Enhancing Early Keratoconus Detection With Multimodal Machine Learning: Integrating Tomography, Biomechanics, and...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Early detection of keratoconus (KC) is crucial for preventing vision loss.
• Current diagnostic methods may not sufficiently identify early or subclinical stages of KC.
• Integrating diverse data sources can potentially improve diagnostic accuracy.

Purpose of the Study:

• To develop and validate a machine learning (ML) system for enhanced early keratoconus (KC) detection.
• To integrate Scheimpflug tomography, corneal biomechanics, and clinical risk factors (CRF) into a diagnostic model.
• To assess the ML system's performance in identifying forme fruste KC (FFKC), subclinical KC, and clinical KC.

Main Methods:

• A prospective, multicenter, cross-sectional study involving patients with KC and refractive surgery candidates.
• Collection of demographic, lifestyle, clinical ophthalmic, Pentacam, and Corvis ST data from 1,035 eyes.
• Development and comparison of six ML models using various feature sets (CRF, device parameters, combined, selected features), evaluated by Area Under the Receiver Operating Characteristic Curve (AUROC).

Main Results:

• The CatBoost ML model, utilizing selected features, achieved the highest AUROC of 0.975 for FFKC detection in the test set.
• This multimodal ML system significantly outperformed CRF-only and device-only models.
• Near-perfect external validation performance was observed for subclinical (AUROC=0.991) and clinical KC (AUROC=1.000).

Conclusions:

• A multimodal ML system integrating clinical risk factors, tomography, and biomechanics significantly improves early KC detection, especially for FFKC.
• This AI-driven approach shows potential to enhance clinical decision-making and screening protocols for refractive surgery candidates.
• The validated ML model offers a promising tool for more accurate and earlier diagnosis of keratoconus.