How can practitioners use AI to assess OrthoK efficacy?
January 15, 2026
By Ashley Wallace-Tucker, OD, FAAO, FSLS
Photo credit: Dreamstime Photos
Myopia progression in children is driven primarily by axial length (AL) elongation, which is the most critical biomarker for long-term visual prognosis and risk of pathology. Orthokeratology lenses are widely recognized to slow myopia progression by reshaping the cornea overnight, but the efficacy of OrthoK varies substantially between individuals. This variation reflects both demographic and ocular factors, as well as the corneal reshaping process itself. Prior studies have linked both corneal topographic changes and baseline ocular measures with AL outcomes, yet quantifying those effects objectively and predicting progression remains a key clinical challenge.
To address this, Rong et al. developed an integrated artificial intelligence (AI) framework combining automated corneal topography analysis, causal inference and predictive modeling to forecast future axial elongation in myopic children undergoing OrthoK. This approach aims to give clinicians an interpretable, evidence-based tool for individualized myopia management and OrthoK efficacy assessment.
Study Design & Methods
- Population: 143 myopic children (276 eyes) who had worn OrthoK lenses for >12 months.
- Automated Feature Extraction: Using advanced digital image processing, researchers automatically quantified three local corneal topographic (CT) features from tangential topography maps after OrthoK wear:
- Treatment Zone Area (TZA): the surface area of the central corneal flattening zone,
- Eccentric Distance (ED): the spatial offset between the pupil center and the treatment zone center,
- Eccentric Angle (EA): the angular orientation of the treatment zone decentration relative to a reference axis.
This automated analysis was validated against manual expert annotation, showing high accuracy (mean absolute percentage errors 2.1% for TZA, 1.2% for ED and 0.7% for EA).
- Causal Inference: Counterfactual inference methods were applied to assess causal effects of these CT features on changes in axial length.
- Predictive Modeling: A CatBoost machine learning model (gradient-boosted decision trees) was trained using baseline characteristics (age and initial AL) and one-month CT metrics to predict axial elongation at six months and one year.
Key Findings
- Automated CT Feature Performance:
The algorithm’s measurements of TZA, ED and EA were highly consistent with manual annotations demonstrating that automated topographic quantification is feasible and reliable for large clinical datasets. - Causal Effects on AL Change:
- Per unit increase in TZA, there was an associated increase in AL of about 0.054 mm (suggesting that larger treatment zones, all else equal, correlated with slightly more elongation).
- Eccentric distance (ED) showed negative association (larger decentration was linked with reduced AL elongation, echoing clinical observations that mild decentration can enhance myopia control).
- Eccentric angle (EA) had negligible effect on AL change in this cohort.
- Prediction Model Accuracy:
The CatBoost model, using just age and baseline AL, predicted six-month and one-year AL with mean absolute errors of ~0.18 mm and ~0.17 mm, respectively. Performance was further improved with inclusion of one-month topographic features.
Clinical & Research Significance
This work is notable for integrating perceptual (image processing), explanatory (causal inference) and prognostic (predictive modeling) components within a single framework – an advance over prior models relying solely on linear regression or demographic predictors. Automated CT extraction reduces observer bias and manual workload, while causal analysis helps interpret how morphological changes in OrthoK lenses relate to growth modulation. The predictive model’s high accuracy in axial forecasting represents clinically meaningful precision, offering potential utility in early identification of poor responders, personalized OrthoK adjustments and enhanced patient counseling.
Overall, this AI approach exemplifies a next-generation decision support tool that may shape personalized myopia control strategies and guide future research on dynamic treatment optimization.
Abstract
Artificial Intelligence for Predicting the Axial Length Response of Orthokeratology in Myopic Children
Xin Rong, Zihui Wu, Zihang Xu, Tianhao Zhang, Guangming Xie and Liu Yang
Objective
This study aimed to automate the extraction of local corneal topography (CT) features in myopic children undergoing orthokeratology (OK), evaluate their causal effects on axial length (AL) control, and develop a predictive model for AL progression.
Approach
We retrospectively analyzed myopic children who had received OK treatment for more than 12 months. Advanced digital image processing techniques were employed to automatically quantify three critical CT parameters: treatment zone area (TZA), eccentric distance (ED), and eccentric angle (EA). Counterfactual inference quantified causal relationships between these parameters and AL changes. Baseline characteristics and one-month CT features were used to train a CatBoost prediction model.
Main results
This study included 143 myopic subjects (276 eyes) treated with OK lenses. The image processing algorithm performed comparably to manual annotation, with mean absolute percentage errors of 2.1% (TZA), 1.2% (ED), and 0.7% (EA). Per unit increase, TZA, ED, and EA were associated with AL changes of 0.054 mm, −0.161 mm, and −0.0003 mm, respectively. The CatBoost model, using initial AL and age, predicted six-month and one-year AL with absolute errors of 0.180 and 0.169 mm.
Significance
This work establishes an integrated artificial intelligence (AI) framework that combines automated CT analysis, causal inference, and predictive modeling. It provides clinicians with an interpretable tool for assessing OK efficacy and forecasting myopia progression, while paving the way for next-generation healthcare AI systems with integrated perceptual, explanatory, and prognostic capabilities.
