, and medical history data. Preprocessing module: preprocessing the historical surgical data set to obtain a training set; Model construction module: Based on the training set, the conditional probability table is learned using the Bayesian estimation method, and the parameters of the conditional probability table are optimized using the structured variational inference algorithm to obtain a dynamic Bayesian network model; the network structure of the dynamic Bayesian network model includes a parent node, an intermediate node, and a child node, wherein the parent node is the preoperative data, the intermediate node is the myopia surgery plan, and the child node is the postoperative data. The connection relationship between the parent node, the intermediate node, and the child node is based on the causal relationship defined by the medical standard documents and the statistical results of the historical clinical database; Plan generation module: collects preoperative data of the patient to be diagnosed, inputs it into the dynamic Bayesian network model, and generates a myopia surgery plan for the patient to be diagnosed; collects intraoperative data of the patient to be diagnosed in real time during the operation, and the intraoperative data includes at least one of the eye movement trajectory, laser cutting parameters, and corneal flap production parameters; inputs the intraoperative data into the dynamic Bayesian network model, updates the conditional probability table of the dynamic Bayesian network model in real time, dynamically corrects the operating parameters in the preoperative surgical plan, and obtains the updated parameter configuration of the surgical plan.

7. A storage medium storing a computer program, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images