Forecasting Myopic Maculopathy Risk Over a Decade: Development and Validation of an Interpretable Machine Learning Algorithm

Abstract

Purpose: The purpose of this study was to develop and validate prediction model for myopic macular degeneration (MMD) progression in patients with high myopia.

Methods: The Zhongshan High Myopia Cohort for model development included 660 patients aged 7 to 70 years with a bilateral sphere of ≤−6.00 diopters (D). Two hundred twelve participants with an axial length (AL) ≥25.5 mm from the Chinese Ocular Imaging Project were used for external validation. Thirty-four clinical variables, including demographics, lifestyle, myopia history, and swept source optical coherence tomography data, were analyzed. Sequential forward selection was used for predictor selection, and binary classification models were created using five machine learning algorithms to forecast the risk of MMD progression over 10 years.

Results: Over a median follow-up of 10.9 years, 133 patients (20.2%) showed MMD progression in the development cohort. Among them, 69 (51.9%) developed newly-onset MMD, 11 (8.3%) developed patchy atrophy from diffuse atrophy, 54 (40.6%) showed an enlargement of lesions, and 9 (6.8%) developed plus signs. Top six predictors for MMD progression included thinner subfoveal choroidal thickness, longer AL, worse best-corrected visual acuity, older age, female gender, and shallower anterior chamber depth. The eXtreme Gradient Boosting algorithm yielded the best discriminative performance (area under the receiver operating characteristic curve [AUROC] = 0.87 ± 0.02) with good calibration in the training cohort. In a less myopic external validation group (median −5.38 D), 48 patients (22.6%) developed MMD progression over 4 years, with the model's AUROC validated at 0.80 ± 0.008.

Conclusions: Machine learning model effectively predicts MMD progression a decade ahead using clinical and imaging indicators. This tool shows promise for identifying “at-risk” high myopes for timely intervention and vision protection.

High myopia is projected to impact a remarkable 9.8% of the world's population by 2050,¹ which translates to nearly one billion people at risk for myopic macular degeneration (MMD)-related vision loss, particularly in their fourth or fifth decade of life.^2–4 Therefore, identifying the risk of MMD development in patients with high myopia decades in advance is of paramount importance for prioritizing prevention and intervention for vision detriment. 

Previous longitudinal studies have primarily focused on a limited number of traditional demographic and ocular risk factors associated with MMD.^5–7 Nevertheless, they ignored other critical variables like information from optical coherence tomography (OCT) imaging. A series of recent investigations have highlighted the potential role of the choroidal thickness derived from the state-of-art swept source OCT (SS-OCT) in the development of MMD.^8–10 The choroidal thickness has been proposed to not only aid in diagnosing different stages of MMD,¹¹ but also to serve as a substantially independent predictor of MMD progression in highly myopic eyes over a 2-year period.¹² Moreover, the application of OCT techniques is essential in the detection of other vision-impairing macular changes associated with MMD, including neovascularization, retinal traction, and dome-shaped macula.¹³ Accordingly, the integration of multimodal data, especially the inclusion of OCT alongside traditional risk factors, is crucial to thoroughly identify the most pivotal predictors for MMD progression. 

Advancements in interpretable machine learning (ML) have thrust artificial intelligence into the limelight as an innovative solution for numerous healthcare challenges, enable to simultaneously combine multiple factors, identifying the most efficient and succinct models for outcome prediction without prior assumptions.^14,15 It has exhibited promising performance in predicting refraction progression and axial length (AL) elongation among myopic children.^16–18 A recent study underscored the effectiveness of ML in predicting visual acuity in highly myopic eyes, spotlighting its potential in high myopia management.¹⁹ However, there remains a knowledge gap in the deployment of ML for predicting long-term MMD progression. 

To fill the knowledge gap, this study aimed to develop and validate ML prediction model for long-term MMD progression over 10-years in the high myopia population. The model's generalizability was further evaluated through external validation on an independent cohort. 

Methods

This study utilized data from two independent cohorts, approved by the Institutional Review Board of Zhongshan Ophthalmic Center (references: 2012KYNL002, 2023KYPJ090, and 2017KYPJ094), with each participant providing written consent prior to inclusion. The study adhered to the principles of the Helsinki Declaration and followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) Statement.²⁰ 

Study Design and Population

The Zhongshan High Myopia Cohort (ISRCTN: 56368396) was formed as the development cohort.²¹ This prospective long-term study aimed to determine the progression trajectory of MMD and to derive the risk prediction system. This cohort included 890 patients with high myopia aged 7 to 70 years with bilateral cycloplegic sphere ≤−6.00 diopters (D) who lived in the area of 2-hour driving distance adjacent to the Zhongshan Ophthalmic Center, Guangzhou, China, between September 2011 and October 2012. Follow-up visits were scheduled every 2 years and the 10th year follow-up was conducted from May 2023 to September 2023. In the development cohort, we excluded participants with cataract or other retinopathy (n = 15), with history of intraocular surgery (n = 9), and absence of fundus photographs or ungradable images at baseline or during 10-year follow-up visit (n = 206), leaving 660 participants with available and assessable fundus color photographs for analysis (Supplementary Fig. S1). 

Participants with high myopia from an independent cohort, Chinese Ocular Imaging Project (COIP), was formed for external validation. This cohort consecutively enrolled healthy subjects and patients with diabetes from the Guangzhou community for detailed annual general and ocular imaging assessments aiming to explore the ocular phenotypic traits in adults.²² The baseline examinations were conducted in 2017 to 2019, and the fourth year of follow-up was conducted in 2022 to 2023. A total of 212 patients with high myopia aged 35 to 80 years with AL ≥25.5 mm²³ without any retinopathy except for MMD throughout the project were included (see Supplementary Fig. S1). The exclusion criteria were similar to participants from the development cohort: a history of intraocular surgery, major systematic or ocular diseases, and absence of fundus photographs. There was no sample overlap between the development cohort and external validation cohort. A summary of study procedures is presented in Supplementary Figure S2. 

Outcome

Classification of MMD severity was based on retinal photographs at each visit by two independent ophthalmologists (authors Y.C. and Y.W.) and a retinal specialist (author W.W.) masked to the participants’ characteristics according to the meta-analysis of pathologic myopia (META-PM) classification system in the reading center.²⁴ MMD was graded into no MMD (category 0 = C0); tessellated fundus only (category 1 = C1); diffuse chorioretinal atrophy (category 2 = C2); patchy chorioretinal atrophy (category 3 = C3); and macular atrophy (category 4 = C4). The “plus” lesions were also evaluated, which included lacquer cracks, Fuchs’ spot and myopic choroidal neovascularization (CNV). The interclass correlation coefficient (ICC) between two ophthalmologists for MMD grading was 0.87 (95% confidence interval [CI] = 0.75–0.94). OCT images were used to assist justification of suspicious signs from fundus photographs, such as CNV. Progression of MMD was defined as an increased MMD category (except for progression from C0 to C1), enlargement of diffuse/patchy chorioretinal atrophy or macular atrophy, first occurrence or increased area of any “plus” signs, or an increased signs number from baseline to follow-up visits.^12,25 

Candidate Features

Five categories of multimodal features for patients with high myopia were selected, encompassing 35 baseline variables: 3 demographic, 4 lifestyle, 4 history related, 14 ophthalmic, and 10 imaging measures (Supplementary Table S1). The same measure protocol was used in the development cohort and external validation cohort.^21,22 Demographics, lifestyle, and medical history were collected from standardized questionnaires administered by trained research staff through face-to-face interviews. Ocular biometrics were obtained by optical low-coherence reflectometry (Lenstar LS900, Haag-Streit AG, Koeniz, Switzerland). If the AL was longer than 32 mm, which exceeded the valid range of measurement in Lenstar, an IOLMaster (Carl Zeiss Meditec, Oberkochen, Germany) was used instead. An autorefractor (Topcon KR8800, Tokyo, Japan) measured the refraction after full cycloplegia by 0.5% tropicamide. Visual acuity was assessed using the Early Treatment Diabetic Retinopathy Study (ETDRS) LogMAR E visual acuity chart (Precision Vision, Villa Park, IL, USA). Imaging information included measures from the SS-OCT device (DRI OCT; Topcon, Tokyo, Japan) and color fundus photographs (Canon CX-1, Tokyo, Japan) under pupil dilation. The methodology of choroidal measure has been reported elsewhere.¹² Briefly, the 12-line 9 mm radial scan mode was performed centered on the fovea, with an average of 4 consecutive scan overlaps. The subfoveal choroidal thickness (SFCT) was automatically segmented and measured as the perpendicular distance between the outer choroid-sclera margin and the retinal pigment epithelium-Bruch's complex. Images were excluded with poor quality, motion artifacts, or segmentation failure. 

Model Development

All variables with missing data <20% were included (Supplementary Fig. S3). The development cohort was split into a training dataset for model training and an internal validation dataset for optimization with a ratio of 7:3. Missing data were imputed using multiple imputation by chained equation in the training, internal, and external validation datasets separately. Collinearity analysis was conducted by assessing the variance inflation factor (VIF) and tolerance among variables, with VIF >10 and tolerance <0.1 indicating significant collinearity. 

Predictors selection was determined through feature ranking and sequential forward selection strategy. Each feature was ranked based on its contribution to model performance measured by information gain, which can be regarded as the predictor's capacity to forecast the future progression of MMD. Considering the potential correlation between the spherical equivalent refraction (SER) and AL as well as between the SER and SFCT, we included AL, SFCT, and other candidate variables (except for SER) for predictor selection. The aim of predictor selection was to gain the stably high cumulative area under the receiver operating characteristics curve (AUROC). The decision to stop including additional predictors was based on reaching an overall AUROC of 0.85 in the initial model, as per previous ML studies, which is considered a prominent result.²⁶ The SHapley Additive exPlanations (SHAP)²⁷ plot was adopted to quantify and interpret the effect of predictor variables on MMD progression. 

Five ML algorithms, including eXtreme Gradient Boosting (XGBoost), logistic regression, random forest, LightGBM, and neural network, were used to predict MMD progression using the average performance of five-fold cross-validation. Hyperparameter optimization was tuned with five-fold cross-validation using the grid search approach that maximized the AUROC curve on the training dataset. Discrimination was evaluated using accuracy, sensitivity, specificity, F₁ score, AUROC, and area under the precision-recall curve (AUPRC) of the five-fold cross validation models with standard deviation (SD). Calibration was visualized through calibration plots. Clinical utility was assessed by the net benefit in decision curve analysis (DCA) to assess its potential clinical use.²⁸ 

Model Validation

In the validation phase, the model was verified in the internal validation dataset which was not included in the training phase. The model was validated in the external cohort to further assess its generalizability in a diverse population. The subgroup analysis in different age groups, including 7 to 18 (children and adolescents), 19 to 40 (young adults), and 41 to 70 (middle-aged adults) years old was conducted. Sensitivity analysis was performed in participants with bilateral cycloplegic sphere ≤−6.00 D from the external validation cohort. To enhance clinical practice, a web-based application was established with the Shiny package. 

Statistical Analysis

Right eye data were adopted in this study. The descriptive statistics were presented as mean ± SD or median (interquartile range [IQR]) for quantitative variables and number (percentage) for categorical variables. Continuous data were compared using unpaired t tests or Mann-Whitney tests, and categorical data were compared using chi-square tests or Fisher exact tests. A two-tailed P value of less than 0.05 was considered significant. All analyses were realized using R version 4.2.1 (R Foundation for Statistical Computing, www.R-project.org) and Python version 3.11.5 (Python Software Foundation). 

Results

Study Population

Table 1 shows the overall study participants’ characteristics. In the development cohort, a total of 660 patients with high myopia were included (462 in the training database and 198 in the internal validation database) with a mean age of 21.28 ± 11.87 years; 352 were female patients (53.3%). The baseline SER and AL were −8.75 (IQR = −10.75 to −7.50) D and 27.32 ± 1.44 mm. During a median follow-up time of 10.9 (IQR = 6.9–11.2) years, 133 (20.2%) patients with high myopia developed MMD progression. Among them, 69 (51.9%) patients developed newly-onset MMD, 11 (8.3%) developed patchy atrophy from diffuse atrophy, 54 (40.6%) showed an enlargement of present lesions, and 9 (6.8%) developed plus signs. Baseline characteristic comparisons between participants with and without MMD progression are shown in Supplementary Table S2. 

Table 1.

 

Baseline Characteristics of the Training, Internal, and External Validation Datasets

 

In the external validation cohort, 212 participants were included, with the mean age of 59.05 ± 10.37 years and 106 (50.0%) female participants. The participants showed the baseline SER of −5.38 (IQR = −6.75 to −3.50) D and AL of 26.50 ± 0.98 mm. Over 4 years, 48 (22.6%) experienced MMD progression (see Table 1). 

The population characteristics of age groups were demonstrated in Supplementary Table S3. The progression of MMD was observed in 14.6% among children and teenagers, 20.9% among young adults, and 51.7% among middled-aged adults within a decade. 

Predictors Selection and Variable Importance

A total of 32 candidate predictors showed positive contribution (importance >0) and were sorted based on their importance to the prediction task in the bar chart (Fig. 1A). The sequential forward selection scheme was demonstrated by the line chart that the cumulative AUROC climbed steeply with the inclusion of the first of several variables to 0.85, whereas it mildly descended or went flat with gentle fluctuation when additional variables were input in. Ultimately, the top six important variables were finally selected for modeling, comprising SFCT, AL, best-corrected visual acuity (BCVA), age, gender, and anterior chamber depth (ACD). The collinearity analysis results showed that the variance inflation factors were less than 5 and tolerance exceeded 0.2 (Supplementary Table S4). 

Figure 1.

 

Predictors selection and SHAP visualization in model training. (A) Sequential forward selection scheme was used for predictor selection. Bar plot demonstrated importance ranking of candidate predictors (left axis) and line chart showed the cumulative AUC upon the addition of predictors one by each iteration (right axis). Finally, the top six predictors (delineated in red) were included in ML development. (B) SHAP plots visualized the selected predictors. Thinner subfoveal choroidal thickness, longer axial length, worse best-corrected visual acuity, older age, female gender, and shallower anterior chamber depth were correlated with long-term MMD progression. SFCT exhibited the broadest SHAP horizontal range, indicating it possessed the most considerable prediction power and a significant impact on the model's output. MMD = myopic macular degeneration; SFCT = subfoveal choroidal thickness; SHAP = SHapley Additive exPlanations.

The SHAP plot visualized the varying contribution of the predictors to the proposed ML model (Fig. 1B). Each participant was represented as a data point, coded with gradient colors to signify the magnitude of the predictor's effect. The SFCT exhibited the widest SHAP horizontal range, indicating it possessed the most considerable prediction power and a significant impact on the model's output. Thinner baseline SFCT, longer AL, worse BCVA, older age, female gender, and shallower ACD were associated with higher risk of MMD progression. 

SFCT in MMD Progression

Figure 2 highlighted representative comparative cases of the outcome with initial SFCT. A highly myopic eye with thin baseline SFCT showed significant progression of MMD, whereas the eye with thick baseline SFCT exhibited stable MMD category without progression from serial fundus photographs over a 10-year period. 

Figure 2.

 

Serial fundus photographs of representative cases differentiated by baseline choroidal thickness and progression patterns. (A) A 9.31-year-old subject with an initial SFCT of 27 µm (AL = 29.58 mm; SER = −14 D), exhibited progressive MMD signs. This progression included diffuse atrophy enlargement and incident lacquer cracks in 2016, transition to patchy atrophy with increased lacquer cracks in 2018, onset of Fuchs' spots in 2020, and further lesion enlargement by 2023. (B) A 27.17-year-old subject with an initial SFCT of 205 µm (AL = 26.28 mm; SER = −8.375 D) showed no MMD progression of tessellated fundus throughout 11 years. AL = axial length; MMD = myopic macular degeneration; SER = spherical equivalent refraction; SFCT = subfoveal choroidal thickness.

Model Development

Five ML models to predict MMD progression were applied and compared (Table 2). The best-tuned hyperparameters for each algorithm are listed in Supplementary Table S5. The XGBoost algorithm showed the best discrimination ability with AUROC of 0.87 ± 0.02 and AUPRC of 0.91 ± 0.01 in the training dataset. This model demonstrated superiority in an accuracy of 0.85 ± 0.02, sensitivity of 0.86 ± 0.05, specificity of 0.83 ± 0.03, and F₁ score of 0.85 ± 0.03. The calibration plots of the XGBoost model indicated strong agreement between the predicted probabilities and the observed actual proportions. DCA plots revealed that when the threshold risk of MMD progression was less than 97.0% in the training set, the net benefit of the model surpassed that of both the “intervene in all” and “intervene in none” strategies (Figs. 3A, 3B, 3C). 

Table 2.

 

Comparative Performance of Different ML Algorithms

 

Figure 3.

 

Performance of XGBoost model in identifying MMD progression risk in the training, internal validation and external validation datasets. The model demonstrated the discrimination abilities with AUROCs of 0.87 ± 0.02, 0.84 ± 0.006, and 0.80 ± 0.008 in the training, internal validation, and external validation datasets, respectively (A, D, G) The model was well calibrated (B, E, H) and showed promising clinical utility (C, F, I) in the training, internal validation, and external validation datasets, respectively. AUROC = area under the receiver operating characteristics curve; MMD = myopic macular degeneration.

Model Validation

In the internal validation dataset, the XGBoost model achieved good discriminative value for predicting MMD progression (AUROC = 0.84 ± 0.006 and AUPRC = 0.87 ± 0.007; see Table 2). Meanwhile, the model was well-calibrated as graphed in the calibration plots. It obtained significant clinical value when the threshold risk was lower than 67.0% (Figs. 3D, 3E, 3F). 

The performance of XGBoost model in the external validation was presented in Table 2 and Figures 3G, 3H, 3I. The model demonstrated clinically satisfactory discriminative capacity with an AUROC of 0.80 ± 0.008 and calibration ability for predicting MMD progression. This model achieved an acceptable accuracy of 0.80 ± 0.01, sensitivity of 0.97 ± 0.01, specificity of 0.63 ± 0.03, F₁ score of 0.83 ± 0.01, and AUPRC of 0.83 ± 0.01. The model showed superior net benefit when the threshold risk was less than 46.0%. 

Subgroup Analysis, Sensitivity Analysis, and Webpage Tool

The XGBoost model deployment in the 3 age groups achieved remarkable discriminative abilities, with AUROCs of 0.92 ± 0.005 in children and adolescents, 0.91 ± 0.008 in young adults, and 0.93 ± 0.02 in middle-aged adults (Fig. 4). Sensitivity analysis was performed in participants with bilateral cycloplegic sphere ≤−6.00 D from the COIP dataset (n = 80) and demonstrated the AUROC of 0.79 ± 0.02 with the AUPRC of 0.81 ± 0.01 (Supplementary Table S6). The XGBoost algorithm has been implemented as a web application, which is publicly available at https://hmzoc.shinyapps.io/mmd_risk_prediction/ (Fig. 5). 

Figure 4.

 

Receiver operating characteristics curves of the XGBoost model for predicting MMD progression for different age groups in the development cohort. The AUROCs of the ML model were 0.92 ± 0.005 (A), 0.91 ± 0.008 (B), and 0.93 ± 0.02 (C) for MMD progression in 7 to 18 years (children and adolescents), 19 to 40 years (young adults), and 41 to 70 years (middle-aged adults) groups. AUROC = area under the receiver operating characteristics curve; MMD = myopic macular degeneration.

Figure 5.

 

Online tool application of the XGBoost model. An example of a 25-year-old male highly myopic participant from the development cohort is demonstrated in this online webpage. The participant's baseline information was shown in the left panel, and the tool outputs that her risk of MMD progression over 10 years is 68.3%. In fact, the example participant developed MMD progression from tessellated fundus (C1) to diffuse atrophy (C2) in the second follow-up year. MMD = myopic macular degeneration.

Discussion

Main Findings

This study for the first time leverages extensive long-term cohort and develops a data-driven ML algorithm that forecasts the decadal risk of MMD progression in high myopia populations. The SFCT variable measured by SS-OCT ranks the most critical factor among the 34 baseline predictors. Furthermore, the prediction model achieves high accuracy and clinical value in predicting future MMD traits for a decade in advance. 

Long-Term MMD Natural Course

The long-term MMD trajectory in our findings agreed with previous longitudinal findings. Up to now, only 4 studies documented long-term MMD progression ≥10 years. Hayashi et al. reported that 327 of the 806 highly myopic eyes (40.6%) demonstrated progression of MMD without available META-PM classification from a 12.7-year medical records analysis.²⁹ An 18-year retrospective case series study revealed that the MMD progression rate was 47.0 per 1000 eye-years among highly myopic eyes with a mean age of 42.3 years.⁵ In the Beijing Eye Study, 35.5% eyes showed 10-year MMD progression among 110 eyes of middle-aged patients with high myopia, where the progression rates were 19% in tessellated fundus (C1), 71% in diffuse atrophy (C2), and 100% in patchy atrophy (C3).⁶ Another recent Singapore longitudinal study depicted the 12-year cumulative MMD incidence was 10.3% in adults with myopia and MMD progression was 19.6% among highly myopic individuals aged over 40 years. In the development cohort, we found that 133 eyes (20.2%) exhibited MMD progression over a median of 10.9 years among high myopes aged 7 to 70 years, which included 6.0% of healthy eyes (C0), 32.0% of tessellated fundus (C1), 50.0% of diffuse atrophy (C2), 100.0% of patchy atrophy (C3), and 100.0% of macular atrophy (C4) at baseline. The presented long-term progression data enhance existing knowledge regarding the natural course of MMD pathology. Utilization of the standardized META-PM classification, combined with a substantial sample size, offers a novel opportunity to develop ML algorithms predicting MMD progression over a decade in advance. 

Predictors for MMD Progression

Identifying the key predictors for MMD progression is vitally important for constructing an interpretable and accurate model. From a comprehensive set of multimodal clinical and imaging variables, SFCT measured by SS-OCT was ranked as the most significant predictor for MMD progression. This prominence likely stems from the choroid's fundamental role in MMD pathogenesis, of which the choroid ischemia from elongated AL is hypothesized to be the primary factor contributing to MMD development.³⁰ A recent OCT-based MMD classification highlighted the application of choroidal thickness thresholds.¹¹ We have summarized the risk factors for MMD incidence and progression in previous studies (Supplementary Table S7). SFCT was recognized as an independent prognostic factor for MMD progression in a 2-year study utilizing SS-OCT measurements.¹² Longer AL, worse BCVA, older age, and female gender were conventional risk factors for MMD pathologies. Additionally, shallower ACD was found to be predictive for progression of MMD, ranking sixth among the factors studied, although the underlying mechanisms are required for more investigation. 

First ML Algorithm for Predicting 10-Year MMD Progression

Determining the risk of sight-threatening complications related to high myopia in younger age holds significant clinical relevance but poses a considerable challenge. A few studies have identified individuals at a greater risk of developing MMD; however, none to our knowledge has endeavored to predict the long-term personalized MMD progression risk in patients with high myopia with all age groups.^7,12 The risk classification in previous studies was often inferred from short-term longitudinal data or focused on MMD incidence in adult myopic population aged over 40 years. Our model showed comparable discrimination abilities with previous models (Supplementary Table S8) and presented the calibration and clinical utility. The excellent discriminative capacity across different ages suggested that the model could be applied in children and teenagers, young adults, middle-aged persons, and elderly persons. Evidence from our cohorts and previous studies indicated that MMD development can occur as early as childhood in patients with high myopia, which emphasized early risk identification in this stage.^31,32 Such accurate long-term predictions were of significance given that targeted screening and timely intervention for preventing visual impairment from MMD in the mild stage was critically important to maximum healthcare benefits. 

External Validation

External validation is an important aspect of developing clinical prediction models with pragmaticism. Although no compromised in model discrimination or calibration observed in the internal validation set suggests a low likelihood of overfitting in our ML model, a more crucial aspect before the widespread implementation of a prediction model is external validation in a distinct population. In our study, we utilized a 4-year high myopia dataset with AL ≥25.5 mm for external validation. The acceptably accurate performance in the external validation (AUROC = 0.80 ± 0.008) confirmed the model effectiveness. Notably, the median SER of external validation cohort was −5.38 D, which was less than −8.75 D observed in the development cohort. In the sensitivity analysis within participants with bilateral cycloplegic sphere ≤−6.00 D from the external dataset, the model also demonstrated acceptable discrimination ability with mean AUROC of 0.79, suggesting the generalizability for visual outcome surveillance across patients with high myopia. Additionally, it is important to note that the average 95% prediction interval width across participants from the external validation cohort was 0.60, indicating the uncertainty in the predictions for an individual participant. This uncertainty may be attributed to the relatively moderate sample size and diverse characteristics of the external validation dataset. Nevertheless, the model is beneficial for identifying overall risk trends when applied to a group of patients with high myopia, providing valuable insights for clinical and epidemiological management of MMD progression. Future studies with larger external validation datasets are warranted to further validate our model. 

Implications of Results and Future Research

The clinical utility of our model implicated it could inform stratified follow-up regimens and have the potential cost effectiveness. The deployment of this algorithm is anticipated to mitigate the substantial physical and economic barriers attributed to MMD associated irreversible visual loss, given that MMD was responsible for visual impairments in approximately 10 million individuals and blindness in 3.3 million people, culminating in a worldwide productivity loss amounting to US $6 billion as of 2015.^33,34 Recent myopia control treatments have undergone significant advancements in children focusing on slowing axial elongation, refraction progression, and choroid thickening.^35–40 Considering SFCT and AL as the top two predictors, these therapeutic interventions may show potential to impede MMD progression and prevent visual impairment in adolescent and adults with high myopia. Although the Atropine Treatment Long-term Assessment Study (ATLAS) reported that 2 to 4 years’ atropine therapy did not reduce long-term MMD incidence with a relatively small sample size,⁴¹ future more extended large prospective clinical trials are warranted in this field. 

Strengths and Limitations

This study had several strengths. First, the prediction model was developed in a large, prospective, population-representative cohort with homogeneous high myopes and 10-year long-term standardized follow-up which enabled capture of more clinical events. Second, predictor selection was performed in a data-driven approach from an extensive set of multimodal variables with six key features ultimately chosen for the final model. Third, the proposed XGBoost model was validated in an independent external cohort. The model has been translated into a webapp for clinical application. 

Several limitations need to be considered when interpreting our findings. First, the model was externally validated in a separate Chinese cohort with only 4-year follow-up visits. Further study is required to evaluate its generalizability to populations with longer follow-up periods and to other ethnic groups. Second, a significant proportion of participants (22.9%) failed to complete the decade follow-up, in part due to the coronavirus disease 2019 (COVID-19) pandemic in the scheduled visits within the development cohort. However, the favorable performance in the external validation cohort suggested that the bias was unlikely to affect our results. Third, genetic data were not included in our analysis. Given the current uncertainty regarding the role of genetic information in MMD risk,^42–44 future genetic studies involving large consortia are expected to reveal the genetic traits associated with MMD, allowing for updating our ML program. Fourth, the choroidal imaging was assessed using SS-OCT and its generalizability to other methods, such as the enhanced depth imaging (EDI) mode on spectral domain OCT with manual or automated off-line measurements cannot be confirmed. Fifth, the assessment of posterior staphyloma was relied on a 45 degrees fundus camera, which may result in a partial missed diagnosis due to the camera's limited imaging range. Additional imaging techniques, such as magnetic resonance imaging or ultrawide field OCT, are essential for a more precise evaluation of staphyloma. 

Conclusions

In summary, ML prediction model was developed and validated with satisfactory discriminative accuracy and promising clinical value to predict 10-year MMD progression in the highly myopic population. Thinner SFCT, longer AL, worse BCVA, older age, female gender, and shallower ACD were the most important predictors. This prediction model will help personalized risk stratification on MMD progression and surveillance of visual outcome in individuals with high myopia. 

Acknowledgments

The authors are grateful to all participants from Zhongshan High Myopia Cohort and Chinese Ocular Imaging Project. 

Funded by the Hainan Province Clinical Medical Center, the National Natural Science Foundation of China (82371086), and Global STEM Professorship Scheme (P0046113). The funder had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. 

Author Contributions: Wang had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. 

Study Concept and Design: Chen, Yang, He, Wang. Acquisition, analysis, or interpretation of data: Chen, Yang, He, Wang. Drafting of the manuscript: Chen, Yang, He, Wang. Critical revision of the manuscript for important intellectual content: All authors. Statistical analysis: Chen, Wang. Obtained funding: Wang, He. Administrative, technical, or material support: Chen, Yang, Liu, Xiong, Wang, Li, Zheng, He, Wang. Study supervision: Wang, He. 

Disclosure: Y. Chen, None; S. Yang, None; R. Liu, None; R. Xiong, None; Y. Wang, None; C. Li, None; Y. Zheng, None; M. He, None; W. Wang, None