Machine Learning-Based Prediction of Cycloplegic Refraction Using the Eyerobo Vision Screener: Design—A Cross-Sectional Comparative Device Study

A cross-sectional comparative device study to compare the Eyerobo Vision Screener (VS), a portable handheld photorefractor, against a conventional auto-refractor for machine-learning-based prediction of cycloplegic spherical equivalent (SE) and power vector components (J0, J45) in a pediatric myopia cohort, to evaluate whether supplementary IOL Master biometry narrows the performance gap between devices, and to report screening-oriented classification metrics (sensitivity, specificity, positive and negative predictive values, area under the receiver operating characteristic curve [AUC]) at clinically meaningful referral thresholds that quantify the device–model combination’s triage performance. Data from 1129 eyes of 574 patients aged 6–18 years were collected using three ophthalmic devices before and after pharmacological dilation. All refractive measurements were decomposed into power vectors (SE, J0, J45) following Thibos et al. Twelve clinically motivated feature scenarios were constructed from a symmetric framework comparing the Eyerobo VS and auto-refractor (each under predilation, post-dilation, and combined conditions), with and without IOL Master biometry. TabPFN, a tabular foundation model requiring no hyperparameter tuning, was evaluated alongside seven conventional machine-learning algorithms using patient-level five-fold grouped cross-validation repeated over three random seeds. Agreement was assessed using Bland–Altman analysis, Pearson correlation, and clinical threshold analysis. In addition to regression metrics, a screening-classification analysis was performed at three prespecified clinical thresholds (cycloplegic SE $$\le -0.50$$ D, $$\le -3.00$$ D, and $$\le -6.00$$ D), reporting sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and AUC with 95% paired bootstrap confidence intervals. Formal paired statistical comparison between TabPFN and the next-best algorithm and a one-eye-per-patient sensitivity analysis are also reported. TabPFN achieved the lowest mean absolute error (MAE) on all Eyerobo scenarios and was competitive with Ridge on auto-refractor scenarios; on the headline Eyerobo Pre + IOL scenario, the per-eye paired difference in MAE between TabPFN (0.425 D) and the next-best gradient-boosting learner (XGBoost, 0.542 D) was 0.117 D in favour of TabPFN, with paired Wilcoxon signed-rank $$p<0.001$$ . For predilation SE prediction, the auto-refractor achieved an MAE of 0.295 D (87.4% within ± 0.50 D), compared with 0.657 D (61.6%) for the Eyerobo VS. Adding IOL Master biometry to the Eyerobo reduced this gap by 64%, yielding an MAE of 0.425 D (72.0%). For astigmatic power vector components, the gap narrowed further: J0 MAE was 0.106 D (auto-refractor) versus 0.143 D (Eyerobo + IOL), and J45 MAE was 0.060 D versus 0.088 D. Auto-refractor post-dilation predictions approached cycloplegic accuracy (MAE = 0.203 D, 96.0% within ± 0.50 D). Subgroup analysis showed the Eyerobo performed best for mild myopia (MAE = 0.398 D, 75.6% within ± 0.50 D) and degraded for high myopia and the small nonmyopic stratum. In the screening-classification analysis, the Eyerobo and the auto-refractor were operationally equivalent at the any-myopia threshold (Eyerobo AUC 0.964 [95% CI 0.939–0.982], sensitivity 0.946, specificity 0.894; auto-refractor AUC 0.969 [0.948–0.986], sensitivity 0.960, specificity 0.886; paired AUC difference 0.005 [95% CI $$-0.022$$ to $$+0.034$$ ], formally noninferior under a prespecified $$\delta =0.05$$ margin). Within this single-center pediatric referral cohort, the Eyerobo VS combined with machine learning produced cycloplegic SE estimates of a precision consistent with use as a triage adjunct to identify children who should proceed to a full cycloplegic examination, with near-equivalent performance to the auto-refractor for astigmatic components. The addition of IOL Master biometry narrowed the device gap by 64%. TabPFN achieved the best performance among the evaluated algorithms within this cohort and requires no hyperparameter tuning, but the modest margin over linear and gradient-boosting baselines and the absence of external validation preclude any recommendation as the definitive algorithm of choice. The proposed approach is not a substitute for cycloplegic refraction; prospective external validation in a general pediatric screening cohort is required before clinical deployment.