Personalized AI System for Maternal Nutrition and Exercise

Abstract

This work presents a personalized AI system for maternal wellness that generates two coordinated, daily outputs for pregnant users: culturally adapted meal plans and risk-aware exercise prescriptions. In its nutrition module, the combination of a library of curated local dishes and trimester-specific energy corridors with condition-aware rules (diabetes/gestational diabetes, anemia, hypertension, allergies) highlights iron, folate, protein, and calcium instead of full nutrient panels. The exercise module comprises a rules-first risk screen (Low/High) corroborated by a calibrated gradient-boosted classifier, followed by a template-based planner that renders day-level sessions with safety guardrails, trimester adaptations, and video guidance. Evaluation on de-identified datasets (2,000 meal profiles; 970 risk records; 1,934 exercise prescriptions) showed high nutritional adequacy and reliable safety enforcement: daily energy corridors were met on 94.7% of days and ≥3/4 essential nutrients on 91.3%, with 0 allergen violations. The risk classifier achieved AUROC 0.86, accuracy 0.82, sensitivity 0.79, and specificity 0.84. The exercise recommender successfully enforced trimester- and risk-based guardrails, with all plans passing safety checks. These findings, from a study focused on Sri Lanka, suggest that an integrated, objective and culturally based method can provide personalized, safe guidance daily in resource-constrained environments, and serve as a basis for future clinical validation and predictive modeling.