Supervised Machine Learning Models for Ocular Sagittal Height Prediction Incorporating Corneoscleral Profile Data
Timoteo González-Cruces1, Miriam Carrillo-Pulido, Francisco Javier Aguilar-Salazar
1Department of Anterior Segment (T.G.-C., F.J.A.-S., A.C.-O.), Cornea and Refractive Surgery, Hospital Arruzafa, Cordoba, Spain; Department of Optics (M.C.-P., R.I.G., S.O.-P.), Faculty of Sciences, University of Granada, Spain; Department of Health and Biomedical Sciences (A.C.-O.), Universidad Loyola, Andalucía, Spain; and Faculty of Biomedical Sciences and Sports (A.C.-O.), European University of Andalucía, Málaga, Spain.
Supervised machine learning (ML) models can predict ocular sagittal height (OC-SAG) using anterior eye topography data. Adding tomography data did not improve predictions, which are clinically acceptable for soft contact lens fitting.
Area of Science:
- Ophthalmology
- Biomedical Engineering
- Data Science
Background:
- Accurate prediction of ocular sagittal height (OC-SAG) is crucial for various ophthalmic applications, including contact lens fitting.
- Traditional methods may lack precision, necessitating advanced predictive tools.
Purpose of the Study:
- To develop and evaluate supervised machine learning (ML) models for predicting OC-SAG.
- To assess the utility of anterior eye topography and tomography data in these predictions.
Main Methods:
- Retrospective analysis of 100 eyes using anterior segment optical coherence tomography (CASIA 2).
- Development of four supervised ML models (including Random Forest) to predict OC-SAG at 10 mm and 14 mm chord lengths.
- Comparison of models using topographical data alone versus combined topographical and tomographical data.
Main Results:
- The Random Forest model demonstrated the highest predictive accuracy for OC-SAG at both 10 mm (r=0.88) and 14 mm (r=0.77) chords using topography data.
- Inclusion of tomographical data, such as corneoscleral junction (CSJ) metrics, did not significantly enhance model performance.
- Mean absolute errors were clinically acceptable for soft contact lens fitting, especially at shorter chord lengths.
Conclusions:
- Supervised ML models, particularly Random Forest, effectively predict OC-SAG using anterior eye topography.
- Tomographical data did not provide significant additional predictive value over topographical data alone.
- These ML models offer a promising, accessible method for clinicians to estimate OC-SAG and aid in decision-making for contact lens fitting.
