TriDoSHAI: Integrating Tridosha Theory with Machine Learning for Personalized Wellness

Abstract

We introduce TriDoSHAI, an end-to-end framework that integrates Ayurveda’s Tridosha theory (Vata, Pitta, Kapha) with modern machine learning for culturally grounded, ethical wellness prediction. The system combines a React/Tailwind conversational frontend with a FastAPI WebSocket backend that orchestrates intent recognition and a supervised feed-forward neural classifier. Using a curated dataset of 1,201 records across 20 encoded features, TriDoSHAI predicts one of six Prakriti labels (including mixed constitutions) based on a structured 20-question interview. We describe the data schema, feature encodings, model architectures, training protocol (80/20 split; validation split 0.2), deployment pipeline, and a comprehensive evaluation suite including accuracy, per-class metrics, normalized confusion matrices, radar plots, and error decompositions. The classifier achieves 0.8833 accuracy with a macro-F1 of 0.8835, demonstrating reliable performance across both single- and mixed-dosha classes. Importantly, the system is explicitly non-diagnostic, providing lifestyle and dietary suggestions under Ahimsa-oriented safety constraints. This work advances the vision of India’s National Education Policy [1] by operationalizing Indian Knowledge Systems in AI-enabled healthcare, while aligning with WHO recommendations on the integration of traditional medicine into modern health ecosystems [2].