Machine learning classification of cycloplegic refractive status in children and adolescents using non-cycloplegic data across ten provincial-level administrative divisions.

Machine learning classification of cycloplegic refractive status in children and adolescents using non-cycloplegic data across ten provincial-level administrative divisions - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMC Med Inform Decis Mak . 2026 Apr 22;26:206. doi: 10.1186/s12911-026-03490-5 Machine learning classification of cycloplegic refractive status in children and adolescents using non-cycloplegic data across ten provincial-level administrative divisions Keke Liu 1, 2 , Huijuan Luo 2 , Yan Chen 1, 2 , Boran E 1, 2 , Huining Kuang 1, 2 , Chenyu Zhang 1, 2 , Xin Guo 3, ✉ Author information Article notes Copyright and License information 1 School of Public Health, Capital Medical University, No. 10, West Toujiao, Right Anmenwai, Fengtai District, Beijing, 100069 People’s Republic of China 2 School Health Center, Beijing Center for Disease Prevention and Control, No. 16, Hepingli Middle Street, Dongcheng District, Beijing, 100013 People’s Republic of China 3 Chinese Center for Disease Control and Prevention, No. 155 Changbai Road, Beijing, 102206 People’s Republic of China ✉ Corresponding author. Received 2025 Dec 22; Accepted 2026 Apr 10; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13235058  PMID: 42021212 Abstract Background Cycloplegic refraction remains the gold standard for assessing refractive error in children, yet its application in large-scale screening is limited. Machine learning offers a promising approach for screening and triage by leveraging non-cycloplegic data for refractive classification. Methods This study utilized a multicenter dataset comprising 54,374 participants aged 5–18 years from ten provincial-level administrative divisions (PLADs) across China. We developed and validated four machine learning models (LightGBM, XGBoost, Random Forest, and MLP) and a multinomial logistic regression model as a conventional baseline comparator to classify refractive status (myopia, emmetropia, hyperopia). Model generalizability was evaluated via leave-one-PLAD-out cross-validation. Feature importance, calibration, and age-stratified trend analyses were conducted to assess interpretability and developmental robustness. Results All models achieved excellent discrimination (macro-AUC 0.910–0.913) and maintained stable performance across PLADs (AUC 0.866–0.950). While all ensemble models achieved comparable discrimination (macro-AUC 0.910–0.913), LightGBM demonstrated slightly lower log-loss and more balanced probability calibration. Therefore, it was selected as the primary model for subsequent age-stratified analyses. Key predictive features included non-cycloplegic SE, AL/CR ratio, UCVA, and axial length. Age-stratified evaluation revealed systematic performance improvement from childhood to adolescence, with accuracy increasing from 0.80 (5–8 years) to 0.89 (16–18 years) and AUC peaking at 0.93 (12–15 years). Linear trend analysis confirmed significant age-related enhancement independent of provincial effects. Conclusions LightGBM-based classification of cycloplegic refractive status from non-cycloplegic optical and biometric data provides a robust, generalizable, and clinically interpretable approach for pediatric vision screening. The model is best positioned as a screening and triage support tool, enabling risk stratification and prioritization for confirmatory cycloplegic refraction rather than replacing it, particularly in large-scale or resource-limited settings. Supplementary Information The online version contains supplementary material available at 10.1186/s12911-026-03490-5. Keywords: Cycloplegic refraction, Non-cycloplegic measurement, Myopia screening, Machine learning, Children and adolescents, Refractive error classification Introduction Refractive errors, predominantly myopia, account for one-fifth of global blindness cases [ 1 ]. The condition has reached pandemic proportions worldwide, with disproportionately high prevalence rates observed among Asian populations [ 2 , 3 ]. Accurate identification and classification of myopia can help identify children at high risk, enabling earlier and more effective interventions to slow the onset or progression of myopia, thereby improving visual outcomes and quality of life [ 4 ]. Cycloplegic refraction is the gold standard for detecting and quantifying refractive error [ 5 ]; however, its large-scale application in epidemiological surveys is limited by potential adverse reactions to cycloplegic agents [ 6 ] and the risk of misclassification in refractive error assessment [ 7 , 8 ]. To address this challenge, there has been increasing interest in developing automated screening tools based on non-cycloplegic ocular biometric parameters, such as axial length (AL) and corneal radius (CR) [ 9 ]. Machine learning (ML) models, which can capture complex nonlinear interactions between covariates and outcomes, have shown great potential in ophthalmology and vision science by achieving higher accuracy and adaptability to data characteristics [ 10 , 11 ]. Previous studies have demonstrated the feasibility of using various ML algorithms, including random forest and support vector machines, for refractive error classification [ 12 , 13 ]. However, most existing studies relied on single-province datasets with limited representativeness and lacked external geographic validation, raising concerns about the generalizability and stability of these models across diverse populations. Cross-validation is a classic method in machine learning used to evaluate model generalization and prevent overfitting. Conventional k-fold or Leave-One-Out cross-validation (LOO-CV) randomly partitions data to estimate internal model performance [ 14 ]. However, such approaches often overlook structural heterogeneity at the geographic or population level, potentially leading to overly optimistic estimates of real-world performance. To overcome this limitation, we employed a more stringent validation framework—Leave-One-PLAD-Out cross-validation (LOPO-CV). In this approach, data from one province are excluded as an external test set while the remaining nine provinces are used for model training. This design simulates model deployment in unseen geographic regions and provides a more accurate assessment of geographic generalizability [ 15 ]. Furthermore, understanding model performance across key developmental stages is crucial for optimizing the implementation of vision screening programs, yet this aspect has not been thoroughly explored in prior research. In this multicenter cross-sectional study, we included children and adolescents from ten provinces across China. Using large-scale, population-based non-cycloplegic data, we developed and validated machine learning models for refractive status classification. Specifically, we aimed to: (1) compare the performance and generalizability of four ML algorithms against a multinomial logistic regression baseline using the rigorous LOPO-CV framework; (2) evaluate model calibration by comparing predicted versus observed refractive error distributions; (3) identify the most influential predictors across algorithms; and (4) assess the performance of the optimal ML model across four key developmental stages from early childhood to late adolescence. Our findings are expected to inform the development of robust and generalizable AI-based tools for scalable refractive error screening and triage in diverse pediatric populations, rather than serving as standalone diagnostic replacements. The primary contribution of this work lies in its large-scale, geographically diverse dataset and the rigorous leave-one-PLAD-out validation framework, which together strengthen the evidence for real-world generalizability. Methods Study population This study utilized data from a nationwide cross-sectional survey conducted in China between 2020 and 2024 and coordinated by the National Disease Control and Prevention Administration. The survey covered ten provincial-level administrative divisions (PLADs): Beijing, Shanxi, Liaoning, Zhejiang, Shandong, Henan, Hunan, Guangdong, Chongqing, and Shaanxi. Within each PLAD, cities or districts were selected to represent both more and less economically developed areas. A multistage cluster sampling design was then applied, with schools and classes serving as sampling units. All eligible students from senior kindergarten to high school within selected classes were invited to participate. A total of 54,374 participants aged 5–18 years were included in the present analysis.The complete data collection and processing workflow is detailed in Supplementary Figure S1 . Ocular examination protocol and refractive classification All participants underwent standardized ocular examinations following a unified protocol administered by trained and certified personnel. Cycloplegic autorefractionation was conducted using identical desktop autorefractor models across all sites, with cycloplegia induced by the administration of 0.5% tropicamide eye drops administered four times at 5-minute intervals. Comprehensive measurements included uncorrected visual acuity (UCVA) (recorded using the Chinese 5-point scale, where higher values indicate better visual acuity), cycloplegic spherical equivalent (SE), AL, CR, anterior chamber depth (ACD), intraocular pressure (IOP), and axial length-to-corneal radius ratio (AL/CR). For statistical consistency and to avoid inter-eye dependency, only data from the right eye of each participant were included in the present analysis. Based on cycloplegic spherical equivalent measurements, participants were classified into three refractive error categories: myopia (SE ≤ -0.50 diopters, D), emmetropia (SE > -0.50 D and ≤ + 0.50 D), and hyperopia (SE > + 0.50 D). The study received ethical approval from the Beijing Municipal Center for Disease Control and Prevention [Approval No. (24), 2022], and all participants provided informed consent. Statistical analysis All statistical analyses were performed using R (version 4.5.1). A fixed random seed (set.seed = 20250828) was applied to ensure reproducibility. Continuous variables are presented as mean ± standard deviation, and categorical variables as counts and percentages. Group comparisons were conducted using ANOVA or Kruskal–Wallis tests for continuous variables and chi-square tests for categorical variables. Univariate logistic regression (α = 0.05), variance inflation factor (VIF), and LASSO regression were performed as exploratory analyses to assess crude associations, multicollinearity, and variable selection stability. These procedures were primarily used to support the interpretation of the multinomial logistic regression baseline model rather than for feature selection. No variables were excluded based on these analyses; all candidate predictors were retained in the final machine learning models, which inherently perform embedded feature selection. Machine learning modeling and validation We implemented a tidymodels-based framework to compare four machine learning algorithms—Random Forest, XGBoost, LightGBM, and MLP—with multinomial logistic regression as a baseline under the same LOPO-CV framework. Data splitting was conducted at the participant level, ensuring independence between training and test sets in each iteration. Model generalizability was evaluated using LOPO-CV, with hyperparameter optimization performed via nested 5-fold cross-validation within training data. Hyperparameter search spaces and tuning details are provided in Supplementary Table S1 . To assess model dependence on key predictors, a sensitivity analysis was conducted by re-training all models after excluding non-cycloplegic SE. Additionally, to evaluate the impact of class imbalance handling, models were re-trained without SMOTE using the same validation framework and evaluation metrics. In each LOPO iteration, one PLAD was held out as an independent test set, while SMOTE was applied only to the remaining training data within the resampling pipeline to prevent data leakage. Model discrimination and calibration were evaluated on the untouched test sets. Performance was assessed using macro-AUC, accuracy, Cohen’s κ, macro F1-score, and log loss, with results aggregated by test-set-size weighting. Age-stratified evaluation of the best-performing model (LightGBM) was conducted across four developmental stages (5–8, 9–11, 12–15, and 16–18 years). Results Demographic and ocular characteristics The study included 54,374 participants, categorized into myopia (13,623), emmetropia (12,067), and hyperopia (28,684) groups. Myopes were significantly older (11.30 ± 3.04 years) than emmetropes (8.98 ± 2.40 years) and hyperopes (7.78 ± 1.82 years; P < 0.001).The proportion of females was higher in the myopia group (51.4%) compared to males (48.6%), whereas the proportion of males was higher in the emmetropia group (53.3%) compared to females (46.7%). These differences in sex distribution across refractive categories were statistically significant ( P < 0.001). Significant variation in refractive error distribution was observed across PLADs ( P < 0.001). Using row percentages to estimate within-province prevalence, Liaoning had the highest prevalence of myopia (43.2%), whereas Hunan had the highest prevalence of hyperopia (66.6%). Non-cycloplegic SE measurements indicated a mean of -0.59 ± 1.38 D overall, with myopes having a more negative SE (-2.22 ± 1.28 D) and hyperopes a positive SE (0.11 ± 0.91 D; P < 0.001). Myopes had better UCVA (4.53 ± 0.34 LogMAR; P < 0.001). Ocular biometric parameters, including axial length (24.18 ± 0.96 mm), AL/CR ratio (3.11 ± 0.11), and anterior chamber depth (3.55 ± 0.33 mm), varied significantly across groups (all P < 0.001) (Table 1 ). Table 1. Baseline characteristics of study participants by refractive error category Characteristic Overall ( n = 54,374) Myopia ( n = 13,623) Emmetropia ( n = 12,067) Hyperopia ( n = 28,684) P -value Age (years) , mean ± SD 8.93 ± 2.73 11.30 ± 3.04 8.98 ± 2.40 7.78 ± 1.82 < 0.001 Sex , n (%) Male 27,815 (51.2) 6,627 (48.6) 6,435 (53.3) 14,753 (51.4) < 0.001 Female 26,559 (48.8) 6,996 (51.4) 5,632 (46.7) 13,931 (48.6) PLAD , n (%) Beijing 8,188 (15.1) 1,855 (22.7) 1,788 (21.8) 4,545 (55.5) < 0.001 Chongqing 1,669 (3.1) 512 (30.7) 316 (18.9) 841 (50.4) Guangdong 9,354 (17.2) 1,877 (20.1) 1,702 (18.2) 5,775 (61.7) Henan 3,577 (6.6) 1,340 (37.5) 786 (22.0) 1,451 (40.6) Hunan 10,817 (19.9) 1,581 (14.6) 2,034 (18.8) 7,202 (66.6) Liaoning 3,310 (6.1) 1,429 (43.2) 830 (25.1) 1,051 (31.8) Shaanxi 3,132 (5.8) 815 (26.0) 670 (21.4) 1,647 (52.6) Shandong 2,728 (5.0) 714 (26.2) 601 (22.0) 1,413 (51.8) Shanxi 8,231 (15.1) 2,465 (29.9) 2,654 (32.2) 3,112 (37.8) Zhejiang 3,368 (6.2) 1,035 (30.7) 686 (20.4) 1,647 (48.9) Non-cycloplegic SE (D) , mean ± SD -0.59 ± 1.38 -2.22 ± 1.28 -0.41 ± 0.75 0.11 ± 0.91 < 0.001 Uncorrected Visual Acuity (Chinese 5-point scale) , mean ± SD 4.86 ± 0.27 4.53 ± 0.34 4.96 ± 0.12 4.97 ± 0.11 < 0.001 Axial length (mm) , mean ± SD 23.18 ± 1.01 24.18 ± 0.96 23.20 ± 0.74 22.69 ± 0.73 < 0.001 Corneal radius (mm) , mean ± SD 7.78 ± 0.26 7.77 ± 0.26 7.78 ± 0.26 7.79 ± 0.26 < 0.001 Axial Length/Corneal Radius ratio , mean ± SD 2.98 ± 0.12 3.11 ± 0.11 2.98 ± 0.07 2.91 ± 0.07 < 0.001 Anterior Chamber Depth (mm) , mean ± SD 3.40 ± 0.38 3.55 ± 0.33 3.42 ± 0.33 3.32 ± 0.39 < 0.001 Intraocular Pressure (mmHg) , mean ± SD 17.54 ± 2.89 17.82 ± 2.84 17.69 ± 2.81 17.34 ± 2.93 < 0.001 Open in a new tab Abbreviations: PLAD, provincial-level administrative division; SE, spherical equivalent; D, diopter; LogMAR, logarithm of the minimum angle of resolution; SD, standard deviation Overall classification performance of machine learning models All models demonstrated excellent generalizability in leave-one-province-out cross-validation (Table 2 ). The multinomial logistic regression baseline achieved a macro-AUC of 0.894 ± 0.023 and accuracy of 0.746 ± 0.041, which were consistently lower than those of the ensemble models. The ensemble methods achieved outstanding discriminative performance, with macro-AUC values of 0.913 for XGBoost, 0.913 for LightGBM, and 0.910 for Random Forest. MLP also showed comparable AUC (0.913). XGBoost yielded the highest accuracy (78.2% ± 3.9%), closely followed by Random Forest (78.1% ± 4.1%) and LightGBM (78.0% ± 3.8%). All models maintained consistent macro F1 scores around 0.752. The tree-based algorithms exhibited superior probability calibration, as evidenced by lower log-loss values (0.504–0.518) compared to MLP (0.803). The minimal performance variations across geographical regions underscore the robustness of our approach. Table 2. Model performance in leave-one-PLAD-out (LOPO) cross-validation Model Accuracy AUC (macro) F1-score (macro) Kappa Log-loss multinomial logistic regression 0.746 ± 0.041 0.894 ± 0.023 0.722 ± 0.034 0.584 ± 0.055 0.594 ± 0.076 LightGBM 0.780 ± 0.038 0.913 ± 0.019 0.752 ± 0.030 0.633 ± 0.047 0.505 ± 0.081 XGBoost 0.782 ± 0.039 0.913 ± 0.019 0.752 ± 0.031 0.635 ± 0.048 0.504 ± 0.084 Random Forest 0.781 ± 0.041 0.910 ± 0.021 0.751 ± 0.033 0.633 ± 0.049 0.518 ± 0.077 MLP 0.772 ± 0.032 0.913 ± 0.019 0.752 ± 0.024 0.625 ± 0.039 0.803 ± 0.020 Open in a new tab Abbreviations: AUC, area under the ROC curve; LightGBM, Light Gradient Boosting Machine; XGBoost, Extreme Gradient Boosting; MLP, Multi-Layer Perceptron Geographical robustness and cross-provincial generalizability All models demonstrated excellent generalizability in leave-one-province-out cross-validation (Fig. 1 ; Supplementary Table S2 ). The ensemble methods consistently achieved superior performance across diverse geographical regions, with macro-AUC values ranging from 0.866 to 0.950 across all provinces. Tree-based algorithms (Random Forest, XGBoost, LightGBM) outperformed MLP in overall accuracy (69.2%-83.8%) and probability calibration, as evidenced by substantially lower log-loss values (0.389–0.531 vs. 0.756–0.840 for MLP). The remarkably low standard deviations in AUC (0.000-0.004) underscore the stability of discriminative performance across geographical hold-outs. Notably, Random Forest achieved the highest accuracy in multiple provinces, while all models maintained consistent performance rankings across different regions. Chongqing showed the strongest overall performance (AUC: 0.949–0.950), whereas Shaanxi presented the most challenging classification scenario. Fig. 1. Open in a new tab Cross-provincial validation of machine learning models for refractive status classification. Abbreviations: AUC, area under the ROC curve; ROC, receiver operating characteristic; LightGBM, Light Gradient Boosting Machine; XGBoost, Extreme Gradient Boosting; MLP, Multi-Layer Perceptron Classification calibration and clinical distribution fidelity Confusion matrices (Fig. 2 A) and predicted-versus-actual distribution plots (Fig. 2 B) indicated strong concordance between predicted and observed refractive state proportions. For hyperopia, the most prevalent condition (52.8% true prevalence), tree-based models demonstrated accurate calibration with predicted proportions ranging from 52.9% to 53.8%. However, MLP substantially under-predicted hyperopia prevalence (49.2% predicted vs. 52.8% true). For myopia (25.1% true prevalence), all models consistently under-predicted the condition, with predicted proportions ranging from 23.7% to 24.0%. This systematic underestimation suggests a potential calibration bias across all algorithms for myopia detection. Most notably, MLP exhibited significant miscalibration for emmetropia, over-predicting its prevalence by 4.7% points (26.9% predicted vs. 22.2% true). In contrast, ensemble methods maintained much closer alignment with the true emmetropia distribution (22.4–23.1% predicted). Among the ensemble approaches, LightGBM showed the most balanced calibration across all three clinical conditions, with deviations within 1.5% points for each class. Fig. 2. Open in a new tab Comparative analysis of machine learning classifiers for refractive status classification. ( A ) Comparative Confusion Matrix Analysis of Four Machine Learning Classifiers for Refractive Status Classification; ( B ) Comparative Analysis of Actual and Classified Refractive Status Distributions Using Four Machine Learning Approaches. Abbreviations: LightGBM, Light Gradient Boosting Machine; XGBoost, Extreme Gradient Boosting; MLP, Multi-Layer Perceptron; RF, Random Forest Feature importance analysis The feature importance analysis (Figure S2) reveals that across the LightGBM, Random Forest, and XGBoost models, non-cycloplegic SE measured by autorefraction, AL/CR, UCVA, and AL are the most critical discriminative factors for all algorithms. These findings highlight the heterogeneity in variable importance across different machine learning algorithms and underscore the necessity for algorithm-specific feature prioritization to optimize predictive performance. Age-stratified model evaluation Our age-specific evaluation revealed a systematic improvement in model performance with increasing age, underscoring the enhanced predictability of the LightGBM model in older children and adolescents (Fig. 3 ). Classification accuracy steadily increased from 0.80 in children aged 5–8 years to 0.89 in adolescents aged 16–18 years, while the discriminative ability, as reflected by macro-AUC, remained consistently high across all age strata (AUC > 0.88), reaching a peak of 0.93 in the 12-15-year group. Parallel gains in the F1-score and Cohen’s κ (from 0.50 to 0.71) and a notable decline in log-loss (from 0.46 to 0.23) collectively indicated better classification consistency, calibration, and class balance with age progression. These results confirm the model’s robust generalizability across pediatric developmental stages, suggesting optimal operational characteristics during adolescence when refractive differentiation becomes more stable. Fig. 3. Open in a new tab Age-stratified performance analysis of the LightGBM model for refractive status classification. ( A ) Radar plot illustrating performance metrics by age groups; ( B ) Receiver operating characteristic (ROC) curves stratified by age groups and refractive status; ( C ) Bar plot comparing classification metrics across age groups. Abbreviations: AUC, area under the curve Sensitivity analysis Excluding non-cycloplegic SE reduced model performance, with macro-AUC decreasing from approximately 0.913 to 0.872–0.877 (Supplementary Table S3 ), although discrimination remained acceptable across all models. Models trained without SMOTE showed comparable performance to the primary analysis (Supplementary Table S4 ), with macro-AUC values above 0.91 and minimal changes in log-loss. Importantly, the relative ranking of models and overall study conclusions remained unchanged, indicating that oversampling did not materially influence model discrimination or calibration. Discussion This large-scale, cross-sectional study, encompassing 54,374 children and adolescents from ten provincial-level administrative divisions (PLADs) across China, establishes that tree-based machine learning models, leveraging routinely collected non-cycloplegic SE and ocular biometric parameters, can achieve high-fidelity concurrent classification of cycloplegic refractive status. Rather than proposing novel algorithms, the primary contribution lies in demonstrating robust and geographically stable performance using a large, multicenter dataset and a stringent leave-one-PLAD-out validation framework. Given the highly similar discriminative performance among ensemble methods, the designation of LightGBM as the primary model was based on its more favorable calibration characteristics. The superiority of ensemble models over multinomial logistic regression further supports the incremental value of nonlinear modeling in capturing complex interactions among biometric features. The pursuit of accurate non-cycloplegic refraction estimation has yielded inconsistent results in prior literature, with reported R² values varying widely from 0.26 to 0.93 [ 12 , 16 – 18 ]. These discrepancies likely stem from limitations in sample size, population homogeneity, and study design. While studies such as that by Du et al. reported promising outcomes (accuracy 80.3%-81.7%), their generalizability to broader, heterogeneous populations remained unproven [ 13 ]. By leveraging a substantially larger and geographically diverse cohort across ten PLADs and employing rigorous regional hold-out validation, we provide compelling evidence for the model’s robustness, with AUCs consistently maintained across PLADs (range: 0.866–0.950). Probability calibration quality is a key determinant of the clinical utility of screening models. Our study found that ensemble methods significantly outperformed MLP in predictive distribution calibration, especially in the identification of hyperopia and emmetropia. The systematic underestimation of hyperopia by MLP (49.2% predicted vs. 52.8% actual) and overestimation of emmetropia (26.9% predicted vs. 22.2% actual) may be attributed to the optimization objectives of neural networks, which tend to optimize global loss functions [ 19 ]. In contrast, tree-based models, which integrate numerous decision trees, generate more robust probability outputs [ 20 ]. A critical observation, consistent across all models, was the slight underestimation of myopia prevalence (23.7–24.0% predicted vs. 25.1% actual). This suggests a fundamental limitation of current non-cycloplegic biometric parameters in identifying the very earliest stages of myopia, where latent accommodative tone may not be fully captured, pointing to the need for incorporating more sensitive, multidimensional indicators in future iterations. Feature importance analysis revealed highly consistent feature rankings across different ensemble algorithms: non-cycloplegic SE, AL/CR, UCVA, and AL consistently emerged as the most critical predictors. Notably, non-cycloplegic SE was the single most dominant predictor across all models, which is expected given its close physiological relationship with cycloplegic refractive status. This dominance, however, also constrains the degree of novelty and limits mechanistic interpretation, as the model partially leverages an already correlated proxy of the outcome rather than entirely independent predictors. Sensitivity analysis excluding non-cycloplegic SE resulted in reduced but still acceptable model performance, confirming that while this feature contributes substantially to predictive accuracy, other ocular biometric parameters also provide meaningful discriminative information. Baseline refractive status and ocular biometric measurements have been identified as significant risk factors for the development and progression of myopia [ 21 ]. Given their ease of collection, they have been employed in previous studies to predict cycloplegic refractive error [ 16 , 18 , 22 ]. Therefore, the primary contribution of this study lies not in introducing novel predictors, but in demonstrating robust, generalizable, and well-calibrated classification performance across diverse populations using a scalable framework. Subtle differences in secondary feature importance among algorithms reflect their distinct feature interaction modeling strategies, providing guidance for algorithm selection in various PLADs. Age-stratified analyses showed improving predictive metrics with age, reflecting biological and measurement factors rather than model superiority. Cycloplegia was induced with 0.5% tropicamide, which provides less complete accommodation paralysis than cyclopentolate, especially in younger children. Thus, residual accommodation may persist in early childhood—where accommodative responses are stronger [ 23 ]—leading to incomplete cycloplegic refraction and potential outcome misclassification. This adds noise to labels, reducing feature separability and model performance, as typical under measurement error [ 24 ]. Conversely, in older children and adolescents, accommodation is weaker and cycloplegia more effective, yielding more accurate refractive classification. Adolescence also features more stable axial elongation and narrower refractive variance [ 25 ], contributing to a less noisy data distribution. These findings support age-specific screening strategies: closer monitoring in younger children given higher measurement noise, and efficient large-scale screening in adolescents where predictive discrimination is higher. Although accuracy was lower in children aged 5–8 years (0.80), macro-AUC remained above 0.88, indicating preserved discrimination. Given biological variability and potential incomplete cycloplegia, ML-based screening in this age group should complement, not replace, cycloplegic assessment. Several limitations warrant consideration. First, the cross-sectional design precludes causal inference and temporal prediction of incident or progressive myopia. Second, despite leave-one-PLAD-out validation, uneven regional sampling may affect model stability. Third, cycloplegia was induced with 0.5% tropicamide, which may result in incomplete cycloplegia, particularly in younger children with stronger accommodative responses. This may introduce systematic misclassification in the reference standard and partially explain the reduced model performance observed in younger age groups. Fourth, although feature importance improves interpretability, ensemble models remain partially “black-box.” Although SMOTE was restricted to training folds, synthetic oversampling may theoretically introduce artificial patterns; however, stable performance on untouched test sets and consistent results in sensitivity analyses without SMOTE suggest that its impact on model discrimination and calibration is likely limited. Finally, slight underestimation of myopia suggests room for refinement. From an implementation perspective, the model is intended as a decision-support and triage tool rather than a standalone diagnostic system. Required inputs (non-cycloplegic refraction, axial length, and basic demographics) are routinely collected in school-based screening programs, enabling integration into existing electronic health platforms. The algorithm can generate real-time risk stratification to prioritize referrals for cycloplegic confirmation, thereby enhancing screening efficiency in high-burden or resource-limited settings. Overall, these findings highlight the potential of tree-based machine learning models as scalable tools for population-level refractive screening. Future work should incorporate longitudinal data, environmental exposures, and dynamic ocular biometrics to enhance temporal prediction and causal inference, and extend validation to international cohorts to assess cross-population transferability. Conclusion Our study demonstrates that ensemble machine learning models, particularly LightGBM, can accurately and robustly classify concurrent cycloplegic refractive status using non-cycloplegic data among Chinese children and adolescents. The models maintained high discriminative capacity and calibration across diverse PLADs and age groups, with performance systematically improving through adolescence. From a medical informatics perspective, the framework enables data-driven classification within existing screening workflows and supports tiered referral, prioritizing high-risk individuals for confirmatory cycloplegic examination rather than replacing clinical diagnosis. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (277.7KB, docx) Acknowledgements The authors gratefully acknowledge the contributions of all staff members involved in data collection and extend their sincere thanks to all the students who participated in this study. Abbreviations AL Axial length CR Corneal radius ML Machine learning LOO-CV Leave-One-Out cross-validation LOPO-CV Leave-One-PLAD-Out cross-validation PLADs Provincial-level administrative divisions UCVA Uncorrected visual acuity SE Spherical equivalent ACD Anterior chamber depth IOP Intraocular pressure AL/CR Axial length-to-corneal radius ratio VIF Variance inflation factors LightGBM Light Gradient Boosting Machine MLP Multilayer Perceptron SMOTE Synthetic Minority Over-sampling Technique macro-AUC Macro-averaged area under the curve Author contributions Keke Liu: Writing-original draft, Visualization, Methodology, Formal analysis, Conceptualization. Huijuan Luo: Software, Methodology. Yan Chen: Data curation, Visualization. Boran E: Data curation, Visualization. Huining Kuang: Data curation, Visualization. Chenyu Zhang: Methodology, Formal analysis. Xin Guo: Writing-review & editing, Supervision, Funding acquisition, Conceptualization. Funding Supported by the Ministry of Science and Technology of the People’s Republic of China, National Key Research and Development Program of China (2021YFC2702102), the Beijing Municipal Health Commission High-level Public Health Technical Talent Construction Project (Leading Talent-01-09), and The Operation of Public Health Emergency Response Mechanism (102393220020010000017). Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate This study received ethics approval from the Institutional Review Board of the Beijing Centers for Disease Prevention and Control (2022 No.24) and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants. For participants aged 16 or older, consent was provided directly. For participants under 16, consent was obtained from their parents or legal guardians. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Bourne R, Flaxman SR, Braithwaite T, Cicinelli MV, Das A, Jonas JB, Keeffe J, Kempen JH, Leasher J, Limburg H, et al. Magnitude, temporal trends, and projections of the global prevalence of blindness and distance and near vision impairment: a systematic review and meta-analysis. Lancet Glob Health. 2017;5(9):e888–97. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Wolffsohn JS, Calossi A, Cho P, Gifford K, Jones L, Jones D, Guthrie S, Li M, Lipener C, Logan NS, et al. Global trends in myopia management attitudes and strategies in clinical practice – 2019 Update. Cont Lens Anterior Eye. 2020;43(1):9–17. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Holden BA, Fricke TR, Wilson DA, Jong M, Naidoo KS, Sankaridurg P, Wong TY, Naduvilath TJ, Resnikoff S. Global Prevalence of Myopia and High Myopia and Temporal Trends from 2000 through 2050. Ophthalmology. 2016;123(5):1036–42. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Sankaridurg P, Tahhan N, Kandel H, Naduvilath T, Zou H, Frick KD, Marmamula S, Friedman DS, Lamoureux E, Keeffe J, et al. IMI Impact of Myopia. Invest Ophthalmol Vis Sci. 2021;62(5):2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Flitcroft DI, He M, Jonas JB, Jong M, Naidoo K, Ohno-Matsui K, Rahi J, Resnikoff S, Vitale S, Yannuzzi L. IMI - Defining and Classifying Myopia: A Proposed Set of Standards for Clinical and Epidemiologic Studies. Invest Ophthalmol Vis Sci. 2019;60(3):M20–30. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Goel S, Maharajan P, Chua C, Dong B, Butcher M, Bagga P. Driving ability after pupillary dilatation. Eye (Lond). 2003;17(6):735–8. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Zhao J, Mao J, Luo R, Li F, Pokharel GP, Ellwein LB. Accuracy of noncycloplegic autorefraction in school-age children in China. Optom Vis Sci. 2004;81(1):49–55. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Sun YY, Wei SF, Li SM, Hu JP, Yang XH, Cao K, Lin CX, Du JL, Guo JY, Li H, et al. Cycloplegic refraction by 1% cyclopentolate in young adults: is it the gold standard? The Anyang University Students Eye Study (AUSES). Br J Ophthalmol. 2019;103(5):654–658. [ DOI ] [ PubMed ] 9. Han X, Liu C, Chen Y, He M. Myopia prediction: a systematic review. Eye (Lond). 2022;36(5):921–9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Tong HJ, Huang ZM, Li YL, Chen YM, Tian B, Ding LL, Zhu LL. Machine learning to analyze the factors influencing myopia in students of different school periods. Front Public Health. 2023;11:1169128. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Beam AL, Kohane IS. Big Data and Machine Learning in Health Care. JAMA. 2018;319(13):1317–8. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Ying B, Chandra RS, Wang J, Cui H, Oatts JT. Machine Learning Models for Predicting Cycloplegic Refractive Error and Myopia Status Based on Non-Cycloplegic Data in Chinese Students. Transl Vis Sci Technol. 2024;13(8):16. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Du B, Wang Q, Luo Y, Jin N, Rong H, Wang X, Nian H, Guo L, Liang M, Wei R. Prediction of spherical equivalent difference before and after cycloplegia in school-age children with machine learning algorithms. Front Public Health. 2023;11:1096330. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Gelfand AE, Dey DK. Bayesian Model Choice: Asymptotics and Exact Calculations. J Royal Stat Soc Ser B: Stat Methodo logy. 1994;56(3):501–14. [ Google Scholar ] 15. Xu H, Bechle MJ, Wang M, Szpiro AA, Vedal S, Bai Y, Marshall JD. National PM(2.5) and NO(2) exposure models for China based on land use regression, satellite measurements, and universal kriging. Sci Total Environ. 2019;655:423–33. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Magome K, Morishige N, Ueno A, Matsui TA, Uchio E. Prediction of cycloplegic refraction for noninvasive screening of children for refractive error. PLoS One. 2021;16(3):e248494. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Sankaridurg P, He X, Naduvilath T, Lv M, Ho A, Smith ER, Erickson P, Zhu J, Zou H, Xu X. Comparison of noncycloplegic and cycloplegic autorefraction in categorizing refractive error data in children. Acta Ophthalmol. 2017;95(7):e633–40. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Wang J, Wang X, Gao HM, Zhang H, Yang Y, Gu F, Zheng X, Gu L, Huang J, Meng J, et al. Prediction for Cycloplegic Refractive Error in Chinese School Students: Model Development and Validation. Transl Vis Sci Technol. 2022;11(1):15. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Pinkus A. Approximation theory of the MLP model in neural networks. Acta Numer. 1999;8:143–95. [ Google Scholar ] 20. Breiman L. Random Forests. Mach Learn. 2001;45(1):5–32. [ Google Scholar ] 21. Morgan IG, Ohno-Matsui K, Saw SM. Myopia. Lancet. 2012;379(9827):1739–48. [ DOI ] [ PubMed ] [ Google Scholar ] 22. Zhao E, Wang X, Zhang H, Zhao E, Wang J, Yang Y, Gu F, Gu L, Huang J, Zhang R, et al. Ocular biometrics and uncorrected visual acuity for detecting myopia in Chinese school students. Sci Rep. 2022;12(1):18644. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Anderson HA, Glasser A, Manny RE, Stuebing KK. Age-related changes in accommodative dynamics from preschool to adulthood. Invest Ophth Vis Sci. 2010;51(1):614–22. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Frénay BT, Verleysen M. Classification in the presence of label noise: a survey. IEEE T Neur Net Lear. 2014;25(5):845–69. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Tideman JWL, Polling JR, Vingerling JR, Jaddoe VWV, Williams C, Guggenheim JA, Klaver CCW. Axial length growth and the risk of developing myopia in European children. Acta Ophthalmol. 2018;96(3):301–9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (277.7KB, docx) Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. Articles from BMC Medical Informatics and Decision Making are provided here courtesy of BMC ACTIONS View on publisher site PDF (1.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• True Beauty#88560
• Frequency of Patho-morphological Classes of Gall Stones in Lahore, Pakistan#842212
• A Novel Arabic-Language Questionnaire to Assess Patient Satisfaction After Refractive Surgery#29738
• Socio-aesthetic Intervention on Body Image and Quality of Life in Women With Breast Cancer.#1059504
• Machine Learning for Reclassification of Obesity#979744