GlycoSmart: AI-Driven Nutritional Information Retrieval and Diabetic Health Evaluator

Abstract

Diabetes management demands precise dietary choices, but interpreting nutritional labels poses a major hurdle. GlycoSmart, an AI-powered mobile platform, ad dresses this by automating label detection, extraction, and analysis to offer real-time personalized diabetic health advice. It employs YOLOv8 for efficient object detec tion, attaining an F1 score of 0.97 at a 0.701 confidence threshold and mAP@0.5 of 0.986 for reliable label identification across varied packaging. The Gemini 1.5 Flash OCR model processes extracted data for accurate text recognition in complex formats. The nutritional analysis integrates XGBoost and Random Forest algorithms to predict Glycemic Index (GI) and compute Glycemic Load (GL), with Random Forest outperforming at an R� of 0.9217 and reduced MAE, selected for GI prediction. Built on a Flask backend with Supabase storage and Nutritionix API integration for additional data, the React Native with Expo frontend ensures an intuitive interface. The fine-tuned Diabetica-7B LLM generates context-specific dietary recommendations based on user profiles. By merging object detection, OCR, and predictive modeling, the system provides accurate insights and empowers informed choices for diabetics. The prototype excels in performance and usability but lacks offline mode, AI meal suggestions, and barcode scanning. Future enhancements will boost accessibility and real-world efficacy in diabetes self-management.