Development and Clinical Validation of Lightweight, Multimodal Machine Learning Models for Smartphone-Based Cataract...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Cataract is the primary cause of global blindness, impacting over 100 million individuals.
• Limited access to ophthalmologists in low- and middle-income countries hinders timely diagnosis and treatment.
• Accessible screening tools are crucial for early detection and intervention.

Purpose of the Study:

• To develop and evaluate lightweight, multimodal machine learning models for smartphone-based cataract classification.
• To enable accessible and immediate cataract screening in resource-limited environments.
• To assess the on-device feasibility and robustness of the developed AI model.

Main Methods:

• Trained and evaluated early and late fusion multimodal models using 6,794 anterior segment images and clinical data from 2,956 patients.
• Classified lens status into clear, immature cataract, mature cataract, or pseudophakia.
• Prospectively validated the model on-device in 210 patients at Aravind Eye Hospital.

Main Results:

• The early fusion model achieved a superior Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.98.
• On-device evaluation demonstrated robust performance with an AUROC of 0.96.
• Model interpretation confirmed alignment with ophthalmologist diagnostic reasoning.

Conclusions:

• The study presents the first prospectively validated, on-device, multimodal AI model for cataract classification.
• This technology facilitates instant, offline cataract detection and referral, particularly in underserved populations.
• The model empowers minimally trained personnel to screen patients, potentially broadening access to eye care and earlier treatment.