Risk assessment and prediction of myopia among school-age children in South China: protocol for a longitudinal follow-up study

Abstract

Introduction

Myopia among children and adolescents has become an important public health concern. The pathogenesis of myopia remains unclear, and myopia may be associated with various ocular complications. Although myopia is preventable and controllable, established myopia is currently irreversible. Therefore, early and comprehensive strategies for myopia prevention and control are needed. This study aims to longitudinally assess fundus images, ocular biometric parameters, and environmental and behavioural factors in children and adolescents and to develop an artificial intelligence (AI)-based model for dynamically assessing the risk of myopia.

Methods and analysis

This is a school-based prospective longitudinal study that will be conducted in four administrative districts in Shenzhen, China, including two economically developed and two economically underdeveloped districts. One primary school will be randomly selected from each district, and all first-grade students in the selected schools will be recruited as study participants. The study will be conducted over the project period from April 2024 to November 2026, with follow-up assessments scheduled every 6 months after enrolment. During the follow-up period, participants will undergo fundus photography, ocular biometric measurements, myopia screening, including visual acuity and non-cycloplegic refraction, and questionnaire assessments of environmental and behavioural factors. AI technology will be used to interpret and quantify findings from bilateral fundus images. A DeepSurv model for predicting myopia risk will be developed using multimodal AI combined with ocular biometric parameters, environmental and behavioural factors, and other relevant indicators. The model will estimate the 1-, 2- and 2.5-year risks of developing myopia. The primary outcome of the study will be the incidence of myopia.

Ethics and dissemination

This study was approved by the Ethics Committee of the Shenzhen Eye Hospital (2024KYPJ023) in accordance with the Declaration of Helsinki. Written informed consent will be obtained from the parents or legal guardians of all participants. The study findings will be disseminated through peer-reviewed publications and presentations at academic conferences. Key findings relevant to myopia prevention and control will also be communicated to participating schools, parents and other relevant stakeholders through appropriate educational and public health communication channels, with the aim of supporting evidence-based strategies for myopia prevention and control.

Trial registration number

ChiCTR2400082680.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This prospective longitudinal study includes assessments every 6 months after enrolment to evaluate changes in ocular parameters, fundus characteristics, and myopia-related environmental and behavioural factors.

• This study integrates multimodal data, including fundus images, ocular biometric parameters, and environmental and behavioural factors, to develop an AI-based model for myopia risk prediction.

• Participants are recruited from four administrative districts of Shenzhen, including economically developed and underdeveloped districts, to capture variation in socioeconomic and environmental contexts.

• Non-cycloplegic refraction may introduce measurement error and potentially overestimate myopia prevalence, particularly in younger children.

• Questionnaire-based assessment of environmental and behavioural factors is prone to recall and self-report bias.

Introduction

Myopia has become increasingly common in children and adolescents owing to a lack of outdoor activities, increased educational pressure and other factors.^{1 2} Estimates indicate that 49.8% of the population will suffer from myopia by 2050; approximately 10% will experience high myopia.³ In 2020, myopia affected approximately 53.4% of the population in the Asia-Pacific region and 51.6% in East Asia.³ A total of 52.7% of children and adolescents in China suffered from myopia in 2020,⁴ ranking first worldwide.⁵ Myopia is now one of the three major public health problems among school students in China.⁶

Myopia has been reported to be associated with reduced attention and an increased risk of visual fatigue, headache and blurred vision, which may have adverse effects on learning among children and adolescents.⁷ Progressive myopia may increase the likelihood of developing high myopia, which is associated with an increased risk of ocular complications, including cataracts, glaucoma, retinal abnormalities and vision impairment.⁸ As the prevalence of myopia continues to increase globally, the escalating harm it poses to society is also increasing. In addition to potentially causing permanent visual impairment, myopia also imposes substantial economic burdens on society and promotes other associated effects.⁹ The annual global economic cost of uncorrected refractive errors was estimated at US$202 billion.¹⁰

Preventing and controlling myopia has many challenges. Epidemiological studies suggest that several factors affect the occurrence and development of myopia. The theory that a combination of environmental, behavioural and genetic factors influences myopia in children and adolescents is widely acknowledged.¹¹ However, the mechanism of myopia is not clear; therefore, predicting and evaluating the risk of myopia is difficult.¹² Traditional diagnosis and intervention methods do not meet the requirements of myopia prevention and control.

Artificial intelligence (AI) technology has developed rapidly and has been increasingly applied in ophthalmology, providing new opportunities for myopia prevention and control.^{13 14} Shu et al¹⁵ used eye images captured by mobile devices to screen for paediatric ophthalmic conditions, including myopia and strabismus. Tan et al¹⁶ developed an AI-based deep learning model using fundus photographs to predict refractive errors, high myopia and myopic macular degeneration. Similarly, Varadarajan et al¹⁷ demonstrated that deep learning algorithms could estimate refractive error from fundus photographs with high accuracy.

Despite these advances, the application of AI to myopia risk assessment remains at an exploratory stage. Existing studies have primarily focused on individual information sources, such as fundus images or ocular biometric parameters, and have often assessed associations between individual risk factors and myopia separately. In addition, many studies have used cross-sectional or short-term follow-up data, limiting the assessment of changes in ocular parameters and environmental and behavioural factors over time and their relationships with myopia development and progression. Studies integrating retinal images, ocular biometric parameters, and environmental and behavioural factors within a longitudinal framework remain limited. Therefore, we aim to apply AI technology in a longitudinal epidemiological study of myopia among children and adolescents to investigate the contributions of retinal images, ocular biometric parameters, and environmental and behavioural factors to myopia risk. A 2½-year longitudinal follow-up study will be conducted among students, with AI-based models used to assess dynamic changes in myopia risk over time.

Methods

Objectives

Primary objective

Through fundus screening, file establishment and regular follow-ups, a myopia prevention and control data platform will be built to explore the longitudinal relationships between fundus imaging parameters, ocular biological parameters, environmental and behavioural factors, and myopia among children and adolescents.

Secondary objectives

Secondary objectives include using AI technology to quantify relevant fundus imaging parameters, examining the relationship between the quantified parameters and myopia, analysing the relationship between environmental and behavioural factors and myopia, and investigating the relationship between ocular biological parameters and myopia.

Study design

This longitudinal cohort study will be conducted from April 2024 to October 2026. Students aged approximately 6–9 years will be recruited from selected primary schools in Shenzhen, China, and will undergo baseline assessment followed by follow-up assessments every 6 months after enrolment. Data analysis will be performed and completed by November 2026. This study was approved by the Ethics Committee of the Shenzhen Eye Hospital based on the Declaration of Helsinki. A flowchart of the study is shown in figure 1.

Figure 1. Study flowchart.

Participants

Study area

This study will be conducted in Shenzhen. Shenzhen is a subprovincial city of Guangdong Province in China, and a special economic zone, economic centre and internationalised city. Shenzhen covers an area of 1997.47 km², with a population of 17.6618 million in 2022. Shenzhen has 10 districts: Nanshan, Futian, Luohu, Yantian, Baoan, Guangming, Longhua, Pingshan, Longgang and Dapeng New District. The socioeconomic situation of each administrative district is presented in table 1.

Table 1. Socioeconomic indicators in Shenzhen in 2022.

Region	Land area (km2)	Permanent population (10 000 people)	Population density (people/km2)	Gross domestic product (billion RMB)
Futian	78.66	151.47	19 256.29	550.36
Luohu	78.75	101.80	12 926.98	260.03
Yantian	74.99	21.15	2820.38	80.96
Nanshan	187.53	181.00	9651.79	815.39
Bao'an	396.61	454.53	11 460.38	468.61
Longgang	388.22	407.36	10 493.02	474.34
Longhua	175.58	249.00	14 181.57	295.19
Pingshan	166.31	60.87	3660.03	112.82
Guangming	155.44	115.09	7404.14	140.32
Dapeng	295.38	16.63	563.00	40.56

^*Data were obtained from the Shenzhen Statistical Yearbook 2023.

This study adopts a stratified sampling strategy to ensure the representativeness and generalisability of the study sample. Significant socioeconomic disparities exist among the 10 administrative districts of Shenzhen City. To fully cover the socioeconomic heterogeneity of urban primary school children, two economically developed districts (Futian and Nanshan) and two economically underdeveloped districts (Guangming and Pingshan) are purposively selected to represent high and low socioeconomic strata in the local region (figure 2). One primary school is randomly sampled from each selected district, which effectively reduces artificial selection bias and ensures the objectivity of sample recruitment. This stratified sampling design enables the enrolled participants to reflect the socioeconomic diversity of first-grade primary school students in Shenzhen.

Figure 2. Location of this study.

Furthermore, quantitative evaluation of sample representativeness is performed in this study. Baseline characteristics of recruited participants, including demographic information, parental myopia prevalence, ocular biometric indicators, as well as environmental and behavioural factors, are compared with the official population data of grade-one students in Shenzhen. Statistical analyses are conducted to verify the demographic and clinical consistency between the study sample and the general paediatric population in the city, so as to confirm the population representativeness and reliability of the research results.

Eligibility criteria

The students must have resided in Shenzhen for at least 6 months to be eligible for inclusion. We will exclude students with a history of retinopathy of prematurity, strabismus, amblyopia, corneal diseases, cataracts, glaucoma, severe opacity of the refractive media or uncontrollable eye tremors. Additionally, those who have undergone previous intraocular or extraocular surgeries and whose parents or legal guardians will not provide consent will be excluded.

Examinations

Demographic information, fundus photography, visual acuity, refractive error, ocular biological parameters, and myopia-related environmental and behavioural factors will be collected from all study participants two times per year. The researchers will include ophthalmologists, optometrists and epidemiologists. Researcher training, fundus image acquisition, ocular biological measurements, visual acuity and refractive error assessments, and epidemiological investigations will be conducted at Shenzhen Eye Hospital. Fundus images will be used to develop a deep learning model based on colour fundus photography. This model aims to achieve automated grading of peripapillary atrophy (PPA) and tessellated fundus, as well as the diagnosis, identification and segmentation of related lesions. All researchers in this study will undergo uniform training and assessment to ensure their qualifications.

Questionnaire survey

An online questionnaire is developed specifically for this study to collect demographic characteristics and information on familial, environmental and behavioural factors potentially associated with myopia in children and adolescents. The questionnaire items are developed with reference to factors commonly investigated in previous epidemiological studies of myopia and are adapted according to the objectives and target population of the present study. During the development process, the content and relevance of the questionnaire items are reviewed by epidemiological experts with relevant research experience, and the questionnaire is subsequently discussed and refined by the research team to improve the clarity and applicability of the items.

The questionnaire collects general information about the students (sex, age, height and weight), their parents (whether they are myopic and the degree of myopia), near-work activities (continuous reading and writing time, time spent using electronic devices, and extracurricular training), time spent outdoors, dietary habits and parents’ cognition of myopia.

Visual acuity test

A standard logarithmic visual acuity chart (Eye Vision 1603-01, China) will be used to evaluate visual acuity with and without correction in compliance with the national standard (GB11533). Distance visual acuity is measured at 5 m. Visual acuity is assessed and recorded for each eye separately. To maintain longitudinal consistency, the same standardised charts and procedures are employed across all data collection rounds for all participants. Visual acuity will be measured according to the criteria for myopia screening in children and adolescents.¹⁸

Non-cycloplegic autorefraction

Non-cycloplegic refraction will be measured using an autorefractor (NIDEK AR-1, Japan). Instrument calibration will be performed daily using a standard simulated eye. Additional measurements will be performed if the difference between the two spherical equivalent refraction (SER) error measurements between the two eyes is ≥0.50 D.

Ocular biological parameters

A multiparameter optical biometer (SW9000, China) will be used to measure the axial length, corneal curvatures, anterior chamber depth, corneal thickness, crystalline lens thickness, corneal diameter and pupil size.

Fundus photography

Fundus images covering potential pathological lesions will be acquired by experienced physicians using a non-mydriatic colour fundus camera (SW-8800, China) approved for clinical use. Subsequently, quantitative analysis of the fundus images will be performed to achieve precise localisation. This process involves the fine extraction and visualisation of tessellated patterns and PPA, ultimately enabling intelligent assessment of fundus images.

Definition of myopia

Students with an uncorrected distance visual acuity <5.0 and SER less than −0.50 D and using orthokeratology lenses will be defined as having myopia.¹⁸ Students with myopia will be divided into high myopia (SER less than −6.00 D), moderate myopia (−6.00 D≤SER less than −3.00 D) and mild myopia (−3.00 D≤SER less than −0.50 D).

Data quality control and statistical analysis

Data quality control

A unified quality control protocol will be implemented throughout the baseline and follow-up surveys. Before commencing the investigation, a standardised survey protocol and operational manual will be developed, and uniform operating guidelines will be adopted. During the training phase, sessions will be conducted with the project leader, technical staff and the personnel responsible for quality control. During the on-site survey process, dedicated quality control personnel will conduct regular inspections to ensure that the research personnel adhere to the survey protocol and perform eye health examinations according to the study guidelines. Periodic checks and feedback on the on-site data quality will be undertaken.

Descriptive statistics

Descriptive statistics will be used to summarise demographic characteristics, ophthalmic parameters, fundus features, relevant covariates, myopia-related lifestyle factors and quantitatively extracted fundus features. Continuous variables will be presented as mean±SD or median (IQR), as appropriate, and categorical variables as frequencies and percentages (%).

Missing data handling

Missing data assessment and handling strategies are prespecified in this longitudinal study. The missing proportion of each variable is calculated, and missing patterns are classified as missing completely at random, missing at random (MAR) or missing not at random. Baseline demographic and clinical characteristics are compared between participants with and without missing data to examine potential systematic selection bias. Multiple imputation is employed to handle missing data under the MAR assumption. The imputation model incorporates demographic factors, ocular biometric parameters, visual and refractive measurements, as well as environmental and behavioural covariates. 20 imputed datasets are generated, and statistical results are pooled according to Rubin’s rules. In addition, complete-case analysis is performed as a sensitivity analysis to validate the robustness of the primary analytical results.

Image preprocessing and association analysis

Each fundus image will undergo standardised preprocessing, including removal of black background areas, centre localisation, normalisation and resizing to a standardised pixel size. Overexposed areas and severe image noise will be addressed when necessary, with black padding applied to image boundaries to preserve peripheral information. Quantitative fundus features will then be extracted from the processed images and assessed longitudinally. We will examine the relationship between the quantitative indicators of fundus images and other indicators (age, sex, vision, refraction, ocular biological parameters, and environmental and behavioural factors). A multiple linear regression model will be established using the quantitative indicators of fundus images as dependent variables and other related factors as independent variables.

Survival analysis and model development

Before survival analysis, the proportional hazards (PH) assumption of the conventional Cox model will be assessed using the Schoenfeld residual test and graphical evaluation of time-dependent residual patterns. Covariates showing evidence of time-varying effects will be considered to violate the PH assumption. For such covariates, stratified Cox regression will be used when the hazard effect is stable over time, whereas time-varying covariate terms will be incorporated when the hazard effect changes substantially during follow-up. These analyses will provide a conventional framework for evaluating associations between candidate predictors and incident myopia.

Beyond the conventional Cox regression analysis of predictor–outcome associations, a separate survival prediction framework will be developed to estimate individualised risks of incident myopia. For this purpose, DeepSurv will be used as the primary machine learning survival model. Unlike the conventional Cox model, DeepSurv does not impose the PH assumption and can flexibly model non-linear associations between predictors and the outcome, as well as complex interactions among covariates, without requiring these relationships to be prespecified.^19–22 This flexibility is particularly relevant to the present study, in which ocular biometric parameters, quantitative fundus image features, demographic characteristics, and environmental and behavioural factors may jointly contribute to myopia development through potentially non-linear and interacting pathways.

The DeepSurv model will be developed to estimate individualised risks of incident myopia at 1, 2 and 2.5 years of follow-up. The model will integrate quantitative features derived from fundus images with ocular biometric, demographic, familial, environmental and behavioural characteristics. The planned predictors will include demographic characteristics, family history of myopia, ocular biometric parameters (including refractive error and axial length), quantitative features extracted from bilateral fundus images, and environmental and behavioural factors. Before feature extraction and model development, fundus images will undergo standardised preprocessing to ensure consistency in image quality and feature representation. The resulting multimodal predictor set will then be used as input to the DeepSurv model for individualised risk estimation at the prespecified follow-up time points. Internal validation will be performed to assess model discrimination and calibration and to evaluate the potential for overfitting.

Model performance evaluation

To assess the predictive performance of DeepSurv, multiple survival prediction models will be developed and evaluated using the same dataset and predictor variables. The conventional Cox PH model will serve as the statistical benchmark, while random survival forest and a gradient boosting survival model (GBM-Cox) will be included as machine learning benchmarks. These approaches provide alternative strategies for modelling associations between predictors and the time-to-event outcome, including potential non-linear relationships and interactions among covariates. To ensure a fair head-to-head comparison, identical predictor variables, data preprocessing procedures and dataset splitting strategies will be applied across all models. Predictive performance for 1-, 2- and 2.5-year myopia risk will be evaluated using complementary measures of discrimination, calibration and clinical utility/net benefit. The comparative analysis will be used to determine whether DeepSurv provides improved or comparable predictive performance relative to the conventional Cox model and other machine learning-based survival models.

Model interpretability and subgroup analysis

Because deep learning-based survival models may have limited interpretability, post hoc model interpretation will be performed for the final DeepSurv model. SHapley Additive exPlanations (SHAP) will be used to quantify the contribution of individual predictors to model-derived myopia risk estimates. SHAP-based analyses will examine the contributions of ocular biometric parameters, quantitative fundus morphological features, demographic characteristics, and behavioural and environmental factors. Both global and individual-level interpretability will be assessed. Global analyses will include feature importance rankings and the overall direction and magnitude of feature contributions, while individual-level SHAP analyses will be used to illustrate how specific predictor profiles contribute to individualised risk estimates. These analyses will provide insight into the relative contribution of different predictor domains and the factors underlying individual predictions.

In addition, stratified subgroup analyses will be conducted according to sex, parental myopia status and baseline refractive status to assess the consistency of model performance and identify potential heterogeneity across clinically relevant population subgroups. Model discrimination, calibration and clinical utility will be examined within these subgroups, and the corresponding interpretability findings will also be compared where appropriate. These analyses will help evaluate the robustness and generalisability of the prediction model across different population characteristics.

Sample size

The sample size of this study is estimated based on a longitudinal follow-up design. Primary schools registered with the Shenzhen Education Bureau are numbered as units. One primary school will be selected randomly from each of the four administrative districts, and all first-grade students from the selected primary schools will be recruited as participants. The sample size for each school will be estimated using the following formula:

$$\mathrm{n}={t_{0.05}}^{2}P(1-P)/d^2$$

P represents myopia prevalence in first-grade students in Shenzhen. The 2019–2021 myopia prevalence in first-grade students in Shenzhen was 13.8%.⁴ The variable d represents the sampling error, set to 5%. We typically use G*Power software²³ to calculate sample sizes; the sample size required for each district is 183. A loss of participants during follow-up is expected; therefore, at least 220 students should be selected from each district (at least 880 students will ultimately be included). The target sample size for the study is 1050 participants.

Follow-up

Participants will be followed up every 6 months for 2.5 years. The follow-up indicators mainly include age, sex, vision, refraction, ocular biological parameters, and environmental and behavioural factors.

Patient and public involvement

The study protocol was developed by a multidisciplinary research team comprising epidemiologists, ophthalmologists and engineers. Patients and members of the public were not involved in the design, planning, conduct or development of the study.

Discussion

Myopia has become a global public health problem.²⁴ The prevalence of myopia has increased worldwide, particularly in children and adolescents in East and Southeast Asia.^{25 26} Myopia development in children and adolescents is gradual. High myopia is often associated with eye complications, including retinal detachment, choroidal neovascularisation, glaucoma and cataracts.^{27 28} Myopia can be corrected using spectacles, contact lenses and refractive surgery; however, fundus lesions caused by myopia cannot be repaired by optical correction.²⁹ Myopia worsens the patients’ quality of life and places a heavy economic burden on their families, society and even the country. The productivity loss caused by uncorrected refractive errors has been estimated to be US$202 billion annually worldwide.³⁰

Children and adolescents are particularly susceptible to myopia. Myopia can be prevented and controlled but is irreversible.³¹ It can develop into pathological myopia, which can cause visual damage or blindness. Therefore, conducting myopia risk assessment in children and adolescents, predicting myopia occurrence and developmental trends, and performing early prevention and control of myopia are crucial. In the present study, we will collect fundus images and data on ocular biological parameters, visual acuity, refractive error, and environmental and behavioural factors to construct a myopia risk prediction model to provide a theoretical basis for personalised prevention and control of myopia.

We will analyse factors that possibly influence the development and diagnosis of myopia from multiple dimensions, including ocular biological parameters, fundus changes, and environmental and behavioural factors (such as parental myopia, near work, outdoor activities, diet and sleep). AI technology will be used to quantitatively process the fundus images of the research participants, providing a basis for exploring myopia’s mechanisms. A DeepSurv model for myopia risk prediction will be constructed using multimodal AI, combining data on ocular biological parameters, quantitative fundus image parameters, and environmental and behavioural factors.

This study has some limitations. First, non-cycloplegic refraction was proven to have high accuracy for myopia screening in large-scale research populations.³² Our study’s use of non-cycloplegic refraction to monitor myopia in students may overestimate the prevalence of myopia.³³ Second, the assessment of environmental and behavioural factors will rely on questionnaire data, which may be subject to self-reporting bias. Participants or their parents may have difficulty accurately recalling past behaviours, resulting in recall bias. In addition, responses may be influenced by social desirability bias, leading to potential overestimation or underestimation of certain behaviours. Although standardised questionnaires and survey procedures will be implemented, the accuracy of self-reported information may still be limited. Future studies will incorporate quantitative and objective monitoring approaches, such as intelligent monitoring devices, to more precisely assess predictive environmental and behavioural factors, including outdoor activity duration. Third, all participants were recruited from coastal regions, which may limit the generalisability of the findings to inland populations. Future multi-centre studies across diverse geographic settings are needed to confirm the robustness of the model. In future studies, we will use quantitative and precise monitoring techniques to evaluate the environmental and behavioural factors deemed to be predictive, including using intelligent monitoring devices to quantify the duration of outdoor activities.

Ethics and dissemination

This study was approved by the Ethics Committee of the Shenzhen Eye Hospital (2024KYPJ023) in accordance with the Declaration of Helsinki. Written informed consent will be obtained from the parents or legal guardians of all participants before enrolment. Participants and their guardians will be informed about the study procedures, the types of data to be collected and the intended use of the study data. All participant information and study data will be handled confidentially and managed in accordance with applicable ethical and institutional requirements.

The findings of this study will be disseminated through publications in peer-reviewed scientific journals and presentations at academic conferences. Key findings relevant to myopia prevention and control will also be communicated to participating schools, parents or guardians, and other relevant stakeholders through appropriate educational and public health communication channels. Where appropriate, summary findings may be used to support evidence-based recommendations for myopia prevention and control. Any sharing or deposition of study data will be conducted in accordance with ethical approval, participant consent and applicable data protection requirements.