Accurate refractive error assessment in pediatric populations necessitates cycloplegic refraction to eliminate accommodation-induced measurement errors, establishing it as the clinical gold standard for both diagnosis and epidemiological studies (Morgan et al., 2015; Doherty et al., 2019; Flitcroft et al., 2019). However, application of cycloplegia faces practical challenges in large-scale pediatric research, including time constraints, resource limitations, and potential contraindications, often precluding its routine use in population-based screenings. To address the conflict between the necessity of cycloplegic refractive examination and the difficulty of performing cycloplegia in practice, emerging studies have explored statistical models to predict cycloplegic refractive error using non-cycloplegic parameters. These models exhibit inconsistent performance (R ² ranging from 0.26 to 0.935) (Ip et al., 2007; Kimura et al., 2007; He et al., 2015; Foo et al., 2016; Sankaridurg et al., 2017; Magome et al., 2021; Wang et al., 2022; Du et al., 2023; He et al., 2023; Ying et al., 2024), likely attributable to methodological heterogeneity, variable predictor selection, and population-specific biases.

Machine learning (ML), a branch of artificial intelligence, offers unparalleled capabilities in identifying complex patterns within high-dimensional datasets. Unlike conventional regression models constrained by linear assumptions, advanced ML algorithms (e.g., random forest [RF], support vector machine [SVM] and gradient boosting models) are capable of modeling intricate and non-linear relationships between predictors and outcomes, and have demonstrated great potential for application in ophthalmology and vision research (Wu et al., 2021; Cao et al., 2022; Wu et al., 2022; Chandra and Ying, 2023; 2024). Recent advancements highlight ML’s ability to predict key refractive parameters such as spherical equivalent refraction (SER) and axial length (AL) in pediatric cohorts. For instance, Zhu et al. (2023) demonstrated that ML models could predict SER and AL changes in children with high accuracy (R ² up to 0.90), leveraging longitudinal data to inform myopia progression. These studies collectively emphasize ML’s capacity to address both diagnostic and prognostic challenges in pediatric ophthalmology, particularly when traditional methods are constrained.

This study aimed to develop and validate ML models for predicting cycloplegic refractive error in Chinese children aged 3–16 years old using readily obtainable non-cycloplegic parameters, encompassing demographics, non-cycloplegic refractive error, and ocular biometric measurements. By leveraging the predictive power of ML, this work seeks to provide a clinically practical solution for determining true refractive status in pediatric populations where cycloplegic examinations are impractical.

This study developed and validated 6 ML models to predict cycloplegic refractive error using non-cycloplegic parameters in a large cohort of Chinese children. The results showed successful predictions of ML models in both the development and validation datasets, demonstrating the promising potential and practical value of ML-based tools to circumvent the challenges of cycloplegia in pediatric refractive assessments while maintaining diagnostic precision.

All the ML models performed well in predicting cycloplegic SER in the validation datasets (R ² ranging from 0.920 to 0.934), with the MLP model performed best. Several studies have employed statistical prediction models to estimate cycloplegic refractive error in pediatric populations, utilizing diverse predictors such as demographics, ocular biometric parameters (Magome et al., 2021), non-cycloplegic refractive error (Sankaridurg et al., 2017; Wang et al., 2022), and uncorrected visual acuity [UCVA](Sankaridurg et al., 2017), alongside modeling methodologies ranging from traditional regression to advanced ML techniques. However, the heterogeneity in study designs and populations has resulted in widely variable predictive performance (R ² : 0.26∼0.935) (Ip et al., 2007; Kimura et al., 2007; He et al., 2015; Foo et al., 2016; Sankaridurg et al., 2017; Magome et al., 2021; Wang et al., 2022; Du et al., 2023; He et al., 2023; Ying et al., 2024). In particular, Du et al. (2023) developed 4 ML models (Support Vector Regression, Random Forest Regression, AdaBoost Regression, and Deep Neural Network) in 2,467 Chinese children aged 6–18 years, achieving R ² of 0.899∼0.927 and MAE of 0.372∼0.436 D. Among them, AdaBoost regression model exhibited the best overall prediction performance, with R ² = 0.927 and MAE = 0.372, slightly inferior to the optimal MLP-based model in our study. However, some predictors used in their models (e.g., the accommodation lag data) may not be easily available in clinical practice, which would limit the models’ applicability. Ying et al. (2024) also developed and validated 6 ML models based on 3,414 Chinese children ages 5–18 years by using age, gender, non-cycloplegic refractive error, biometric measures (AL, CR, AL/CR ratio, CCT, and anterior chamber depth), UCVA, intraocular pressure, and glasses-wearing status. The models predicted cycloplegic SER with R ² of 0.913∼ 0.935, and MAE of 0.393∼0.480 D, with top performance comparable to ours. Notably, while the ML methods used in our study were largely similar to those employed by Ying et al., the optimal model diverged (MLP vs. XGBoost). This discrepancy may be attributed to cohort differences: their population comprised older children (5∼18 years), with nearly 50% was low hyperopia (+0.50∼+3.00 D); whereas ours encompassed younger ages (3∼16 years old), dominated by low myopia (−3.00∼-0.50 D) and low hyperopia (+0.50∼+2.00 D). In comparison to prior work, our study prioritizes clinical practicality by exclusively incorporating parameters easily obtainable without cycloplegia, while maintaining robust accuracy (MLP model: R ² = 0.934). This balance between simplicity and precision positions our framework as a scalable solution for population-based studies, particularly in resource-limited settings where cycloplegic examinations are challenging. Furthermore, the included younger children (3–16 years old) with varied refractive status (from −16.13 to +7.88 D) enhances the model’s clinical relevance for early myopia risk stratification.

Substantial evidence demonstrates clinically significant discrepancies between cycloplegic and non-cycloplegic measurements, with non-cycloplegic SER systematically overestimated myopia (Zhao et al., 2004; Fotedar et al., 2007; Hu et al., 2015; Sankaridurg et al., 2017; Gu et al., 2022). Consistent with previous studies, such bias was also observed in our study: the mean SER changed from −0.04 ± 2.11 D to −0.43 ± 1.85 D under the non-cycloplegic condition with a mean difference of 0.39 ± 0.71 D toward myopia. Our systematic analysis of paired differences between cycloplegic and non-cycloplegic refraction across ages 3–16 years and all types of refractive error revealed that the overestimation of myopia or underestimated hyperopia with non-cycloplegia is greater in younger individuals, emmetropic or hyperopic eyes, and shorter AL. Accommodation significantly impacts the accurate measurement of refraction data in children. Age-related norms for accommodative responses in eyes without significant refractive error suggest that accommodative responses are quite variable under high accommodative demands, especially in younger individuals (McClelland and Saunders, 2004). Consequently, predicting refractive error without cycloplegia is challenging in young children. The inclusion of children under 5 years of age in this study provides a basis for addressing the difficulty of predicting cycloplegic refractive error in this age group.

The MLP model demonstrated robust clinical accuracy, with 73.08% and 94.29% of predictions falling within ±0.50 D and ±1.00 D of measured cycloplegic SER, respectively. This performance outperformed the traditional regression-based approach by He et al. (2023), which achieved an accuracy of 47% within ±0.50 D and 79% within ±1.00 D, using limited predictors (age, gender, AL and AL/CR ratio). While Du et al. (2023) reported marginally higher precision (75.2% within ±0.50 D) in their best-performing model, our framework maintains comparable efficacy while prioritizing clinically accessible parameters. Notably, the accuracy of MLP model peaked in children aged 7∼10 years (77.17∼79.70% within ±0.50 D) and those with low myopia (83.09∼83.56% within ±0.50 D). These subgroups represent pivotal stages for myopia surveillance and early intervention to mitigate progression to high myopia. These results position the ML model as a pragmatic tool for large-scale screenings in resource-limited settings.

There were several limitations in the present study. First, this study used the AB-800 optical biometer in a cohort of Chinese children, which may restrict the generalizability of the findings to other types of biometers or children of different races or ethnicities. External validation across diverse populations and devices is warranted. Second, the reduced accuracy in high myopia (≤−6.0 D) and moderate/high hyperopia (>+2.0 D) limits utility for these subgroups. Expanding data with extreme refractive errors and retraining models on enriched cohorts could improve performance and enhance the model’s applicability. Finally, the current ML models predict the cycloplegic refractive error at a single timepoint. Future longitudinal studies can evaluate the models’ capacity to track refractive progression over time.

This study developed and validated 6 ML models for predicting cycloplegic refractive error using readily available demographics and non-cycloplegic biometric parameters in a large cohort of Chinese children (3∼16 years). Among the evaluated algorithms, the MLP model demonstrated superior performance (R ² = 0.934), and exhibited enhanced accuracy in children aged 7–10 years and those with low myopia—key populations for myopia surveillance and intervention. These ML models address practical challenges in pediatric refractive assessments, offering a scalable tool for clinical practice and large-scale epidemiological research where administering cycloplegia is not feasible.