Efficacy of Machine Learning Models for Predicting Cycloplegic Refractive Error Based on Non-Cycloplegic Parameters in Adults With Myopia

Myopia is a highly prevalent, irreversible refractive disorder with substantial impact on quality of life. Cycloplegic refraction is the gold standard for assessing refractive error in adults considering optical or surgical correction, but it is time-consuming, slow to recover from, and frequently associated with ocular discomfort. Non-cycloplegic refraction is therefore used routinely in clinical practice, despite known differences from cycloplegic values in a subset of adult myopes.

Critically, this discrepancy varies substantially between individuals and cannot be anticipated from non-cycloplegic measurements alone. Clinicians have no reliable way to identify, prior to dilation, which patients are likely to be overcorrected if cycloplegia is omitted, potentially leading to overcorrected prescriptions, asthenopia, and myopic progression.

Machine learning approaches that capture non-linear relationships between clinical predictors and refractive outcomes have shown promise in children, but comparable models for adults remain largely unexplored, and most rely on axial length, which is unavailable in routine optometric settings. Refractive surgery centers offer a uniquely suitable data source, as every candidate undergoes standardized paired non-cycloplegic and cycloplegic refraction with detailed anterior segment biometry during routine preoperative evaluation. This study leverages such data to develop and validate models estimating cycloplegic refractive error from non-cycloplegic parameters, providing a decision-support tool that reduces unnecessary cycloplegia while flagging patients for whom dilated refraction remains indicated.