Development and Validation of a Model to Predict Who Will Develop Myopia in the Following Year as a Criterion to Define Premyopia

Abstract

Purpose: 

To develop and validate models to predict who will develop myopia in the following year based on cycloplegic refraction or ocular biometry and to identify thresholds of premyopia.

Methods: 

Prospective longitudinal data of nonmyopic children at baseline from the Guangzhou Twins Eye Study and the Guangzhou Outdoor Activity Longitudinal Study were used as the training set, and the Singapore Cohort Study of the Risk factors for Myopia study formed the external validation set. Age, sex, cycloplegic refraction, ocular biometry, uncorrected visual acuity, and parental myopia were integrated into 3 logistic regression models to predict the onset of myopia in the following year. Premyopia cutoffs and an integer risk score system were derived based on the identified risk.

Results: 

In total, 2896 subjects with at least 2 visits were included. Cycloplegic refraction at baseline is a better predictor to identify the children with myopia onset [C-statistic=0.91, 95% confidence interval (CI), 0.87–0.94; C-statistic=0.92, 95% CI, 0.92–0.92 for internal and external validation, respectively], comparing to axial length, corneal curvature radius (CR) and anterior chamber depth (C-statistic=0.81, 95% CI, 0.73–0.88; C-statistic=0.80, 95% CI, 0.79–0.80, respectively), and axial length/CR (C-statistic=0.78, 95% CI, 0.71–0.85; C-statistic=0.76, 95% CI, 0.75–0.76). With a risk of >70%, the definitions of premyopia indicating approaching myopia onset were 0.00 D for 6–8 years and −0.25 D for ≥9 years in children with 2 myopic parents.

Conclusions: 

Either cycloplegic refraction or ocular biometry can predict 1-year risk of myopia. Premyopia can be successfully defined through risk assessments based on children’s age and predict who would require more aggressive myopia prophylaxis.

Myopia has become a major source of vision impairment worldwide, with its prevalence and severity growing rapidly in recent decades.^1–3 Myopia has been predicted to affect ~50% of the global population by 2050, imposing a total global health cost of $1.7 trillion.^4,5 Alongside this tremendous cost is the personal burden of myopia, which from its irreversible complications carries high rates of morbidity and impacts quality of life.

Prevention of the onset of myopia is fundamental to reducing the prevalence of myopia among the societies. Additional time spent outdoors in school children have dose-dependent prophylactic effects on preventing the development of myopia, with randomized clinical trials showing reductions in cases of myopia onset of 20% with an additional outdoor time of around 40 minutes.⁶ Greater benefits were observed with even more time spent outdoors.⁷ Although effective, the feasibility of spending this amount of time outdoors is low in competitive schooling environments, where academic focus is regarded as incompatible with so much time spent outdoors. Reforms in schooling to reduce educational pressure in early school years seemed to be pursued in some areas, such as China and Japan.⁸ However, students still suffer the risk of lack of time outdoors because of the coronavirus disease 2019 home quarantine.⁹ Therefore, prophylactic measure would be more impactful if it can prioritize those who are in greatest need.

Recently, premyopia was defined as ≤+0.75 diopters (D) and >−0.50 D based in part on risk prediction models.¹⁰ In the Collaborative Longitudinal Evaluation of Ethnicity and Refractive Error (CLEERE) study, predicted cutoffs for premyopia were ≤+0.75 D for 6 years of age, ≤+0.50 D for 7–8 years, ≤+0.25 D for 9–10 years, and ≤0 D for 11 years,¹¹ and similar cutoffs were also suggested in Asian populations.^12,13 Considering performing cycloplegic refraction massively among the children with normal vision or without myopia is challenging, noninvasive parameters including ocular biometry and additional known risk factors should be considered as alternative predictors in myopia prediction. Zhang et al¹⁴ developed a prediction model utilizing ocular biometry, in combination with age, sex, and height but not with baseline refraction, as predictors and further proved this model could achieve high accuracy to predict who will develop >−0.75 D myopia in 3 years among children in China and Singapore; however, the model was complex and likely requires computer assistance for evaluation, and it might be location specific with rates of change varied with the environment.

The objectives of this study were to develop and validate prediction models assessing the risk of developing myopia in next year through cycloplegic refraction and ocular biometry, such models in combination of risk triaging system could be used as criteria to define “premyopia” —an indication for identifying the children who require more aggressive prophylaxis.

METHODS

Study Populations

Three cohorts were included in this study: the Guangzhou Twins Eye Study (GTES), the Guangzhou Outdoor Activity Longitudinal Study (GOALS), and the Singapore Cohort Study of the Risk factors for Myopia study (SCORM). The methods of data collection in the 3 cohorts are reported elsewhere.^6,15–17 In the GTES cohort, twins living in Guangzhou, China, were invited to have annual follow-up examinations from 2006 to 2016, and data of the first-born twin were included in the study. In the GOALS cohort, children from 6 schools were enrolled in a randomized clinical trial, and the control group with 3-year follow-up visits (from 2009 to 2012) was selected. In SCORM, children were recruited from 3 schools and were followed-up for a period of 9 years (from 1999 to 2007). The inclusion age was 7–15 years, 6 years, and 7–9 years for GTES, GOALS, and SCORM, respectively.

All examinations were conducted in accordance with the Tenets of the World Medical Association’s declaration of Helsinki. Written informed consent was obtained from all participants, their parents or statutory guardians. Ethical approval was obtained from the ethics committees at the Zhongshan University Ethical Review Board, Zhongshan Ophthalmic Center, and Singapore Eye Research Institute.

Measures

Autorefraction was measured using autorefractors (GTES and GOALS: Topcon KR8800, SCORM: Canon RK5) for each participant at baseline and each follow-up visit. Cycloplegia was achieved before refraction measurement in all the 3 studies with 3 drops of 1% cyclopentolate. Ocular biometry, including axial length (AL), corneal curvature radius (CR), and anterior chamber depth (ACD), were measured using the IOL Master (GTES and GOALS: Zeiss IOL Master 500) or an ultrasound biometry machine (SCORM: Nidek Echoscan model US-800). Visual acuity was assessed using the logarithm of the minimum angle of resolution charts following standard procedures.^18,19 Refractive error in parents of participants were recorded from refraction measurements (GTES) or questionnaires (GOALS and SCORM).

Definition and Statistical Methods

Participants’ right eyes were arbitrarily selected for analysis. Spherical equivalent (SE) was defined as the sum of sphere and 1/2 cylinder power. Myopia was defined as SE ≤ −0.50 D. All participants without myopia at baseline with ≥2 visits at ≥1-year interval were included in the analysis. Myopia onset in the following year was defined as that a nonmyopic (SE > −0.5 D) child at baseline developed myopia in the next 12+2 months.

Three logistic regression models were developed with myopia onset in the next year as a main outcome (with each visit as a predictive time point). Model 1 included baseline age, sex, SE, and myopic parent (categorized as 0, 1, or 2). Model 2 included baseline age, sex, AL, CR, ACD, uncorrected visual acuity (UCVA), and myopic parent. Model 3 included baseline age, sex, AL/CR, ACD, UCVA, and myopic parent.

Data from GETS and GOALS were assigned as model building sets, and a 5-fold cross-validation was conducted. External validation was performed using data from SCORM, and C-statistics with 95% confidence interval (CI) were calculated. To simplify the prediction model to practical use, we developed a integer risk score system following the methods of the Framingham Heart Study.²⁰ In brief, integer points to each level of each risk factor were computed using β coefficients of the logistics model and category intervals. Predicted risk of myopia using risk score systems was reported and compared in validation cohort. P-values of <0.05 were defined as significant. All the analysis was conducted using R software (version 4.1.2, https://www.r-project.org/).

RESULTS

A total of 2678 children aged 6–16 years were included in the current study (252 nonmyopic subjects from GTES, 652 from GOALS, and 1774 from SCORM). Mean age at baseline was 7.58+1.96 years in the model training set (10.18+2.03 y in GTES, 6.57+0.32 y in GOALS), and 7.85+0.85 years in external validation set (P<0.001). Female sex made up 46.8% in GTES, 45.7% in GOALS, and 49.3% in SCORM (P=0.115). Other baseline characteristics are summarized in Table 1. Subjects in the model training set were less myopic, had shorter ALs at baseline, and had a lower incidence of high myopia at last visit compared with those in external validation set. The change of SE during the next 1 year was −0.42 D (95% CI, −1.66 to 0.82 D) for age of 6 years, and −0.25 D (95% CI, −1.10 to 0.59 D) for age of 15 years (Supplementary Digital Content Table 1, https://links.lww.com/APJO/A218).

TABLE 1 - Baseline Characteristics of the Training and External Validation Cohorts

	Model training set			External validation
Variable	GTES (n=252)	GOALS (n=652)	Total (n=904)	SCORM (n=1774)	P
Baseline age, mean (SD), y	10.18 (2.03)	6.57 (0.32)	7.58 (1.96)	7.85 (0.85)	<0.001
Female, n (%)	118 (46.8)	298 (45.7)	416 (46.02)	875 (49.3)	0.115
Baseline SE, mean (SD), D	0.84 (0.80)	1.13 (0.84)	1.04 (0.84)	−0.38 (1.68)	<0.001
Baseline AL, mean (SD), mm	22.94 (0.75)	22.66 (0.70)	22.74 (0.73)	23.32 (0.95)	<0.001
Baseline CR, mean (SD), mm	7.75 (0.26)	7.78 (0.25)	7.77 (0.26)	7.75 (0.25)	0.565
Follow-up period, mean (SD), y	5.63 (2.56)	3.65 (0.67)	4.49 (1.99)	4.63 (2.27)	0.105
Baseline UCVA (logMAR), mean (SD)	0.09 (0.14)	0.07 (0.08)	0.08 (0.11)	0.07 (0.15)	0.058
Having 2 myopic parents, n (%)	50 (19.84)	152 (23.31)	202 (22.35)	392 (22.10)	0.923
High myopia at last visit, n (%)	16 (6.35)	3 (4.60)	19 (2.10)	107 (6.03)	<0.001

ACD indicates anterior chamber depth; AL, axial length; CR, corneal curvature radius; GOALS, Guangzhou Outdoor Activity Longitudinal Study; GTES, Guangzhou Twin Eye Study; logMAR, logarithm of the minimum angle of resolution; SE, spherical equivalent; SCORM, Singapore Cohort Study of the Risk factors for Myopia study; UCVA, uncorrected visual acuity.

The parameters for the 3 models are summarized in Supplementary Digital Content Table 2, https://links.lww.com/APJO/A219. A younger baseline age and a greater number of myopic parents were significantly associated with increased risk of myopia onset in the following year in all the 3 models (P<0.001 for all). Significant predictors for myopia onset in model 1 included less baseline hyperopia (β=−3.84, P<0.001). In model 2, female sex (β=0.79, P<0.001), a longer AL (β=3.40, P<0.001), a steeper cornea (β=−7.40, P<0.001), a shallower ACD (β=−1.33, P<0.001), and worse UCVA (β=2.63, P<0.001) and in model 3, female sex (β=−0.13, P<0.001), a higher ratio of AL/CR (β=17.86, P<0.001), and worse UCVA (β=2.21, P<0.001) predicted myopia.

After a 5-fold cross-validation of the 3 linear regression models, a C-statistic of 0.91 (95% CI, 0.87–0.94) in model 1, 0.81 (95% CI, 0.73–0.88) in model 2, and 0.78 (95% CI, 0.71–0.85) in model 3 was observed (Table 2). After external validation, model 1 had the highest predictive accuracy (C-statistic=0.92, 95% CI, 0.92–0.92), followed by model 2 (C-statistic=0.80, 95% CI, 0.79–0.80) and model 3 (C-statistic=0.76, 95% CI, 0.75–0.76).

TABLE 2 - Predictive Performance of Models for Risk of Myopia in Internal and External Validation Cohorts

	Internal validation		External validation
Model	C-statistic	95% CI	C-statistic	95% CI
1. Age, sex, SE, myopic parent	0.91	0.87–0.94	0.92	0.92–0.92
2. Age, sex, AL, CR, ACD, UCVA, myopic parent	0.81	0.73–0.88	0.80	0.79–0.80
3. Age, sex, AL/CR, ACD, UCVA, myopic parent	0.78	0.71–0.85	0.76	0.75–0.76

ACD indicates anterior chamber depth; AL, axial length; CI, confidence interval; CR, corneal curvature radius; SE, spherical equivalent; UCVA, uncorrected visual acuity.

The ability to predict myopia onset next year of different premyopia cutoffs were tested. As seen in Fig. 1 and Supplementary Digital Content Figures 1 and 2, https://links.lww.com/APJO/A220, overall the risk of developing myopia in the following year decreased with age when using a single cutoff. The −0.25 D cutoff presented the highest risk of myopia onset ranging from 66% to 77% in children aged 6–14 years with 2 myopic parents (Fig. 1), whereas the cutoff of +0.75 D only generated <25% of risk. When a risk of ≥70% was used, the definitions of premyopia indicating approaching myopia onset were 0.00 D for age of 6–8 years and −0.25 D for age of 9 years or above among children with 2 myopic parents, 0.00 D for 6 years and −0.25 D for 7 years or above among children with 1 myopic parent (Supplementary Digital Content Fig. 1, https://links.lww.com/APJO/A220), and −0.25 D for children with no myopic parent (Supplementary Digital Content Fig. 2, https://links.lww.com/APJO/A220).

FIGURE 1: Risk prediction of myopia onset in the next year using different premyopia cutoffs among children with 2 myopic parents.

An integer-based risk score was developed from the 3 models to facilitate premyopia risk assessments in clinical practice. As shown in Figures 2 and 3, risk factors were categorized with a reference point value for each category. Individual total risk scores were computed and risk of developing myopia next year was estimated. Risk scores of <−6 points using cycloplegic SE, 8 points using individual ocular biometry, and 0 points using AL/CR ratio were associated with a 70.3% (Fig. 2), 73.6% (Fig. 3A), and 60.3% (Fig. 3B) higher risk of developing myopia, respectively. The risk score system showed similar discrimination, as well as curves of prediction risk over scores in internal and external validation cohorts across the models.

FIGURE 2: Risk score system for predicting the presence of myopia in the following year based on cycloplegic refraction.

FIGURE 3: Risk score system for predicting the presence of myopia in the following year based on ocular biometry, including axial length, corneal curvature radius, anterior chamber depth (A), axial length/corneal curvature radius (AL/CR) (B).

DISCUSSION

The present study predicted myopia onset in the following year in school-aged children using cycloplegic refraction and ocular biometry. Cycloplegic refraction showed superior predictive ability after internal and external validation compared with ocular biometry. The number of myopic parents was a strong predictive factor, which allowed for age-adjusted SE cutoffs to be determined, where younger ages combined with one or more parents with myopia substantially increased the likelihood of myopia development. In addition, this study developed a risk score for myopia, which evaluates risk based on commonly measured eye parameters and patient demographics. Prediction models and risk indexes proposed by this study may be valuable clinical tools for the identification of premyopia and facilitate early intervention strategies which will ultimately reduce the burden and morbidities of myopia in the long term.

The predictive accuracy in the current model using cycloplegic SE was 0.91 in internal validation and 0.92 in the external validation cohorts, which is comparable with previously reported models.^11,12,21 This supports that baseline SE is a strong predictor of incident myopia if adequate cycloplegia can be achieved, as suggested in other studies.^21–23 Accurate, cycloplegic refraction is only available in eye specialist clinics, and dilatory agents such as atropine and cyclopentolate are not accessible to general physicians. This limits accessibility to predictions based on cycloplegic refraction, and general physicians should provide early referral to eye vision specialists for premyopia assessments if parental history is evident. Alternatively, more easily attainable parameters could be integrated into prediction models for risk assessment by non–eye vision specialists.

With the aim of improving accessibility of myopia risk assessments, this study developed the first risk score system that used ocular biometry alongside patient characteristics such as age and number of myopic parents, to provide a scalable, population-based clinical tool for premyopia. Without the need for cycloplegic refraction data inputs, the calculator can provide a score range that gives an estimated risk of myopia that will hopefully encourage early referral to eye vision specialists.^24,25 It is noteworthy that when cycloplegic refraction is not available, UCVA and noncycloplegic refraction are commonly used as referral criteria for school screening or general physicians. However, based on the current and previous results,^24,25 their effects are inadequate in risk prediction for myopia onset, and may need to be combined with AL or AL/CR. Incorporation of more ocular biometric parameters seems to increase the accuracy of prediction model. Zhang et al’s¹⁴ model including ocular biometrics of AL, ACD, lens thickness (LT), vitreous chamber depth (VCD), corneal curvature and their multiplication (ACD*LT*VCD) achieved high accuracy (0.974), but the area under the curve (AUC) decreased to 0.500 when LT or VCD was removed. Different from the 3-year prediction and a myopia definition of ≤−0.75 D in Zhang et al’s model, we conducted prediction over a shorter period and used a lower value of myopia definition, and the accuracy has generated an sufficient accuracy (AUC 0.78–0.81 in internal validation and 0.76–0.80 in external validation, respectively) for the purpose of population-based screening, which even surpasses AUC’s of other risk calculators such as the Framingham Risk Score²⁶ and Chads-Vasc scores²⁷ for atrial fibrillation. Further validation and refining of this risk score system in larger-scale multiethnic populations should allow for the translation of these scores to clinical use in future.

The results of our analysis led to the proposal of premyopia cutoffs that are different to those described by previous studies. CLEERE proposed cutoffs of <+0.75 D, +0.5 D, +0.25 D, and 0 D for ages 6, 7–8, 9–10, and 11+ years, respectively; however, our proposed cutoffs were 0 D for 6–8 years and −0.25 D for 9+ years. Although the cutoff of +0.75 D is associated with high predictive ability in some populations,^21,28 it was only associated with a <25% risk to develop myopia next year based on our analysis. This may mainly because of that in the CLEER study, there was a much longer time frame for onset of myopia that the predicted period was up to 6 years, whereas only a 12-month risk was assessed in the current study. A less hyperopic premyopia definition may be more valuable to identify children who require aggressive treatments for myopia prevention. It should be noted that the definition of myopia in the current study is SE ≤−0.5 D, based on the criterion suggested by the consensus of the International Myopia Institute.¹⁰ However, the capacity of emmetropization may be impaired with all negative refraction, that is, −0.25 D. Whether treatments for preventing myopia should be implemented earlier based on a less negative definition needs to be explored in future studies. In addition, the relationship between SE and AL/CR was nonlinear and weaker around the emmetropic values,²⁹ and the myopic progression varied across ethnicities.³⁰ Therefore, the use of a definition based on ocular biometry or age-specific criteria for myopia should be considered in further investigations.

The strengths of this study included the external validation of our prediction model using an independent cohort (SCORM) from Singapore, which has a higher myopia incidence compared with China. The high accuracy in the external validation indicates these models should translate well for use in other populations with a high prevalence of myopia. In addition, this study established a risk score system which is easy to operate as a screening tool, and may have applications for clinical practice and clinical trials.

Some limitations must also be acknowledged. First, the study’s sample size is relatively small, especially the sample within the training set. Future studies including more nonmyopic children at baseline are needed to validate these prediction models. Second, the data set comprised mostly Chinese ethnicity, and multiethnic studies may be needed to validate these tools across different racial profiles. Third, the generalizability of the model may be limited in populations with low prevalence of myopia, as the model was trained by a data set comprising high rates of myopia. Lastly, this study did not conduct a Youden index test, which would indicate the sensitivity and specificity of our proposed cutoffs.

In conclusion, this study proved cycloplegic refraction and ocular biometry predicts the 1-year risk of developing myopia with validated efficacy. On the basis of the prediction model, premyopia definitions varied depending on age and the number of myopic parents. The premyopia cutoffs and risk score systems offered by this study provide practical solutions for the screening of at-risk children who will require premyopic interventions to preserve vision and prevent myopia.