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 Multimedia Appendix 1.

We established detailed inclusion and exclusion criteria for this systematic review. To enhance visualization, these criteria are presented in tabular form (Textbox 1).

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 Multimedia Appendix 2.

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.

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 [16]. 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.

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 [17]. 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).

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.

The included studies were published from 2010 to 2023. Four of the studies [18-21] 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 [22-26] were about the diagnosis of high myopia, of which 1 study [22] also involved the diagnosis of pathological lesions of high myopia. Six studies focused on the diagnosis of high myopia-associated glaucoma [27-32]. 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 [4,15,22,33-55] were based on DL (deep learning), while 5 studies [56-60] 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 [4,33,34,38,39,49,50,52,61] 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 [23,24,34,40,47,59] used k-fold cross-validation, 29 [15,19-22,25-29,35-38,41,42,45,48-58,61] used random sampling, and 6 [4,18,32,33,44,60] applied a combination of k-fold cross-validation and random sampling. The detailed characteristics of the eligible studies are shown in Tables 1 and 2.

This review incorporated 67 models. There were 36 retrospective studies [4,15,22-26,28-30,32-42,44-50,52-55,57-59,61] that constructed 39 models, indicating a high bias in the choosing of study participants. Five case-control studies [21,27,31,51,56] 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 [19,20,23,24,27,30,31,56-60] constructed 22 models based on manually coded ML, with a high bias in predictive factors. In terms of statistical analysis, 2 studies [21,45] 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 [4,15,18,21-23,25,26,28,29,32-42,44-55,61] could not estimate event per variable due to the use of the DL method. Additionally, 10 studies [19,20,24,27,30,31,56,58-60] with 29 models in ML did not report on the complexity of the data, rendering it difficult to determine their bias risk. Five studies [20,27,30,31,60] 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.

Out of the included studies, 9 [4,33,34,38,39,49,50,52,61] 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 [33,34,38,39,50] 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 [49,61] 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 [4] 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 [52] applied fundus image–based DL to distinguish between normal, fundus tessellation, and pathologic myopia, with an accuracy of 94%, as illustrated in Table 3.

This study comprehensively described the accuracy of ML in detecting high myopia, high myopia-associated glaucoma, and pathologic myopia. ML demonstrated exceptionally favorable performance in detecting high myopia, while DL was highly accurate in diagnosing pathologic myopia.

Previous studies have also explored the detection accuracy of ML in this field. A systematic review has reported that fundus image– or OCT image–based DL can effectively diagnose and classify myopic maculopathy. Additionally, ML examination of the optic disc area can detect myopic maculopathy that may not be easily identified during clinical examination [14]. A recent meta-analysis based on only 11 studies evaluated the performance of DL in identifying pathological myopia based on fundus images. The SROC, specificity, and sensitivity were found to be 0.9905, 0.959 (95% CI 0.955-0.962), and 0.965 (95% CI 0.963-0.966), respectively [62]. In the previous meta-analysis, the 11 original studies all constructed fundus images-based DL models, and studies on conventional ML (non-DL) were not incorporated. The number of studies included in our review was further expanded, with a total of 20 studies on the performance of ML in diagnosing pathological myopia. Moreover, subgroup analysis was executed between conventional ML (non-DL) and DL. Our finding also indicated that DL demonstrated exceptionally favorable efficiency in detecting pathological myopia.

As the understanding of the etiology of myopia deepens, growing evidence reveals risk factors for the onset or progression of myopia, including age, sex, parental myopia, susceptibility genes, and outdoor activities. For high myopia, early prediction appears to be more beneficial. Among the included studies, one incorporated 135 myopia-related single nucleotide polymorphisms to forecast the progression and onset of high myopia. ML for the prediction of high myopia was mainly based on genetic factors, environmental factors, and ocular clinical characteristics. ML showed an SROC of 0.96, sensitivity of 0.91, and specificity of 0.87, respectively [20], suggesting that ML methods can effectively identify high-risk individuals with high myopia, thus effectively preventing this condition, especially in minors.

Glaucoma is a significant contributor to irreversible vision impairment and blindness all over the world. A 10-year study in Chinese individuals over the age of 40 years found that every 1 mm increase in axial length increased the risk of open-angle glaucoma by 1.72 times. In comparison to emmetropic and hyperopic eyes, highly myopic eyes had a 7.3 times higher risk of developing open-angle glaucoma [63]. Due to the changes in retinal structure caused by myopia, diagnosing glaucoma in myopic patients, especially those with high myopia, is challenging. Six studies were included to evaluate the diagnosis of high myopia glaucoma. Of them, 3 studies [28,29,32] used fundus OCT image-based DL techniques, while the remaining 3 [27,30,31] used non-DL ML (Lagrange multiplier, fully connected network, radial basis function network, decision tree, RF) approaches using OCT parameters, Heidelberg Retina Tomograph parameters, and ocular biometric parameters of patients. The findings indicated that ML yielded highly promising results in the detection of high myopia glaucoma.

It was also noted that different ML methods, conventional ML and DL, showed significant differences in their ability to identify positive or outcome events. Conventional ML is often used to construct models with interpretable clinical features. Lately, various image-based ML methods have emerged. However, a significant challenge in this context is the requirement for manual annotation to facilitate ML. From this standpoint, manual annotation poses a formidable barrier to effectively mitigating the risk of bias. DL, on the other hand, enables intelligent processing of medical images and has been widely applied in various fields, including detecting diabetic retinopathy [8-10], retinopathy of prematurity [64,65], age-related macular degeneration [10], and glaucoma [11-13]. With the rapid development of ML, imaging data are increasingly becoming a valuable source for medical analysis. Multiple studies have demonstrated that images from various sources, including fundus images [66], external eye appearance [67], and refractive images [68], can effectively estimate a patient’s spherical refractive error, indicating the potential of imaging data in predicting the risk of myopia. This study also finds that image-based DL is more accurate than conventional ML, providing a theoretical basis for the creation of future intelligent tools.

Additionally, the dataset used in ML demands considerable attention. Many studies are hampered by a limited number of cases, raising concerns about the robustness of the findings. Additionally, validation methods often depend heavily on internal validation, which may not fully capture the model’s generalizability. Hence, incorporating comprehensive patient data is essential for building a robust large-scale database, which will enable the development of ML models that are applicable to a broader population. Among the studies included, 7 [15,22,39,42,44,55,59] established ML models based on publicly available large databases.

Although our review includes a larger number of studies than previous meta-analyses and provides an evidence-based basis for subsequent studies, this study has limitations. First, there were few studies on the prediction of high myopia, which limits the interpretation of our results, and clinically interpretable variables for predicting high myopia were not explained. Second, we did not conduct a subgroup analysis on the type of ML (conventional ML vs DL) owing to the insufficient number of included studies based on high myopia glaucoma and high myopia. Third, the majority of the models included in this study were assessed as having a high risk of bias, which may impact the interpretation of our results. Most included studies adopted a retrospective design, which might lead to selection bias.

In conclusion, this study comprehensively reviews and meta-analyzes the performance of ML in the diagnosis and prediction of high myopia, high myopia-associated glaucoma, and pathological myopia, providing valuable guidance and references for future research. Challenges exist within the emerging field of myopia prediction. With the development of new analytical methods and the accumulation of real medical datasets, future research holds the promise of improving the prediction of myopia onset and progression. This advancement brings us closer to the ultimate goal of identifying high-risk individuals promptly and implementing targeted interventions in clinical practice.