Answers are grounded in this paper and cite their sources.
Introduction Personalized optical simulation and machine learning provide a computational approach to estimating spectacle prescriptions for pediatric astigmatism. Methods This single-center, prospective, within-subject agreement study with randomized examination order evaluated prescriptions for 77 right eyes at a post-cycloplegia follow-up visit. The model used structured numerical wavefront exports and patient information for individualized optical reconstruction, virtual lens simulation, and clinical prescription residual calibration, and was developed using a separate dataset of 5,630 eyes. Artificial intelligence (AI) estimates and conventional autorefraction were compared with subjective refraction performed by an experienced optometrist masked to AI outputs. Children and the trial-lens outcome assessor were masked to prescription source, and lens presentation order was randomized. Agreement was evaluated for spherical equivalent (SE), J0, and J45. Inferential analyses were exploratory, without adjustment for multiple comparisons. Results SE mean bias ± standard deviation of paired differences was 0.01 ± 0.48 D for AI and 0.22 ± 0.57 D for conventional autorefraction; conventional Bland-Altman limits of agreement were −0.94 to 0.96 D and −0.89 to 1.33 D, respectively. The between-method difference in signed bias was −0.21 D (95% confidence interval: −0.35 to −0.07; P = 0.004). SE mean absolute errors were 0.30 D and 0.39 D, with a paired difference of −0.09 D (95% bootstrap confidence interval: −0.23 to 0.04). Astigmatic error analyses did not establish an overall advantage for AI. In the immediate comparison, trial-lens logMAR visual acuity was lower with AI, and 66.2% of participants preferred the AI prescription. Discussion A smaller SE mean bias does not establish greater individual accuracy, clinical interchangeability, or the effectiveness of AI-assisted clinical decision-making. Long-term benefits and safety during spectacle wear require further evaluation.
No takes yet. Share an insight, caveat, or question.
甄毅 et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: