Publication:

A multimodal artificial intelligence framework for the early detection of diet-related maculopathy using trophic biomarkers and fundus imaging

Background Diet-related maculopathy (DRM) is a progressive retinal condition associated with chronic nutritional deficiencies, oxidative stress, and impaired macular metabolism. Early detection is challenging because biochemical deterioration precedes clinically visible retinal changes. This study aimed to develop and validate a multimodal artificial intelligence (AI) framework integrating fundus imaging with trophic biomarkers to enhance early DRM detection.

Methods A prospective diagnostic model validation study was conducted among 580 adults aged 30-65 years at a Asim eye care Center in Ghaziabad, Delhi-NCR between January 2023 and March 2025. Fundus images were analysed using a fine-tuned ResNet-50 convolutional neural network, whereas plasma concentrations of lutein, zeaxanthin, vitamins A, C, and E, zinc, and docosahexaenoic acid (DHA) were processed using a Random Forest classifier. A fusion architecture integrated both outputs. Model performance was assessed through accuracy, sensitivity, specificity, and area under the ROC curve. Longitudinal follow-up assessed predictive lead time.

Results The multimodal AI model achieved an accuracy of 95.2%, sensitivity of 93.1%, specificity of 96.4%, and an AUC of 0.972. Lutein, zeaxanthin, and DHA were the most significant biochemical predictors, whereas macular reflectance patterns and early drusen signatures were the strongest image-derived features. The model detected DRM on average 11.2 months before clinical diagnosis. Nutritional insufficiencies were present in 42.1% of participants.

Conclusions The multimodal AI framework demonstrated excellent diagnostic capability and substantial predictive lead time, enabling early identification of DRM and supporting personalized nutritional intervention. This integrative approach may improve preventive retinal care and reduce long-term visual impairment.