Myopia Prediction Using Machine Learning: An External Validation Study

Rajat S. Chandra, Bole Ying, Jianyong Wang, Hongguang Cui, Gui‐Shuang Ying, Julius T. OattsVision 2025 · DOI

TL;DR

This study validates previously developed machine learning (ML) models (XGBoost for refractive error, Random Forest for myopia status) for predicting myopia using non-cycloplegic data. The models demonstrated robust performance even with variations in cycloplegic agents and ocular biometers, suggesting their potential for real-world clinical and epidemiological applications.

What Problem Does This Paper Address?

Accurately measuring refractive error, especially for myopia, typically requires cycloplegia (dilating eye drops) to relax the eye's focusing muscle. However, administering cycloplegia is time-consuming, resource-intensive, and can be challenging in large-scale epidemiological studies or clinical settings, potentially leading to underestimation of myopia prevalence and severity. Previous research has shown that ML models can predict cycloplegic refractive error using non-cycloplegic data, but these models were validated under standardized conditions. This study addresses the critical question of whether these ML models maintain their predictive power when applied in a more diverse, real-world setting with variations in the agents used for cycloplegia and the devices used for ocular biometry.

Methods

The researchers employed two machine learning models that had previously shown strong performance: XGBoost for predicting cycloplegic spherical equivalent refraction (SER) and Random Forest for predicting myopia status. These models were trained on non-cycloplegic data (demographics, non-cycloplegic SER, ocular biometrics like axial length, corneal curvature, etc.) and then externally validated on a new cohort. The performance of the models was evaluated using standard metrics: R-squared and Mean Absolute Error (MAE) for SER prediction, and Area Under the ROC Curve (AUC), sensitivity, and specificity for myopia prediction. Feature importance analysis was also conducted to understand which non-cycloplegic factors were most influential in the predictions.

Datasets

The study utilized data from 614 Chinese students aged 8-13 years. This cohort was distinct from the original training and validation datasets. Crucially, this external validation dataset incorporated variations in:

  • Cycloplegic Agents: 0.5% tropicamide (used in 82.2% of students) and 1% cyclopentolate (used in 17.8%).
  • Ocular Biometers: IOLMaster 700 (used in 33.7% of students) and Optical Biometer SW-9000 (used in 66.3%).

This heterogeneity in data collection protocols is a key aspect of the study's external validation design.

Key Insights and Ideas

  • Robustness Across Heterogeneity: The most significant finding is that ML models can maintain high predictive accuracy even when the data comes from different cycloplegic agents and biometry devices. This is crucial for real-world applicability, as standardized protocols are not always feasible.
  • Non-Cycloplegic Data is Powerful: The study reinforces that readily available non-cycloplegic measurements (like non-cycloplegic SER, axial length, and visual acuity) are strong predictors of cycloplegic refractive error and myopia status.
  • Addressing the "Gold Standard" Challenge: The research offers a practical solution for situations where obtaining the "gold standard" cycloplegic refraction is difficult, providing a reliable alternative for myopia assessment.
  • Feature Importance Consistency: Key predictive features identified in previous studies (non-cycloplegic SER, axial length, etc.) remained important in this diverse cohort, highlighting their fundamental role in myopia prediction.

Key Results

  • XGBoost for SER Prediction: Achieved an R-squared of 0.95 and a Mean Absolute Error (MAE) of 0.32 D, indicating excellent prediction of cycloplegic spherical equivalent refraction.
  • Random Forest for Myopia Prediction: Demonstrated high performance with an AUC of 0.991, sensitivity of 93.4%, and specificity of 96.4% for per-eye analysis. For per-person analysis, sensitivity was 93.3% and specificity was 94.7%.
  • Myopia Prevalence Prediction: The models accurately predicted the observed myopia prevalence rate (62.9%), with XGBoost predicting 58.8% and Random Forest predicting 60.6%. This contrasts with non-cycloplegic SER alone, which significantly overestimated myopia prevalence (76.7%).
  • Performance Across Subgroups: Both models performed consistently well across different age groups (8-13 years), cycloplegic agents (tropicamide vs. cyclopentolate), and ocular biometers (IOLMaster 700 vs. SW-9000).

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Age Range: The study cohort was restricted to students aged 8-13 years.
  • Specific Agents/Biometers: Only two cycloplegic agents and two ocular biometers were used.
  • Geographic Scope: Participants were from a single city.

Future work should focus on:

  • Broader Validation: Expanding validation to include students from urban and rural schools, a wider range of cycloplegic agents, and different ocular biometers.
  • Diverse Populations: Testing these models in other populations to assess generalizability.
  • Longitudinal Studies: Investigating the models' ability to predict myopia progression over time.

Summary updated

Abstract

We previously developed machine learning (ML) models for predicting cycloplegic spherical equivalent refraction (SER) and myopia using non-cycloplegic data and following a standardized protocol (cycloplegia with 0.5% tropicamide and biometry using NIDEK A-scan), but the models' performance may not be generalizable to other settings. This study evaluated the performance of ML models in an independent cohort using a different cycloplegic agent and biometer. Chinese students (N = 614) aged 8-13 years underwent autorefraction before and after cycloplegia with 0.5% tropicamide (<i>n</i> = 505) or 1% cyclopentolate (<i>n</i> = 109). Biometric measures were obtained using an IOLMaster 700 (<i>n</i> = 207) or Optical Biometer SW-9000 (<i>n</i> = 407). ML models were evaluated using R<sup>2</sup>, mean absolute error (MAE), sensitivity, specificity, and area under the ROC curve (AUC). The XGBoost model predicted cycloplegic SER very well (R<sup>2</sup> = 0.95, MAE (SD) = 0.32 (0.30) D). Both ML models predicted myopia well (random forest: AUC 0.99, sensitivity 93.7%, specificity 96.4%; XGBoost: sensitivity 90.1%, specificity 96.8%) and accurately predicted the myopia rate (observed 62.9%; random forest: 60.6%; XGBoost: 58.8%) despite heterogeneous cycloplegia and biometry factors. In this independent cohort of students, XGBoost and random forest performed very well for predicting cycloplegic SER and myopia status using non-cycloplegic data. This external validation study demonstrated that ML may provide a useful tool for estimating cycloplegic SER and myopia prevalence with heterogeneous clinical parameters, and study in additional populations is warranted.
BibTeX
@inproceedings{chandra2025myopia,
  author = {Rajat S. Chandra and Bole Ying and Jianyong Wang and Hongguang Cui and Gui‐Shuang Ying and Julius T. Oatts},
  title = {{Myopia Prediction Using Machine Learning: An External Validation Study}},
  booktitle = {Vision},
  year = {2025},
  doi = {10.3390/vision9040084},
}
Figure