Introduction

Myopia is currently widely regarded as a significant public health issue, leading to substantial vision loss and serving as a risk factor for a range of other serious ocular diseases. It is estimated that by 2050, 4.758 billion people (49.8% of the world population) and 938 million people (9.8% of the world population) will suffer from myopia and high myopia, respectively []. A recent meta-analysis study proposed that the global economic burden due to productivity losses from uncorrected myopia and myopic macular degeneration is estimated to reach US $250 billion []. Therefore, the prevention of high myopia as well as the diagnosis and treatment of pathological myopia remain a formidable societal challenge.

High myopia is defined as the spherical equivalent ≤–6.0 diopter [] when the accommodation of the eye is relaxed. However, increased severity of myopia and elongation of the eye’s axial length could alter the posterior segment structures, causing posterior scleral staphyloma, myopic macular degeneration, and optic neuropathy related to high myopia, potentially leading to the loss of best-corrected visual acuity []. High myopia-related fundus lesions stand as an important contributing factor to blindness across the world as well as in China []. The detection of high myopia hinges primarily on artificial auxiliary techniques, like refraction detection, fundus examination, measurement of axial length, and fundus photography. Nevertheless, manual examination and analysis by ophthalmologists are still essential, necessitating a significant investment of time and effort []. Additionally, in regions with limited medical resources, the shortage of ophthalmologists and medical equipment impedes the early and accurate identification of high-risk patients with high myopia, resulting in missed opportunities for optimal treatment. Therefore, forecasting the risk of high myopia and precisely diagnosing pathological myopia are currently major research focus.

With the rapid advances in computing technology and the ongoing refinement of statistical theory, machine learning (ML) has gradually been promoted and applied in clinical practice. For instance, ML can not only improve image quality, reduce misregistration, and simulate attenuation correction imaging in core cardiology [], but also be used for cancer screening (detection of lesions), characterization and grading of tumors, and prognosis prediction, thus facilitating clinical decision-making []. Since fundus images are noncontact, noninvasive, low-cost, easily accessible, and easy to process, ML has been extensively used to diagnose common retinal diseases, including diabetic retinopathy [-], macular degeneration [], and glaucoma [-]. ML has been applied to various image-processing tasks. Novel techniques for analyzing fundus images of high myopia and pathological myopia are continuously emerging [,]. However, the accuracy of these ML detections has not been systematically studied. Consequently, the present study was executed to comprehensively describe the accuracy of ML in detecting different degrees of lesions in high myopia, furnishing an evidence-based reference for subsequent lesion management.

MethodsStudy Registration

This study was implemented as per the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines and prospectively registered with PROSPERO (ID: CRD42023470820). The PRISMA checklist is available in .

Inclusion and Exclusion Criteria

We established detailed inclusion and exclusion criteria for this systematic review. To enhance visualization, these criteria are presented in tabular form ().

Textbox 1. Inclusion and exclusion criteria.

Inclusion criteria

Study type: (1) case-control, cohort, nested case-control, and case-cohort studies and (2) studies reported in English.Machine learning (ML): studies that fully constructed ML models for the prediction or diagnosis of high myopia, the diagnosis of pathological myopia, or the diagnosis of high myopia-associated glaucoma.Outcome measures: at least one of the following outcome indicators were reported: receiver operating characteristic (ROC), c-index, sensitivity, specificity, accuracy, recovery rate, accuracy rate, confusion matrix, F1-score, and calibration curve.Datasets: (1) some studies lacked independent validation sets, and only k-fold cross-validation was leveraged to verify the effect of the constructed mode; and (2) in some publicly available datasets, particularly those involving medical imaging, different studies have reported the efficiency of varying ML methods.

Exclusion criteria

Study type: (1) meta, review, guide, expert opinion; and (2) studies with too few samples (less than 20 cases).ML: literature that only executed the risk factor analysis but did not develop a complete ML mode.Outcome measures: none of the following outcomes were reported: ROC, c-index, sensitivity, specificity, accuracy, recovery rate, accuracy rate, confusion matrix, F1-score, and calibration curve.Data Sources and Search Strategy

PubMed, Cochrane, Embase, and Web of Science were thoroughly retrieved up to September 3, 2023, using the form of MeSH (Medical Subjects Headings) + free term, without any restrictions on region or publication period. The specific search strategy is depicted in .

Study Selection and Data Extraction

Duplicates were excluded from the retrieved literature, and titles and abstracts were reviewed to delete obviously irrelevant studies. The full texts of the remaining studies were then downloaded and thoroughly read to determine the final included studies in the systematic review. A standard electronic data extraction spreadsheet was prepared prior to extracting data. The extracted data encompassed the title, first author, type of study, year of publication, author’s country, patient source, target event, number of cases of the target event, the total number of cases, number of training set cases, the total number of training set cases, method of validation set generation, number of events in the validation set, total number of cases in the validation set, type of models, and modeling variables.

Two researchers (HZ and LF) independently screened the literature and extracted data. Upon completion, their findings were cross-checked. A third reviewer (JH) was consulted for resolution in case of any dissents.

Risk of Bias in Studies

The risk of bias in the eligible studies was appraised by two independent reviewers (HZ and LF) using the prediction model risk of bias assessment tool []. This tool is comprised of a large number of questions in four domains (participants, predictors, outcomes, and analysis), which reflect overall bias risk and applicability. The 4 domains involve 2, 3, 6, and 9 specific questions, respectively, and each question may be answered by yes or probably yes, no or probably not, or no information. Following the quality evaluation, a cross-check was carried out. In the event of any disputes, a third researcher (JH) was consulted for resolution.

Synthesis Methods

In some of the original studies included in our research, there was not only 1 validation set. Therefore, the number of models included in the meta-analysis does not equal the number of studies. The meta-analysis of sensitivity and specificity was executed using a bivariate mixed-effects model []. Sensitivity and specificity were meta-analyzed as per the diagnostic 2×2 table. However, most included studies did not provide the diagnostic 2×2 table. In such cases, the following two approaches were used to calculate the diagnostic 2×2 table: (1) it was computed based on sensitivity, specificity, and precision, combined with the number of cases; and (2) sensitivity and specificity were extracted based on the optimal Youden index, and then combined with the number of cases for calculation. The meta-analysis was implemented using R (version 4.2.0; R Foundation for Statistical Computing).

ResultsStudy Selection

A total of 4214 records were retrieved from the databases, of which 582 were duplicates. After reading the titles and abstracts, 3561 studies unrelated to ML in high myopia were excluded, leaving 71 studies. Of these, 13 only conducted image segmentation without constructing ML models, 5 did not provide full extractable outcome indicators, and 8 analyzed risk factors. Ultimately, 45 studies were incorporated into this review. The literature screening process is depicted in .

‎Figure 1. Flowchart of literature screening. Study Characteristics

The included studies were published from 2010 to 2023. Four of the studies [-] were about the prediction of high myopia, and the predicted variables were mainly derived from life characteristics, environmental and genetic factors, and routinely interpretable ocular clinical characteristics. Five of the studies [-] were about the diagnosis of high myopia, of which 1 study [] also involved the diagnosis of pathological lesions of high myopia. Six studies focused on the diagnosis of high myopia-associated glaucoma [-]. Out of the included studies, 31 studies focused on the diagnosis of pathological myopia, primarily using optical coherence tomography and fundus imaging to construct artificial intelligence models. Of these, 26 studies [,,,-] were based on DL (deep learning), while 5 studies [-] required manually coded ML for construction. Additionally, it was noted that in the 45 original studies, all 45 studies included binary classification tasks, with 9 studies [,,,,,,,,] additionally incorporating multiclassification tasks. Regarding validation methods, 31 studies provided an external validation set, and 23 used a combination of internal and external validation sets. In terms of the generation method of validation set, 6 studies [,,,,,] used k-fold cross-validation, 29 [,-,-,-,,,,-,] used random sampling, and 6 [,,,,,] applied a combination of k-fold cross-validation and random sampling. The detailed characteristics of the eligible studies are shown in and .

Table 1. Fundamental features of included studies.First authorYear of publicationCountry of authorsStudy typePatient sourceTarget eventsTotal number of casesTang et al []2022China, United StatesRetrospective studyMulticenterDiagnosis of pathologic myopia1395 fundus photographs, 895 patientsLi et al []2023ChinaNested case-control studySingle centerDiagnosis and prediction of pathological myopia20,870 patientsDu et al []2021ChinaRetrospective studySingle centerDiagnosis of pathologic myopia313 patients with high myopia and 457 eyesFoo et al []2023SingaporeProspective studyMulticenterPrediction of high myopia965 children with 1878 eyes and 7456 fundus photographsKim et al []2021KoreaRetrospective studyMulticenterDiagnosis of pathologic myopia860 eyesZhang et al []2013SingaporeRetrospective studyRegistry databaseDiagnosis of pathologic myopia2258 patientsZhu et al []2023ChinaRetrospective studySingle centerDiagnosis of pathologic myopia6078 photographsWu et al []2022ChinaRetrospective studySingle centerDiagnosis of pathologic myopia1853 photographsYe et al []2021ChinaRetrospective studySingle centerDiagnosis of pathologic myopia1041 patients with pathologic myopia and with 2342 eligible OCTa macular imagesWang et al []2023ChinaRetrospective studySingle centerDiagnosis of pathologic myopia7606 patients with 10,347 fundus photographsWang et al []2022ChinaProspective, longitudinal, observational studyWenzhou large-scale surveyPrediction of myopia and high myopia15,765 patientsWan et al []2021ChinaRetrospective studySingle centerDiagnosis of pathologic myopia858 photographsWan et al []2023ChinaRetrospective studySingle centerDiagnosis of pathologic myopia1750 photographsTan et al []2021SingaporeRetrospective multicohort studyMulticenter + registry databaseDiagnosis of high myopia + pathological myopia125,421 patients with 251,349 photographsSun et al []2023ChinaRetrospective multicohort studyMulticenter + registry databaseDiagnosis of pathologic myopia1514 fundus photographsSogawa et al []2020JapanRetrospective studySingle centerDiagnosis of pathologic myopia910 patients with 910 imagesDu et al []2022JapanRetrospective studySingle centerDiagnosis of pathologic myopia1327 patients with 2400 high myopia eyes and 9176 OCT imagesHou et al []2023ChinaProspective cohort studySingle centerDiagnosis of pathologic myopia576 patientsLi et al []2022ChinaRetrospective cohort studyMulticenterPathologic myopia29,230 patients with 57,148 fundus photographsLi et al []2021ChinaCase-control studyMulticenterDiagnosis of glaucoma in high myopia2731 participants with 2731 eyesChen et al []2019ChinaProspective studySingle centerPrediction of high myopia1063 patientsChoi et al []2021KoreaRetrospective studySingle centerPrediction of high myopia492 patients with 690 eyesCui et al []2021China, TaiwanRetrospective studyRegistry databaseDiagnosis of pathologic myopia800 imagesGuan et al []2023ChinaRetrospective studyMulticenterPrediction of high myopia1,285,609 participantsHe et al []2022ChinaRetrospective studyMulticenterDiagnosis of pathologic myopia2866 patients with 3945 OCT imagesHemelings et al []2021BelgiumRetrospective studyRegistry databaseDiagnosis of pathologic myopia1200 photographsRauf et al []2021PakistanRetrospective studyRegistry databaseDiagnosis of pathologic myopia840 photographsPark et al []2022KoreaRetrospective studySingle centerDiagnosis of pathologic myopia367 eyesLu et al []2021ChinaRetrospective studySingle centerDiagnosis of pathologic myopia and diagnosis of pathologic myopia17,330 photographs
17,330 photographs
Lu et al []2021ChinaRetrospective studyMulticenterDiagnosis of pathologic myopia32,419 patients with 37,659 imagesLiu et al []2010SingaporeRetrospective studySingle centerPathologic myopia80 photographsLi et al []2022China, United StatesRetrospective studySingle centerDiagnosis of pathologic myopia1139 patients with 5917 imagesLee et al []2023KoreaRetrospective studySingle centerDiagnosis of glaucoma in high myopia260 eyes and 260 imagesKim et al []2023KoreaRetrospective studySingle centerDiagnosis of glaucoma in high myopia2607 eyesJeong et al []2023KoreaRetrospective cross-sectional studySingle centerDiagnosis of glaucoma in high myopia274 patientsHuang et al []2022ChinaCase-control studySingle centerPrediction of high myopia1298 patientsHuang et al []2023China, United KingdomRetrospective studySingle centerDiagnosis of pathologic myopia1131 patients with 3441 imagesDu et al []2021JapanRetrospective studySingle centerdiagnosis of pathologic myopia4432 eyes and 7020 imagesCrincoli et al []2023ItalyCase-control studyMulticenterdiagnosis of pathologic myopia84 patients with 84 eyes and 252 photographsAsaoka et al []2014JapanCase-control studyMulticenterDiagnosis of glaucoma in high myopia242 patients and 242 eyesBowd et al []2023United States, GermanyRetrospective studySingle centerDiagnosis of glaucoma in high myopia593 eyesZhao et al []2022ChinaRetrospective studySingle centerPrediction of high myopia546 patientsLiu et al []2010SingaporeRetrospective studySingle centerDiagnosis of pathologic myopia80 photographsDai et al []2020ChinaRetrospective studySingle centerPrediction of high myopia319 patients with 932 imagesBaid et al []2019IndiaRetrospective studyRegistry databaseDiagnosis of pathologic myopia481 photographs

aOCT: optical coherence tomography.

Table 2. Fundamental features of included studies.Total number of cases in training setGeneration of validation setTotal number of cases in validation setTotal number of cases in test setModel typeModeling variables727 fundus photographs5-fold cross-validation + random sampling238 fundus photographs238 fundus photographsDLaFundus photographs2069 patientsRandom sampling1382 patientsUnclearACPb, MLcClinical features319 eyesRandom sampling138 eyesUnclearML-based
radiomics analysis methodFundus photographs769 children with 1502 eyes and 5945 photographsInternal validation (5-fold cross-validation + random sampling) + multicenter external validation196 children with 376 eyes and 1511 fundus photographs99 children with 189 eyes and 821 photographsDLFundus photographs + clinical features602 eyesRandom sampling258 eyesunclearSVMd, MLFundus photographs2258 patientsStratified 20-fold cross-validationunclearunclearSVM, MLSNPe + clinical features + fundus photographs4252 photographsStratified 20-fold cross-validationunclear1826 photographsDLFundus photographs1483 photographsRandom samplingunclear370 photographsDLFundus photographs1874 photographsInternal validation (random sampling) + external validation（multicenter）468 photographs450 photographsDLFundus photographs5003 patients with 7389 photographsRandom sampling775 patients with 821 photographs1828 patients with 2137 photographsDLFundus photographs11,350 patientsInternal validation (random sampling) + external validation(prospective)4415 patients6168 patients (prognostic cohort)LRf, GBDTg, NNhClinical features758 photographs5-fold cross-validation + random sampling100 photographsUnclearDLFundus photographs1402 photographsRandom sampling174 photographs174 photographsDLFundus photographs226,686 photographsInternal validation (random sampling) + external validation（multicenter）11,303 photographs213,475 photographsDLFundus photographs400 fundus photographsMulticenter400 fundus photographs714 fundus photographsDLFundus photographsUnclear5-fold cross-validationUnclearUnclearDLFundus photographs7865 photographsrandom sampling1311 photographsUnclearDLFundus photographs516 patients10-fold cross-validation + random sampling60 patientsUnclearXGBoosti, SVM, LRClinical features + metabolic characteristics29,213 photographsInternal validation (random sampling) + external validation（multicenter）7302 photographs16,554 photographsDCNNj, DLFundus photographs2223 participants with 2223 eyesRandom sampling508 participants with 508 eyesUnclearFCNkOCTl images + clinical features638 patientsRandom sampling425 patientsUnclearLRGenetic factors + clinical features434 patients with 600 eyes and 1200 images5-fold cross-validationUnclear58 patients with 90 eyes and 180 imagesCNNm, DLOCT images400 imagesRandom sampling200 images200 imagesDLFundus photographs1600 participantsInternal validation (5-fold cross-validation)Unclear400 patientsRFn, LR, SVMClinical features2380 imagesRandom sampling680 photographs340 photographsDLOCT images400 photographsRandom sampling400 photographs400 photographsDLFundus photographs400 photographs10-fold cross-validation + random sampling40 photographs400 photographsDLFundus photographs293 eyesrandom sampling37 eyes37 eyesDL3D OCT images11,502 photographsUnclear3284 photographs1642 photographsDLFundus photographs2457 photographsUnclear707 photographs372 photographsDLFundus photographs32,010 imagesInternal validation (5-fold cross-validation) + external validation (multicenter)Unclear732 patients with 1000 imagesDLFundus photographs40 photographsRandom samplingUnclear40 photographsSVM, DLFundus photographs838 patients with 4338 imagesInternal validation (random sampling) + external validation (prospective)210 patients with 1167 photographs91 patients with 174 eyes and 412 photographsDLOCT macular images165 imagesRandom sampling46 photographs49 photographsDLOCTAo and OCT images1416 eyesInternal validation (random sampling) + external validation471 eyes720 eyesDLOCT imagesUnclearUnclearUnclearUnclearDecision treeOCT images325 patientsRandom sampling973 patientsUnclearDLGenetic + clinical features2264 imagesInternal validation (random sampling) + external validation (prospective)501 photographs604 photographsDLOCT images4140 photographsRandom sampling1036 photographs1844 photographsDLFundus photographs176 photographsRandom sampling25 photographs51 photographsDLOCT imagesUnclearUnclearUnclearUnclearRFHRT parameters347 eyes5-fold cross-validation + random sampling87 eyes159 eyesCNNOCT images928 fundus photographsRandom sampling232 photographsUnclearDLFundus photographs40 photographsRandom samplingUnclear40 photographsSVM or DLFundus photographs792 photographsRandom samplingUnclear140 photographsDLFundus photographs374 photographsRandom sampling80 photographs27 photographsCNNFundus photographs

aDL: deep learning.

bACP: algorithm of conditional probability.

cML: machine learning.

dSVM: support vector machine.

eSNP: single nucleic polymorphism.

fLR: logistic regression.

gGBDT: gradient boosted decision tree.

hNN: neural network.

iXGBoost: extreme gradient boosting

jDCNN: deep convolutional neural networks

kFCN: fully connected network

lOCT: optical coherence tomography.

mCNN: convolutional neural networks

nRF: random forest

oOCTA: optical coherence tomography angiography.

Risk of Bias in Studies

This review incorporated 67 models. There were 36 retrospective studies [,,-,-,-,-,-,-,] that constructed 39 models, indicating a high bias in the choosing of study participants. Five case-control studies [,,,,] constructed 13 models, also showing high bias in the selection of study participants. Since the predictors were evaluated in the context of a known outcome in the case-control studies, there was a high bias in the assessment of predictive factors. Twelve studies [,,,,,,,-] constructed 22 models based on manually coded ML, with a high bias in predictive factors. In terms of statistical analysis, 2 studies [,] with 5 models did not meet the requirement of having an event per variable>20%, indicating a high risk of bias. In the statistical analysis, 32 models in 34 studies [,,,-,,,,,-,-,] could not estimate event per variable due to the use of the DL method. Additionally, 10 studies [,,,,,,,-] with 29 models in ML did not report on the complexity of the data, rendering it difficult to determine their bias risk. Five studies [,,,,] with 11 models were identified as having a high risk of bias in statistical analysis because they did not perform cross-validation to adjust the stability of models with different parameters. In summary, in terms of research participants, 14 models had a low risk of bias; 52 models had a high risk of bias, and 1 model had an unclear risk of bias. In terms of predictors, 37 models had a low risk of bias and 30 models had a high risk of bias. In terms of outcomes, all 67 models had a low risk of bias. In terms of statistical analysis, 3 models had a low risk of bias, 16 models had a high risk of bias, and 48 models had an unclear risk of bias.

Meta-Analysis of ML for Binary Classification TasksPathological Myopia

Twenty studies [,-,,,,-,,,,] reported ML for diagnosing pathological myopia. Modeling algorithms included algorithms of conditional probability, support vector machines (SVMs), logistic regression (LR), extreme gradient boosting, convolutional neural networks (CNNs), and deep convolutional neural networks (DCNNs). The overall sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), and summary receiver operating characteristic (SROC) were 0.91 (95% CI 0.89-0.92), 0.95 (95% CI 0.94-0.97), 19.7 (95% CI 13.8-28.2), 0.10 (95% CI 0.08-0.12), 201 (95% CI 122-331), and 0.97 (95% CI 0.95-0.98), respectively. The Deek funnel plot indicated no substantial evidence of publication bias in the included ML models. Assuming that the prior probability of pathological myopia was 20% if the result of ML was pathological myopia, then the probability of true pathological myopia would be 83%. If the result of ML was nonpathological myopia, then the probability of true pathological myopia would be 2% (ie, the probability of true nonpathological myopia was 98%; and Figures S1-S3 in ).

‎Figure 2. Forest plot for the meta-analysis of sensitivity and specificity of machine learning in detecting pathological myopia [,-,,,,-,,,,]. Note: the pooled sensitivity and specificity of 44 models from 20 machine learning studies on the diagnosis of pathological myopia were 0.91 (95% CI 0.89-0.92) and 0.95 (95% CI 0.94-0.97), respectively.

Five studies [,,,,] reported conventional ML (non-DL) for diagnosing pathological myopia. Modeling algorithms included algorithms of conditional probability, SVM, extreme gradient boosting, and LR. The overall sensitivity, specificity, PLR, NLR, DOR, and SROC curve were 0.77 (95% CI 0.69-0.84), 0.85 (95% CI 0.75-0.92), 5.2 (95% CI 2.8-9.8), 0.27 (95% CI 0.18-0.39), 20 (95% CI 7-51), and 0.86 (95% CI 0.75-0.92), respectively. The Deek funnel plot indicated the presence of publication bias in the conventional ML (non-DL) models. Assuming that the prior probability of pathological myopia for conventional ML (non-DL) was 20% if the result of conventional ML (non-DL) was pathological myopia, then the probability of true pathological myopia would be 57%. If the result of conventional ML (non-DL) was nonpathological myopia, then the probability of true pathological myopia would be 6% (ie, the probability of true nonpathological myopia was 94%; and Figures S4-S6 in ).

‎Figure 3. Forest plot for the meta-analysis of sensitivity and specificity of conventional machine learning (non-deep learning) in detecting pathological myopia [,,,,]. Note: the pooled sensitivity and specificity of 6 models from 5 conventional machine learning (non-deep learning) studies on the diagnosis of pathological myopia were 0.77 (95% CI 0.69-0.84) and 0.85 (95% CI 0.75-0.92), respectively.

Fifteen studies [,-,,,,-,] mentioned DL for diagnosing pathological myopia. Modeling algorithms included CNN and DCNN. The overall sensitivity, specificity, PLR, NLR, DOR, and SROC were 0.92 (95% CI 0.90-0.93), 0.96 (95% CI 0.95-0.97), 23.7 (95% CI 16.5-34.0), 0.09 (95% CI 0.07-0.11), 271 (95% CI 168-437), and 0.97 (95% CI 0.95-0.98), respectively. The Deek funnel plot revealed no remarkable publication bias in the DL models. Assuming that the prior probability of pathological myopia for DL was 20% if the result of DL was pathological myopia, then the probability of true pathological myopia would be 86%. If the result of DL was nonpathological myopia, then the probability of true pathological myopia would be 2% (ie, the probability of true nonpathological myopia was 98%; and Figures S7-S9 in ).

‎Figure 4. Forest plot for the meta-analysis of sensitivity and specificity of deep learning in detecting pathological myopia [,-,,,,-,]. Note: the pooled sensitivity and specificity of 38 models from 15 deep learning studies on the diagnosis of pathological myopia were 0.92 (95% CI 0.90-0.93) and 0.96 (95% CI 0.95-0.97), respectively. High Myopia

Six studies [,,,-] discussed ML for diagnosing and forecasting high myopia. Modeling algorithms included DCNN, CNN, LR, SVM, random forest (RF), and linear mixed models. The sensitivity, specificity, PLR, NLR, DOR, and SROC were 0.94 (95% CI 0.90-0.96), 0.94 (95% CI 0.88-0.97), 16.2 (95% CI 7.7-33.8), 0.06 (95% CI 0.04-0.11), 255 (95% CI 79-822), and 0.98 (95% CI 0.96-0.99), respectively. The Deek funnel plot indicated no substantial evidence of publication bias in the included ML models. Assuming that the prior probability of high myopia for ML was 20% if the result of ML was high myopia, then the probability of true high myopia would be 80%. If the result of ML was non-high myopia, then the probability of true high myopia would be 2% (ie, the probability of true non-high myopia was 98%; and Figures S10-S12 in ).

Three studies [,,] focused on diagnosing high myopia, while 3 studies [,,] focused on predicting high myopia. Due to the limited number of studies included, we did not perform a meta-analysis for the diagnostic and prediction tasks. In the validation sets of the diagnostic tasks, sensitivity ranged from 0.91 to 1.00 and specificity ranged from 0.85 to 1.00, while in the validation sets of the prediction tasks, these values were 0.85-0.94 and 0.86-0.94, respectively. We found that both diagnostic and prediction tasks demonstrated highly favorable performance.

‎Figure 5. Forest plot for the meta-analysis of sensitivity and specificity of machine learning in detecting high myopia [,,,-]. Note: the pooled sensitivity and specificity of 9 models from 6 machine learning studies on the diagnosis and prediction of high myopia were 0.94 (95% CI 0.90-0.96) and 0.94 (95% CI 0.88-0.97), respectively. High Myopia–Associated Glaucoma

Six studies [-] mentioned ML for diagnosing high myopia-associated glaucoma. Modeling algorithms included Lagrange multiplier, fully connected network, radial basis function network, decision tree, RF, and CNN. The sensitivity, specificity, PLR, NLR, DOR, and SROC curve were 0.92 (95% CI 0.85-0.96), 0.88 (95% CI 0.67-0.96), 7.6 (95% CI 2.4-23.8), 0.09 (95% CI 0.04-0.20), 84 (95% CI 13-555), and 0.96 (95% CI 0.94-0.97), respectively. The Deek funnel plot indicated no substantial evidence of publication bias in the included ML models. Assuming that the prior probability of high myopia–associated glaucoma was 20% if the result of ML was high myopia-associated glaucoma, then the probability of true high myopia–associated glaucoma would be 65%. If the result of ML was non-high myopia–associated glaucoma, then the probability of true high myopia–associated glaucoma would be 2% (ie, the probability of true non-high myopia–associated glaucoma was 98%; and Figures S13-S15 in ).

‎Figure 6. Forest plot for the meta-analysis of sensitivity and specificity of machine learning in detecting high myopia-associated glaucoma [-]. Note: the pooled sensitivity and specificity of 9 models from 6 machine learning studies on the diagnosis of high myopia-associated glaucoma were 0.92 (95% CI 0.85-0.96) and 0.88 (95% CI 0.67-0.96), respectively. Review of ML for Multiclassification Tasks

Out of the included studies, 9 [,,,,,,,,] used ML for multiclassification tasks. Due to significant variations in the diagnostic differences across these multiclassification tasks, a quantitative analysis was not feasible. Five studies [,,,,] focused on fundus images–based DL to detect different types of myopic atrophy maculopathy in high myopia, with an accuracy ranging from 88% to 97%. Two studies [,] used optical coherence tomography (OCT) image–based DL to detect different types of myopic traction maculopathy in high myopia, with an accuracy ranging from 91% to 96%. One study [] used fundus image–based DL to differentiate between normal, low-risk high myopia, and high-risk high myopia, with an accuracy of 99%. One study [] applied fundus image–based DL to distinguish between normal, fundus tessellation, and pathologic myopia, with an accuracy of 94%, as illustrated in .

Table 3. Results of machine learning for multiclassification tasks.First authorYearDiagnostic purposeTypes of artificial intelligenceModeling variablesGeneration of validation setAccuracy rate, %Tang et al []2022Classification of atrophic macular lesions in myopicCNNsa; DLbFundus photographs5-fold cross-validation + random sampling94Zhu et al []2023Classification of atrophic macular lesions in myopicNeural network; DLFundus photographsStratified 20-fold cross-validation90Wan et al []2021Normal, low, and high risk of high myopiaDCNNsc; DLFundus photographs5-fold cross-validation + random sampling99Wan et al []2023Classification of atrophic macular lesions in myopicDLFundus photographsRandom sampling95-97Sun et al []2023Classification of atrophic macular lesions in myopicDLFundus photographsExternal validation (multicenter)89.2Li et al []2022Differential diagnosis of normal, leopard print fundus, and pathological myopiaDCNN; DLFundus photographsInternal validation (random sampling) + external validation (multicenter)94He et al []2022Differential diagnosis of tractive macular degeneration and neovascular macular degeneration in high myopia, and othersDLOCTd imagesRandom sampling91-96Huang et al []2023Classification of tractive macular degeneration in high myopiaDLOCT imagesInternal validation (random sampling) + external validation (prospective)96Du et al []2021Classification of atrophic macular lesions in myopicDL