A System for Evaluating the Efficacy of End-Stage Renal Disease and Predicting the Risk of Complications

Abstract

The present invention discloses a system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications, comprising: a data input preparation module for acquiring and preprocessing the ophthalmic examination data and metadata of hemodialysis and peritoneal dialysis patients; a model training and evaluation module for training a model using machine learning algorithms and classical statistical regression, and selecting a model using the cross-validation method and combining the accuracy rate of the model; and a prediction result output module for respectively outputting corresponding classification results and prediction probabilities through the efficacy evaluation model and the complication and prognosis risk prediction model. By adopting the embodiment provided by the present invention, a comprehensive evaluation can be made on the treatment conditions and long-term prognosis of end-stage patients before treatment, during treatment, and during long-term follow-up.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and particularly to a system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications. Background Art

[0002] With the aging of the population and the increasing incidence of hypertension, diabetes, etc. year by year, the number of patients with chronic kidney disease (CKD) progressing to end-stage renal disease (ESRD) is also increasing. For patients with end-stage renal disease, renal replacement therapy including hemodialysis and peritoneal dialysis is essential. Dialysis treatment can replace part of the excretory function of the kidneys and is widely used clinically. However, the insufficient dialysis efficacy and the occurrence of complications (such as hypotension, adverse cardiovascular events, etc.) are the main reasons for dialysis failure and low patient survival rate. Therefore, accurately and conveniently evaluating the dialysis efficacy of patients and predicting the occurrence of complications has important clinical significance.

[0003] At present, the examination methods for end-stage renal disease patients have the following disadvantages: 1. Low monitoring frequency, unable to meet the diagnosis and treatment requirements: The current requirement in China is to monitor blood samples once every 3 months, and it should be once a month if conditions permit. However, in actual situations, due to limitations in technology and medical levels, in many areas, monitoring is only carried out once every six months or so, far from meeting the requirements; 2. Invasive operations, poor patient compliance: At present, the evaluation of the efficacy of dialysis patients, including monitoring indicators such as urea clearance rate, hemoglobin, and serum albumin levels, all rely on invasive venous blood sampling to extract a certain amount of blood, which may cause infections, thrombosis, and transmission of blood-borne infectious diseases. For dialysis patients with low immune function themselves, frequent invasive examination operations increase the infection risk of the patients themselves; 3. Inadequate evaluation and insufficient timeliness: The current evaluation mainly focuses on the hemodynamic monitoring during dialysis and the single adequacy evaluation after dialysis. The routine blood tests, biochemical tests, etc. performed when patients visit the doctor take half an hour to 1 day to issue reports, and the available evaluation information is insufficient, making it impossible to adjust the dialysis plan in a timely and personalized manner before dialysis or when the patient visits the doctor. 4. Strong subjectivity and low efficiency: The evaluation of end-stage renal disease patients is carried out by medical staff to formulate dialysis goals and risk stratification based on experience, and the accuracy depends on their professional knowledge level, and the diagnosis is highly subjective. In addition, due to the shortage of medical resources, the workload of a single doctor is too large, and missed diagnoses and misdiagnoses are likely to occur. Generally speaking, the current examinations for end-stage renal disease patients, including blood tests, etc., are mostly invasive operations, time-consuming, costly, and unable to evaluate the patient's overall condition in a timely manner. Summary of the Invention

[0004] An embodiment of the present invention provides a system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications, which uses simple and non-invasive eye examination information to comprehensively evaluate the treatment status and long-term prognosis of end-stage patients before treatment, during treatment, and in long-term follow-up.

[0005] To achieve the above object, a system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications according to an embodiment of the present application includes:

[0006] A data input preparation module for acquiring and preprocessing the eye examination data and metadata of hemodialysis and peritoneal dialysis patients; the metadata includes patient medical record information, efficacy evaluation indicators, complication and prognosis standard data;

[0007] The data input preparation module includes an image processing unit and a metadata processing unit; the image processing unit is used to standardize the eye examination images in the eye examination data according to the examination type used in the eye examination data; the metadata processing unit is used to clean and uniformly format the metadata according to the formatted data type;

[0008] A model training and evaluation module for training a model using machine learning algorithms and classical statistical regression, and selecting a model using the cross-validation method and combining the accuracy of the model;

[0009] The model training and evaluation module includes a feature selection unit, a model training unit, a model evaluation and selection unit, and a model adjustment unit;

[0010] The feature selection unit is used to select the input feature variables of the model from the eye examination data, metadata, and multi-modal data; the multi-modal data is combined data about the eye examination data and metadata;

[0011] The model training unit is used to train a model using machine learning algorithms or statistics in combination with the input feature variables;

[0012] The model evaluation and selection unit is used to screen the trained model according to the AUC value of the accuracy of the model;

[0013] The model adjustment unit is used to adjust the relevant variable parameters of the screened model according to the efficacy evaluation target and the complication risk prediction target, respectively, using the features and algorithms used in the model, and train the final model on all development data sets to obtain an efficacy evaluation model, a complication and prognosis risk prediction model;

[0014] A prediction result output module for outputting the corresponding prediction probabilities and classification results through the efficacy evaluation model and the complication and prognosis risk prediction model, respectively.

[0015] In a possible implementation, cleaning and formatting the eye examination data according to the examination type adopted for the eye examination data specifically includes:

[0016] Extracting the image features of the eye examination data;

[0017] Performing duplicate data exclusion, missing value and outlier processing, categorical variable recoding, and data transformation on the formatted eye examination data;

[0018] If the examination type adopted for the eye examination data is fundus photography and the image features meet the requirements that the position is within a preset coordinate range and there is no occlusion, the brightness is greater than a preset photography threshold, and the resolution is greater than a preset resolution threshold, retain the eye examination data;

[0019] If the examination type adopted for the eye examination data is fundus optical coherence angiography and the image features meet the requirement that the image signal intensity is greater than a preset intensity threshold, retain the eye examination data.

[0020] In a possible implementation, selecting the input feature variables of the model from the eye examination data, metadata, and multimodal data specifically includes:

[0021] Using the Lasso or forward stepwise regression method to select the model input feature variables from the eye examination data, metadata, and multimodal data.

[0022] In a possible implementation, the selection conditions for the input feature variables include: meeting clinical indicators, the variable missing rate being less than a preset threshold, the variable variance not being equal to 0, and selecting one of two variables with a correlation index greater than the threshold.

[0023] In a possible implementation, training the model by combining the input feature variables using a machine learning algorithm or a classical statistical algorithm specifically includes:

[0024] Combining the input feature variables to train the model using a machine learning algorithm or a Logistic regression model.

[0025] In a possible implementation, screening the trained model according to the accuracy AUC value of the model specifically includes:

[0026] Taking the input patient eye examination data and metadata, and using the method of cross-validation. Randomly divide the patients into 10 or 5 folds with equal sample sizes. Each time, use 9 or 4 folds of the data as the training set and 1 fold as the validation set;

[0027] Select according to the ROC curves and AUC values of different models cross-validated by the training set and the validation set.

[0028] In a possible implementation, for the selected models, the features and algorithms used by the models are adopted to adjust the relevant variable parameters of dialysis adequacy and blood pressure control, and the models are trained on all the development data sets while adjusting the relevant variable parameters of dialysis adequacy and blood pressure control to obtain a therapeutic effect evaluation model;

[0029] For the selected models, the features and algorithms used by the models are adopted to adjust the relevant variable parameters of short-term anemia, malnutrition, and long-term cardiovascular complications, and the models are trained on all the development data sets to obtain a complication and prognosis risk prediction model;

[0030] For the therapeutic effect evaluation model and the complication and prognosis risk prediction model, tests on an external validation set are performed to respectively evaluate the generalization ability.

[0031] In a possible implementation, the corresponding prediction probabilities and classification results are respectively output through the therapeutic effect evaluation model and the complication and prognosis risk prediction model, specifically including:

[0032] The therapeutic effect evaluation model and the complication and prognosis risk prediction model output probability values, binary classification results, and risk stratification results for the judgment of blood pressure control, dialysis adequacy, anemia status, malnutrition status, short-term and long-term complications.

[0033] Compared with the prior art, a system for evaluating the therapeutic effect of end-stage renal disease and predicting the complication risk provided by an embodiment of the present invention uses the non-invasively obtained eye examination data and patient metadata to establish an artificial intelligence model (therapeutic effect evaluation model, complication and prognosis risk prediction model) based on eye information in the system to evaluate the dialysis therapeutic effect of end-stage renal disease and predict the dialysis complications and prognosis risk of end-stage renal disease. The therapeutic effect evaluation model predicts various therapeutic effect indicators for end-stage renal disease patients without invasive blood sampling tests, etc., and is not limited by the personal knowledge and experience of users. The complication and prognosis risk prediction model predicts multi-period, multi-system complications and long-term prognosis through eye information combined with formatted information of the nephrology specialty, and has a higher accuracy rate.

[0034] BRIEF DESCRIPTION OF THE DRAWINGS FIG.

[0035] Figure 1 is a schematic structural diagram of a system for evaluating the therapeutic effect of end-stage renal disease and predicting the complication risk provided by an embodiment of the present invention;

[0036] Figure 2It is a flowchart of the operation of a system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications provided by an embodiment of the present invention. Detailed implementation manners

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0038] Currently, the examinations for end-stage renal disease patients, including blood tests, etc., are mostly invasive operations, time-consuming, costly, and unable to evaluate the patient's overall condition in a timely manner. The vascular structure and perfusion conditions of the fundus are closely related to the overall state. Previous studies have shown that significant changes have occurred in relevant ophthalmic indicators before and after hemodialysis, and they are related to systemic indicators such as blood pressure, blood urea nitrogen, and weight changes. Corresponding ophthalmic indicators of peritoneal dialysis patients have also changed significantly. Ophthalmic examinations are non-invasive and fast. Optical coherence tomography angiography of the fundus can observe the blood flow signals of the fundus in real time and intuitively. After the input eye information is calculated by a machine learning model, the efficacy of the dialysis patients seeking medical treatment can be evaluated and the prediction of complications can be made.

[0039] In the following embodiments, in view of the deficiencies in the evaluation of the dialysis efficacy and the prediction of the risk of complications of existing end-stage renal disease patients, an efficacy evaluation model for end-stage renal disease and a complication and prognosis risk prediction model based on eye information are proposed to construct a system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications.

[0040] Please refer to Figure 1 , an embodiment of the present application provides a system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications, including: a data input preparation module 1, a model training and evaluation module 2, and a prediction result output module 3.

[0041] The data input preparation module 1 is used to obtain and preprocess the eye examination data and metadata of hemodialysis and peritoneal dialysis patients; the metadata includes patient medical record information, efficacy evaluation indicators, complication and prognosis standard data.

[0042] The data input preparation module 1 includes an image processing unit 10 and a metadata processing unit 11; the image processing unit 10 is used to standardize the eye examination images in the eye examination data according to the examination type adopted by the eye examination data; the metadata processing unit 11 is used to clean and uniformly format the metadata according to the formatted data type;

[0043] The model training and evaluation module 2 is used to train a model using machine learning algorithms and classical statistical regression, and select a model using the cross-validation method in combination with the accuracy of the model.

[0044] The model training and evaluation module 2 includes a feature selection unit 20, a model training unit 21, a model evaluation and selection unit 22, and a model adjustment unit 23.

[0045] The feature selection unit 20 is used to select the input feature variables of the model from the eye examination data, metadata, and multi-modal data; the multi-modal data is combined data regarding the eye examination data and metadata;

[0046] The model training unit 21 is used to train a model using machine learning algorithms or statistical algorithms in combination with the input feature variables.

[0047] The model evaluation and selection unit 22 is used to screen the trained model according to the AUC value of the accuracy rate of the model.

[0048] The model adjustment unit 23 is used to adjust the relevant variable parameters of the screened model according to the efficacy evaluation target and the complication risk prediction target respectively, using the features and algorithms used by the model, and train the final model on all development data sets to obtain an efficacy evaluation model, a complication and prognosis risk prediction model.

[0049] The prediction result output module 3 is used to output the corresponding prediction probabilities and classification results through the efficacy evaluation model and the complication and prognosis risk prediction model respectively.

[0050] In this embodiment, for the dialysis efficacy evaluation of end-stage renal disease, the prediction of multiple complications and prognosis risks, an artificial intelligence system centered on a series of machine learning models based on eye information is developed. By inputting the corresponding eye data (with or without the results of nephrology specialist examinations), a comprehensive assessment of the patient's condition can be made before treatment, during treatment, and during long-term follow-up, which helps to reduce the invasive examination items of the patient, formulate an effective follow-up plan, and provide an objective decision-making basis for the selection of treatment plans, timely adjustment, and the formulation of follow-up plans. Among them, the efficacy evaluation model can evaluate and classify and predict multiple indicators for end-stage renal disease patients, including but not limited to: blood pressure, body weight, BMI, complete blood count (white blood cell count, red blood cell count, platelets, etc.), hemoglobin, nutritional indicators (albumin, blood lipids, etc.), electrolyte indicators, and dialysis adequacy evaluation (urea clearance rate, urea reduction rate, β2-MG clearance rate), etc. The complication risk prediction model includes cardiovascular system complications, vascular access complications, skeletal muscle system complications, and nervous system complications, etc.

[0051] Please refer to Figure 2, the data input preparation module 1 is responsible for data preparation and image processing in the system.

[0052] Data preparation: Extract and integrate the eye examination data and metadata of hemodialysis and peritoneal dialysis patients respectively. The metadata includes the basic information of the patients and the examination data of end-stage renal disease (potential input features), the efficacy evaluation indicators, and the standard data of complications and prognosis (label format, output format).

[0053] The image processing work is completed by the image processing unit 10.

[0054] Exemplarily, in the image processing unit 10, the eye examination data is cleaned and formatted according to the examination type adopted for the eye examination data, specifically including:

[0055] Extract the image features of the eye examination data;

[0056] Exclude duplicate data, handle missing values and outliers, re-encode categorical variables, and perform data conversion on the formatted eye examination data;

[0057] If the examination type adopted for the eye examination data is fundus photography, and the image features meet the requirements that the position is within the preset coordinate range and there is no occlusion, the brightness is greater than the preset photography threshold, and the resolution is greater than the preset resolution threshold, the eye examination data is retained;

[0058] If the examination type adopted for the eye examination data is optical coherence angiography of the fundus, and the image features meet the requirement that the image signal intensity is greater than the preset intensity threshold, the eye examination data is retained.

[0059] The image processing unit 10 performs quality assessment on the images according to different examination types. Taking the fundus as an example, the global image features of the fundus image are extracted: brightness, contrast, integrity; and local features: sharpness, gradient, local blood vessel density, distribution, etc. The data cleaning process includes excluding duplicate data, handling missing values and outliers, re-encoding categorical variables, data conversion, etc. for the formatted data. Images and data that do not meet the quality standards are excluded. Finally, preprocessing and image analysis are performed on the images to obtain formatted eye examination data.

[0060] As a non-invasive ophthalmic examination, it causes no damage to patients, is convenient and fast, and can rely on portable equipment, such as handheld slit lamps, smartphone fundus photography, and OCT handheld imaging, and can also be carried out in remote areas and at home, which is convenient and fast. Ophthalmic information can be obtained, analyzed, and diagnosed immediately, with strong timeliness, and the dialysis plan can be adjusted promptly and personalized according to the eye-related information before dialysis or when the patient visits.

[0061] Please refer to Figure 2, the model training and evaluation module 2 is responsible for training the model using machine learning algorithms and classical statistical regression, and selecting the model using the cross-validation method in combination with the accuracy of the model.

[0062] Exemplarily, in the feature selection unit 20, the selection of the input feature variables of the model from the eye examination data, metadata, and multi-modal data specifically includes:

[0063] Use the Lasso or forward stepwise regression method to select the input feature variables of the model from the eye examination data, metadata, and multi-modal data.

[0064] Exemplarily, in the feature selection unit 20, the selection conditions for the input feature variables include: conforming to clinical indicators, the variable missing rate being less than a preset threshold, the variable variance not being equal to 0, and selecting one of the two variables with a correlation index greater than the threshold for retention.

[0065] Exemplarily, in the model training unit 21, the combination of using the input feature variables and using machine learning algorithms or classical statistical algorithms for model training specifically includes:

[0066] Combine the input feature variables and use machine learning algorithms or the Logistic regression model for model training.

[0067] Exemplarily, in the model evaluation and selection unit 22, the screening of the trained model according to the AUC value of the model accuracy specifically includes:

[0068] Take the input patient eye examination data and metadata, adopt the cross-validation method, randomly divide the patients into 10 or 5 folds with equal sample sizes, and each time use 9 or 4 folds of the data as the training set and 1 fold as the validation set;

[0069] Select according to the ROC curve and AUC value of different models cross-validated by the training set and the validation set.

[0070] Exemplarily, in the model adjustment unit 23, the internal variable condition adjustment and selection of the screened model are carried out separately according to the efficacy evaluation target and the complication risk prediction target to obtain the efficacy evaluation model and the complication and prognosis risk prediction model, specifically including:

[0071] For the screened model, adopt the features and algorithms used by the model, adjust the relevant variable parameters of dialysis adequacy and blood pressure control, and train on the entire development data set to obtain the efficacy evaluation model;

[0072] For the screened model, using the features and algorithms used by the model, adjust the relevant variable parameters of short-term anemia, malnutrition, and long-term cardiovascular complications, and train on the entire development dataset to obtain a complication and prognosis risk prediction model.

[0073] For the efficacy evaluation model and the complication and prognosis risk prediction model, conduct tests on the external validation set to evaluate the generalization ability respectively.

[0074] Exemplarily, output the corresponding prediction probabilities and classification results through the efficacy evaluation model and the complication and prognosis risk prediction model respectively, specifically including:

[0075] Output the judgment of blood pressure control, dialysis adequacy, anemia status, malnutrition status, probability values of short-term and long-term complications, binary classification results, and risk stratification results through the efficacy evaluation model and the complication and prognosis risk prediction model.

[0076] It should be noted that the direct output forms of the efficacy evaluation model and the complication and prognosis risk prediction model are all probability values. In practical applications, according to needs, a probability threshold can be set to divide the probability values output by the efficacy evaluation model and the complication and prognosis risk prediction model into binary classification or multi-classification (risk stratification).

[0077] Users of the efficacy evaluation model and the complication and prognosis risk prediction model do not need to undergo professional ophthalmology-related training, nor do they rely on the personal professional knowledge and diagnosis and treatment experience of users. They only need to input relevant information, and the model can give evaluations and predictions. The results are objective and reproducible, which is conducive to unifying the clinical diagnosis and treatment standards. In addition, the two models cover many indicators and a wide range of complication spectra, including various efficacy evaluation indicators for end-stage renal disease patients and the risk of multi-system complications, covering the entire treatment cycle of end-stage renal disease, and providing a comprehensive clinical decision support system.

[0078] The application scenario of the above embodiments is that, before dialysis treatment and without blood sampling for laboratory tests, non-invasive ophthalmic examinations (including fundus color photography, intraocular pressure measurement, optical coherence tomography angiography, etc.) are performed on patients. After the data is cleaned and processed, it is uploaded to the cloud server where the efficacy evaluation model and the complication and prognosis risk prediction model are deployed, and input into the intelligent diagnosis and treatment model to infer the patient data input by the client in real time, obtaining output results including the probability of dialysis efficacy evaluation and complication occurrence, and triggering suggestions for adjusting dialysis plans such as treatment time and ultrafiltration volume before dialysis treatment, dynamic evaluation of efficacy during the treatment process, risk stratification results, and personalized follow-up plan suggestions according to the output results. The model output results and treatment suggestions are transmitted back to the medical staff application terminal, and after review, the doctor's orders and follow-up plans are automatically distributed to the patient client terminal, and follow-up data collection and feedback are carried out.

[0079] Compared with the prior art, an end-stage renal disease efficacy evaluation and complication risk prediction system provided by an embodiment of the present invention establishes an artificial intelligence model (efficacy evaluation model, complication and prognosis risk prediction model) based on eye information in the system to evaluate the dialysis efficacy of end-stage renal disease and predict the dialysis complications and prognosis risks of end-stage renal disease through non-invasively obtained eye examination data and patient metadata. The efficacy evaluation model predicts various efficacy indicators for end-stage renal disease patients without invasive blood sampling for laboratory tests and is not limited by the personal knowledge and experience of users. The complication and prognosis risk prediction model predicts multi-period, multi-system complications and long-term prognosis through eye information combined with formatted information in the nephrology specialty, and has a higher accuracy rate.

[0080] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A system for evaluating the efficacy of end-stage renal disease and predicting the risk of complications, characterized in that, Comprising: A data input preparation module for obtaining and preprocessing the eye examination data and metadata of hemodialysis and peritoneal dialysis patients; the metadata includes patient medical record information, efficacy evaluation indicators, complications, and prognosis standard data; The data input preparation module includes an image processing unit and a metadata processing unit; the image processing unit is used to standardize the eye examination images in the eye examination data according to the examination type used for the eye examination data; the metadata processing unit is used to clean and uniformly format the metadata according to the formatted data type; A model training and evaluation module for training a model using machine learning algorithms and classical statistical regression or statistics, and selecting a model using the cross-validation method and combining the accuracy of the model; The model training and evaluation module includes a feature selection unit, a model training unit, a model evaluation and selection unit, and a model adjustment unit; The feature selection unit is used to select the input feature variables of the model from the eye examination data, metadata, and multimodal data respectively; The multimodal data is combined data regarding the eye examination data and metadata; The model training unit is used to train a model using machine learning algorithms or statistical algorithms in combination with the input feature variables; The model evaluation and selection unit is used to screen the trained model according to the AUC value of the model's accuracy; The model adjustment unit is used to adjust the relevant variable parameters according to the efficacy evaluation target and the complication risk prediction target for the screened model, using the features and algorithms used by the model, and training the final model on all development data sets to obtain an efficacy evaluation model, a complication and prognosis risk prediction model; A prediction result output module for outputting the corresponding prediction probabilities and classification results through the efficacy evaluation model and the complication and prognosis risk prediction model respectively.

2. The end-stage renal disease treatment efficacy evaluation and complication risk prediction system according to claim 1, wherein The cleaning and formatting of the eye examination data according to the examination type used for the eye examination data specifically includes: Extracting the image features of the eye examination data; Performing duplicate data exclusion, missing value and outlier processing, categorical variable recoding, and data conversion on the formatted eye examination data; If the examination type used for the eye examination data is fundus photography, and the image features meet the conditions that the position is within a preset coordinate range and there is no occlusion, the brightness is greater than the preset photography threshold, and the resolution is greater than the preset resolution threshold, the eye examination data is retained; If the examination type used for the eye examination data is optical coherence angiography of the fundus, and the image features meet the condition that the image signal intensity is greater than the preset intensity threshold, the eye examination data is retained.

3. The end-stage renal disease treatment efficacy evaluation and complication risk prediction system according to claim 1, characterized in that, The selection of the input feature variables of the model from the eye examination data, metadata, and multimodal data specifically includes: Using the Lasso or forward stepwise regression method to select the input feature variables of the model from the eye examination data, metadata, and multimodal data.

4. The end-stage renal disease treatment effect evaluation and complication risk prediction system according to claim 3, wherein The selection conditions for the input feature variables include: meeting clinical indicators, the variable missing rate being less than a preset threshold, the variable variance not being equal to 0, and retaining one of the two variables with a correlation index greater than the threshold.

5. The end-stage renal disease treatment efficacy evaluation and complication risk prediction system according to claim 1, characterized in that, Using a machine learning algorithm or a classical statistical algorithm for model training in combination with the input feature variables, specifically including: Using a machine learning algorithm or a Logistic regression model for model training in combination with the input feature variables.

6. The end-stage renal disease treatment efficacy evaluation and complication risk prediction system according to claim 1, wherein Screening the trained model according to the accuracy AUC value of the model, specifically including: For the input patient eye examination data and metadata, adopting the method of cross-validation, randomly dividing the patients into 10 or 5 folds with equal sample sizes, and each time taking 9 or 4 folds of the data as the training set and 1 fold as the validation set; Selecting according to the ROC curves and AUC values of different models through cross-validation of the training set and the validation set.

7. The end-stage renal disease treatment efficacy evaluation and complication risk prediction system according to claim 1, characterized in that According to the efficacy evaluation objective and the complication risk prediction objective, respectively, for the screened model, using the features and algorithms used in the model to adjust the relevant variable parameters, and training the final model on all development data sets to obtain an efficacy evaluation model, a complication and prognosis risk prediction model, specifically including: For the screened model, using the features and algorithms used in the model to adjust the relevant variable parameters of dialysis adequacy and blood pressure control, and training on all development data sets to obtain an efficacy evaluation model; For the screened model, using the features and algorithms used in the model to adjust the relevant variable parameters of short-term anemia, malnutrition, and long-term cardiovascular complications, and training on all development data sets to obtain a complication and prognosis risk prediction model; Testing the efficacy evaluation model and the complication and prognosis risk prediction model on an external validation set to evaluate the generalization ability respectively.

8. The end-stage renal disease treatment efficacy evaluation and complication risk prediction system according to claim 1, wherein Outputting the corresponding prediction probabilities and classification results through the efficacy evaluation model and the complication and prognosis risk prediction model respectively, specifically including: Outputting the probability values, binary classification results, and risk stratification results of blood pressure control situation judgment, dialysis adequacy judgment, anemia status judgment, malnutrition status judgment, short-term and long-term complications through the efficacy evaluation model and the complication and prognosis risk prediction model.

Images