Abstract
Globally, cataract remains the leading cause of blindness, affecting over 100 million people, with a disproportionate burden in low- and middle-income countries where access to ophthalmologists is limited. Although cataract surgery can restore vision almost immediately, timely diagnosis and referral remains a major barrier to care. We developed and prospectively evaluated lightweight, multimodal machine learning models capable of classifying lens status on a smartphone, enabling accessible screening in low-resource environments. We trained and evaluated both early and late fusion model architectures to classify lens status as clear, immature cataract, mature cataract, or pseudophakia using 6,794 anterior segment eye images captured using Scout™ smartphone-based diffuse illumination system paired with clinical data (age, visual acuity, and pinhole acuity) from 2,956 patients from three eye hospitals in India. The early fusion model, which jointly integrates image and clinical features via a learnable gating mechanism, achieved superior performance (AUROC=0.98) compared to late fusion. Model interpretation using feature importance and Grad-CAM revealed that early fusion effectively balanced visual and clinical parameters, mirroring ophthalmologist diagnostic reasoning. Prospective on-device evaluation in 210 patients at the Aravind Eye Hospital demonstrated equivalent performance, achieving an AUROC of 0.96, confirming robustness and real-time feasibility on mobile hardware. These results demonstrate the first prospectively validated, on-device, multimodal cataract machine learning model, demonstrating the feasibility of instant, offline cataract classification and referral in low-resource environments. This advance has potential to broaden cataract screening, allowing minimally trained workers to screen and refer patients, and enabling earlier diagnosis, referral, and treatment in underserved populations.