PedRAG: Retrieval-Augmented Generation for Pediatric Medical Question-Answering

RAG framework for pediatric medical QA combining dense encoders and sparse retrieval (BM25) with cross-encoder reranking, improving answer accuracy by 34% and reducing hallucination rates by 42% over existing baselines.

Built multi-class age-group segmentation models in PyTorch achieving 94% accuracy across four pediatric cohorts (0-2, 3-5, 6-12, 13-18) to enable targeted, age-appropriate medical QA.

Work conducted as Graduate Research Assistant with Prof. Cornelia Paulik at UC Berkeley.

Manuscript in preparation. Target venue: ICML 2026 (Poster).

Recommended citation: Vankadaru, V. et al. (2026). PedRAG: Retrieval-Augmented Generation for Pediatric Medical QA. Target: ICML 2026.