CN114639460A

CN114639460A - Cycloplegic demand prediction and paralysis post-diopter refractive state prediction method

Info

Publication number: CN114639460A
Application number: CN202210531703.1A
Authority: CN
Inventors: 魏瑞华; 梁猛; 王清鑫; 杜蓓; 王景慧; 金楠; 郭丽; 骆源
Original assignee: TIANJIN MEDICAL UNIVERSITY EYE HOSPITAL
Current assignee: TIANJIN MEDICAL UNIVERSITY EYE HOSPITAL
Priority date: 2022-05-16
Filing date: 2022-05-16
Publication date: 2022-06-17

Abstract

The invention relates to a method for predicting the paralysis requirement of ciliary muscle and predicting the refractive state of paralysis postflexion, which comprises the steps of measuring parameters before paralysis of ciliary muscle, and predicting whether the paralysis of ciliary muscle is required according to a binary classification model by using the measured parameters; predicting the paralysis retroflexion state of the ciliary muscle of the child and the distribution of the refraction state of the crowd according to the three classification models by using the measured parameters; and predicting the diopter of the paralyzed ciliary muscles of the children and the diopter distribution of the crowd according to the measured parameters and the regression model. The invention has the beneficial effects that: judging whether cycloplegia needs to be implemented or not by predicting the retroflexion degree of the cycloplegia, reducing the workload of the cycloplegia for medical personnel, and avoiding unnecessary cycloplegia measures for patients; the scheme of the invention can also be used for analyzing data of ciliary muscle states and predicting the average diopter and the refractive state distribution of people through three classifications and a regression model in an unbiased mode.

Description

Cycloplegic demand prediction and paralysis post-diopter refractive state prediction method

Technical Field

The invention belongs to the field of medical detection, and particularly relates to a method for predicting the paralysis demand of ciliary muscles and predicting the refractive state of paralysis postflexion.

Background

In ocular optics, the refractive state of a human eye is usually evaluated by diopter, and the human eye is classified into emmetropia, myopia and hypermetropia. The human eye diopter is the basis of further refractive correction and is also an index for measuring the refractive development condition, and the accurate acquisition of the human eye diopter value is very important in clinical and related epidemiological researches. Because the diopter measurement is influenced by the adjusting condition, a ciliary muscle paralytic agent is required to be used for fully relaxing and adjusting, and the diopter is accurately obtained. However, it presents a minor challenge in clinical and epidemiological research work to administer cycloplegia to all patients due to the extra waiting time required for cycloplegia, examination costs, photophobia after paralysis, and blurred near vision.

In the present research, it is found that the regulation tension has a high incidence in the juvenile population, but still has a large individual difference, and the regulation tension phenomenon does not exist in all the juveniles. At present, aiming at a cycloplegia strategy of teenagers and children, the cycloplegia strategy proposes that the retroflexion examination is carried out on all people in the age group, people needing the cycloplegia can be distinguished through a proper decision process in theory, the cycloplegia is carried out in a targeted mode, and unnecessary paralysis processes are reduced. In a big data era, a machine learning method is used for processing a large amount of multidimensional data, a complex prediction model is built, and high accuracy can be obtained for prediction problems.

Disclosure of Invention

In order to solve the technical problems, the invention provides a method for predicting the paralysis demand of ciliary muscles and predicting the refractive state of paralysis post-diopter.

The technical scheme adopted by the invention is as follows: a method for predicting a paralytic demand of ciliary muscles and predicting a paralytic post-refractive state, the method comprising the steps of:

measuring and obtaining numerical parameters before cycloplegia;

establishing a two-classification model according to the numerical parameters before cycloplegia, and predicting whether cycloplegia is needed or not through the two-classification model;

establishing a three-classification model according to the numerical parameters before cycloplegic, wherein the three-classification model can calculate the refractive state after cycloplegic according to the numerical parameters;

establishing a regression model according to the numerical parameters before cycloplegion, wherein the regression model can calculate diopter after cycloplegion according to the numerical parameters;

and summarizing the calculated values of the three classification models and the regression model, and performing statistical analysis on the calculated values to obtain the distribution of the non-offset refraction state and the distribution of the diopter in the test target.

Further, the numerical parameters before cycloplegia comprise optical biological data, optometry data and basic information.

Further, the optical biometric data includes one or more of ocular axial length, corneal thickness, anterior chamber depth, crystal thickness, vitreous cavity depth, and corneal curvature values;

the optometry data comprises one or more of spherical power, cylindrical power and astigmatic axis;

the basic information includes age and/or gender;

the other examinations include one or more of intraocular pressure, accommodative lag, eye position, living vision, and on-the-lens information.

Further, the two classification models are one of SVC _ W, RF _ W, DNN _ W and EEC, the model performance index is Accuracy, Sensitivity and Specificity, and the index calculation formula is:

wherein, TP: a true positive sample; FP: a false positive sample; FN: a false negative sample; TN: true negative samples.

Further, the three classification models are SVC _ W, RF _ W, DNN _ W and EEC, model performance indexes are Precision, Recall and F1Score, and an index calculation formula is as follows:

Further, the regression model is one of SVR, RFR, DNN and Adaboost, and the index calculation formula is:

wherein, the total amount m of the sample, the real equivalent sphere power is𝑦The machine learning model predicted value is

。

The invention has the advantages and positive effects that: judging whether cycloplegia needs to be implemented or not by predicting the retroflexion degree of the cycloplegia, reducing the workload of the cycloplegia for medical personnel, and avoiding unnecessary cycloplegia measures for patients; the scheme of the invention can also be used for analyzing data of ciliary muscle states, and the average diopter and the refractive state distribution of people can be predicted unbiased through three classifications and a regression model.

Drawings

FIG. 1 is a schematic diagram of model feature selection in accordance with certain embodiments of the present invention;

FIG. 2 is a schematic diagram of a model workflow in accordance with certain embodiments of the present invention;

FIG. 3 illustrates the performance of a binary prediction model in accordance with certain embodiments of the present invention;

FIG. 4 is a graph comparing the performance of a three-class prediction model according to some embodiments of the present invention with that of a conventional method;

FIG. 5 is a graph comparing the performance of a regression prediction model with that of a conventional approach in accordance with certain embodiments of the present invention.

Detailed Description

Embodiments of the present invention will be described below with reference to the accompanying drawings.

The invention designs a method for predicting paralysis of ciliary muscle and predicting refractive state of paralysis postflexion, which comprises the steps of firstly training a two-classification model/three-classification model/regression model through existing data, measuring parameters before paralysis of ciliary muscle, and predicting whether paralysis of ciliary muscle is needed according to the two-classification model by using the measured parameters; and/or predicting the light bending state of the paralyzed ciliary muscles of the children and the distribution of the crowd refractive state according to the three classification models by using the measured parameters; and/or predicting the paralytic retroflexion of the child ciliary muscle and the diopter distribution of the crowd according to the regression model by using the measured parameters.

Among them, the ophthalmic examination before cycloplegia includes optical biometry, optometry and other examinations. The optical biological data comprises one or more of ocular axial length, corneal thickness, anterior chamber depth, crystal thickness, vitreous cavity depth, and corneal curvature values; the optometry data comprises one or more of spherical power, cylindrical power and astigmatism axis; other examinations include one or more of intraocular pressure, accommodative lag, eye position, living vision, and information on wearing glasses; the basic information includes age and gender; and respectively recording the information in a data mode.

In this embodiment, four supervised learning machine algorithms, namely, a support vector machine algorithm (SVM), a random forest algorithm (RF), a deep neural network algorithm (DNN) and an integrated learning algorithm (EEC), are used for training the binary models at each threshold of 0.50D (the difference between the equivalent spherical power before and after cycloplegia is greater than 0.05D, which is considered to require retrorefractive examination of the cycloplegia). In order to avoid overfitting caused by sampling errors of the verification set during parameter adjustment, the optimal parameter search is carried out on the training set by using a ten-fold cross verification method through a Hyperopt package; and after the optimal parameters are obtained, setting the optimal parameters on the full training set to train the model newly, and evaluating the prediction capability of the model on the test set.

The support vector machine is a generalized linear classifier which carries out two-classification or multi-classification on data through a supervised learning method, and can also be used for modifying an algorithm to complete a multi-classification task. And nonlinear modeling can be carried out through a kernel method, so that the method has robustness and sparsity.

The random forest algorithm combines a plurality of base learners with accuracy and diversity by adopting a voting method to obtain the generalization performance which is obviously superior to that of a single learner.

The neural network algorithm is a parallel interconnection network consisting of simple units with adaptability, and can simulate the interactive reaction of a biological nervous system on real world objects to learn and obtain characteristic representation about the essence of original data, thereby completing a prediction task.

The ensemble learning algorithm is a machine learning algorithm specially proposed for label category unbalanced learning, a plurality of Adaboost base classifiers are trained for ensemble learning by repeatedly combining positive samples and negative samples of the same number sampled randomly, and a final predicted value is obtained by voting the result of each base learner.

Due to the fact that imbalance of positive and negative samples exists, decision bias can occur in the three machine learning algorithms, a balancing strategy is used in the model training process to reduce the occurrence of the bias, and the corresponding algorithms with balancing measures are SVC _ W, RF _ W, DNN _ W and EEC respectively.

Model performance evaluation indexes include Accuracy, Sensitivity and Specificity, and an ROC curve is drawn to calculate the area under the ROC curve (AUC). AUC is a common evaluation index in classification problems, and reflects the proportion of the number of correctly classified samples to the total number of samples. Sensitivity represents the proportion of samples that are true positive classes, and Specificity represents the proportion of samples that are true negative classes, and is predicted as a positive class. The ordinate of the ROC curve is the True case Rate (True Positive Rate: TP/(TN + FP)). AUC is the area under the ROC curve for the generalization performance of quantitative evaluation classifier, and the more the ROC curve is close to the upper left corner, the larger the area under the curve AUC is, and the better the model prediction performance is.

The index calculation formula is as follows:

TP: a true positive sample; FP: a false positive sample; FN: a false negative sample; TN: true negative samples.

According to the equivalent sphere power of the sample after cycloplegic, the collected sample is divided into three types of myopia, emmetropia and hyperopia, the labels are respectively set to be 0, 1 and 2, and a machine learning three-classification model is established to predict the retroflexion state of the three cycloplegic. Because the supervised learning classification task is also adopted, a balanced machine learning algorithm similar to that adopted when the two classification models are established is adopted in the process of establishing the three classification models: SVC _ W, RF _ W, DNN _ W and EEC. In order to avoid overfitting caused by sampling errors of the verification set during parameter adjustment, the optimal parameter search is carried out on the training set by using a ten-fold cross-validation method through a Hyperopt package. And after the optimal parameters are obtained, setting the optimal parameters on the full training set to train the model newly, and evaluating the prediction capability of the model on the test set.

The study uses ACC, Precision, Recall, F1score indices to evaluate the efficacy of the machine learning three-class model and uses the chi-square test to compare whether there is a statistical difference between the refractive state ratios predicted by the machine learning model in the sample and the refractive state ratios after true cycloplegia in the sample. The ACC calculates the proportion of the samples which are truly near-sighted and predicted to be near-sighted and the samples which are truly far-sighted and predicted to be far-sighted and the samples which are truly emmetropic and predicted to be emmetropic in the total test samples, and the samples are used for representing the overall prediction accuracy of the model.

When calculating Precision, Recall, and F1score, one of the labels 012 can be regarded as positive class sample, and the other two labels can be regarded as negative class sample, so that the 3 × 3 confusion matrix is converted into three confusion matrices similar to those in the binary task. And during model efficiency evaluation, calculating a Precision value and a Recall value for each converted two classification tasks respectively, wherein Precision is the proportion of the real positive type samples in the positive type samples predicted by the model, Recall is the proportion of the real positive type samples predicted pairs in the real positive type samples, and the model efficiency is comprehensively evaluated by calculating the harmonic mean F1score of the Precision and the Recall.

The index calculation formula is as follows:

And establishing a machine learning regression model for predicting the equivalent sphericity after cycloplegia by a regression form support vector machine algorithm (SVR), a regression form random forest algorithm (RFR), an Adaboost regression Algorithm (ABR) and a neural network connection algorithm (DNN). And when the sample label is a continuous value, the support vector machine, the random forest and the neural network machine learning calculation can establish a corresponding regression model and continue to have the specific advantages of various algorithms. The Adaboost regressor is an iterative algorithm, also known as reinforcement learning or boosting learning. And performing optimal parameter search on the training set by using a ten-fold cross validation method through a Hyperopt package. After the optimal parameters are obtained, designing an optimal parameter new training model on the full training set, and evaluating the model prediction efficiency on the test set.

The study uses R2, R, MAE, MSE indices for machine learning regression model performance evaluation. And respectively drawing a scatter diagram by taking the equivalent spherical power predicted by the machine learning model and the equivalent spherical power after the real cycloplegia as the abscissa and ordinate axes, and calculating the proportion of samples accounting for the test lumped samples, wherein the difference between the equivalent spherical power predicted by the machine learning model and the equivalent spherical power after the real cycloplegia is less than 0.50D. And predicting whether the distribution condition of the equivalent spherical power is different from the distribution condition of the equivalent spherical power after real cycloplegia by using a wilcoxon symbolic rank sum test analysis machine learning model.

The decision coefficient R2 is the second power of the correlation coefficient R, and the greater the model goodness of fit, the greater the degree of interpretation of the dependent variable by the independent variable. When the average absolute error MAE is calculated, the error is calculated to be an absolute value, then the average is calculated, the mean square error MSE is similar to the absolute error, but the influence of the prediction abnormal value on the index value is amplified by the power operation, and the prediction value is more sensitive to abnormality.

The index calculation formula is as follows:

total sample amount m, true equivalent sphere power of𝑦The machine learning model predicted value is

。

Establishing a two-classification, three-classification and regression prediction model by using juvenile ophthalmic examination database data and combining a machine learning algorithm, and predicting whether the child is necessary to implement cycloplegia and post-paralysis refractive status or not by using parameters of the anterior eye part of the cycloplegia and the basic information of the child. And optimizing steps such as z-score standardization, missing value interpolation, data balance and the like are implemented in the modeling training process to improve the model prediction performance.

In the training and testing of the model of one embodiment, a teenager eye health examination database of Tianjin medical university is used, and the database has about 4600 samples, so that the requirements of model training and independent testing can be met. Other similar eye health check databases may also be used to train and test the model. As shown in fig. 1, the input characteristics for the three models were identical, including pre-cycloplegic ciliary muscle for the sample eye and the contralateral eye: sphere, cylinder, astigmatism, equivalent sphere, axial length, corneal thickness, anterior depth, crystal thickness, corneal mean K value, corneal steepness K value, accommodative lag, intraocular pressure, vision, and IPO/CCT, and further including subject age, gender, whether glasses are worn, eye position, and sample eye SE-contralateral eye SE. And performing a sample ophthalmoplegia retroflexion examination to obtain SE difference before and after cycloplegia (greater than 0.50D is regarded as the need of cycloplegia), and the cycloplegia retroflexion state and the cycloplegia equivalent sphericity data are respectively used as labels of three training tasks, so as to train and test a two-classification/three-classification model and a regression model.

Example 1:

and collecting data of 19 patients according to characteristics of the cycloplegic muscle required by the model, and obtaining actual cycloplegic retrodiopter degree, refractive state and cycloplegic muscle requirements of the patients. The pre-paralytic features are input into the trained model to obtain the prediction results of the three tasks.

TABLE 1

Note: the sex place 0 is female, and 1 is male; the lens information position 0 is an unmated lens, and 1 is a matched lens;

TABLE 2

TABLE 3

TABLE 4

TABLE 5

The data in tables 1-5 above are input into the trained EEC model, and the following results are obtained.

TABLE 6

Note: whether cycloplegia is needed or not/whether the model predicts whether the cycloplegic part needs 0 as unnecessary or 1 as needed;

as can be seen from the data in Table 6, the data predicted by the model substantially corresponds to the results of the clinical procedure.

Through a large number of data experiments, the performance of each model is detected, and the prediction efficiency of each model is better. In the binary model, each prediction model ACC was about 0.80 and AUC about 0.83. The highest AUC of the ECC model is 0.86, the highest sensitivity is 0.73, the specificity is 0.80, and the identification capability of the ECC model to positive and negative samples is strong, as shown in FIG. 3. In the three classification models, the prediction efficiencies of the machine learning prediction models are similar, ACC is between 0.82 and 0.85, precision is between 0.79 and 0.83, recall is between 0.78 and 0.84, F1-score is between 0.78 and 0.83, the ratio of each refraction state predicted by the machine learning model and the ratio of each refraction state after cycloplegia have no statistical difference, and a scatter diagram is drawn by taking the equivalent sphere power after paralysis and the equivalent sphere power predicted by the model as horizontal and vertical coordinates respectively in a regression model, as shown in FIG. 4. As shown in fig. 5, the sample ratio of the predicted value to the value after paralysis within the range of 0.50D is 81.3%, and the data distribution of the equivalent sphere power predicted by the machine learning model has no statistical difference with the data distribution of the equivalent sphere power after cycloplegic.

The embodiments of the present invention have been described in detail, but the description is only for the preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. All equivalent changes and modifications made within the scope of the present invention shall fall within the scope of the present invention.

Claims

1. A method for predicting the paralytic demand of ciliary muscle and predicting the refractive state of post-paralytic diopter, characterized in that the method comprises the following steps:

measuring and obtaining numerical parameters before cycloplegia;

and summarizing the calculated values of the three classification models and the regression model, and performing statistical analysis on the calculated values to obtain the distribution of the unbiased refraction state and the distribution of the diopter in the test target.

2. The method for demand prediction of cycloplegia and prediction of refractive state of retroplegic refraction according to claim 1, wherein: the numerical parameters before cycloplegia comprise optical biological data, optometry data and basic information.

3. The method for demand prediction of cycloplegic and prediction of refractive state with post-paralysis refractive power as claimed in claim 2, wherein:

the optical biological data comprises one or more of ocular axial length, corneal thickness, anterior chamber depth, crystal thickness, vitreous cavity depth, and corneal curvature values;

the basic information comprises age and/or gender;

the other examinations include one or more of intraocular pressure, accommodative lag, eye position, living vision, and on-lens information.

4. The method for predicting the paralytic demand of ciliary muscle and predicting the paralytic post-refractive power as set forth in any one of claims 1-3, wherein: the binary model is one of SVC _ W, RF _ W, DNN _ W and EEC, the model performance indexes are Accuracy, Sensitivity and Specificity, and the index calculation formula is as follows:

5. The method for predicting the paralytic demand of ciliary muscle and predicting the paralytic post-refractive power as set forth in any one of claims 1-3, wherein: the three-classification model is one of SVC _ W, RF _ W, DNN _ W and EEC, the model performance indexes are Precision, Recall and F1Score, and the index calculation formula is as follows:

6. The method for predicting the paralytic demand of ciliary muscle and predicting the paralytic post-refractive power as set forth in any one of claims 1-3, wherein: the regression model is one of SVR, RFR, DNN and Adaboost, and the index calculation formula is as follows:

wherein, the total amount m of the sample and the real equivalent sphere power are𝑦The machine learning model predicted value is

。