A Multimodal Multitask Artificial Intelligence Model for Orthokeratology Contact Lens Fitting: An Integrated Framework to Enhance Lens Centration and Myopia Control Effect.

PURPOSE: To develop a multimodal, multitask artificial intelligence (AI) model to classify postfitting corneal topography patterns and predict axial length (AL) growth, while simultaneously outputting optimal orthokeratology (ortho-K) lens parameters and the predicted probability of axial growth to support clinical decisions.

DESIGN

A retrospective analysis.

SUBJECTS: Clinical data and corneal topography maps from 3529 myopic eyes fitted with ortho-K lenses (Euclid or Alpha designs) from 2018 to 2023 were collected.

METHODS

A novel AI model was built comprising (1) a ResNet50-based "recall" model to predict key lens parameters and (2) a multimodal multitask "ranking" model (comparing ResNet50, ViT-B/16, and CLIP-ViT architectures) to classify postlens corneal topography and predict axial elongation. Postlens topography pattern was classified by centration and plus power ring pattern. Annual axial elongation rate ≥0.3 mm/y is defined as fast myopia progression. The recall model output gave a list of optimal lens parameter candidates, and the ranking model output the likely topography classification and the probability of AL growth for each candidate at 1 year.

MAIN OUTCOME MEASURES: Mean squared error, recall rate, and accuracy were measured to evaluate model performance.

RESULTS

Of the 3529 ortho-K fits, 2643 eyes (74.89%) wore Euclid lenses and 886 (25.11%) wore Alpha lenses. All subjects wore their ortho-K lenses over the study period without serious complications. The average spherical equivalent refraction was -2.91 ± 1.14 diopters, and mean lens wear period was 11 months (range 9-13 months). Uncorrected visual acuity ≥0.8 was achieved in 96.43% of eyes at 1 month. The overall annual axial elongation rate observed was 0.24 ± 0.20 mm/y. Class 1 axial growth (<0.3 mm/y) in 62.79% of eyes, and class 1 postlens topography was achieved in 80.0% of eyes at 1-year follow-up. All 3 ranking models performed comparably: topography classification accuracies were 0.95, 0.96, and 0.96 for ResNet50, ViT, and CLIP-ViT models, respectively; while axial growth prediction accuracies were 0.880, 0.882, and 0.883, respectively.

CONCLUSIONS

The proposed multimodal multitask AI model performed well in classifying corneal topography patterns, predicting AL growth rate, and recommending lens parameters, offering valuable decision support in ortho-K lens fitting.

FINANCIAL DISCLOSURES: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.