How We Built Healthcare AI You Can Trust: The Science Behind Amazon Connect Health

This article details the scientific approach behind Amazon Connect Health, AWS's agentic AI solution for healthcare that automates clinical documentation, medical coding, and patient engagement while maintaining clinical safety and trust.

• Pre-visit insights: Multi-step agentic pipeline summarizes patient history with 99%+ source accuracy
• Ambient documentation: Fine-tuned smaller LLM outperforms larger models by 2.2x speed, 50% fewer errors
• Medical coding: Joint ICD/CPT training achieves 0.89 F1 (ICD) and 0.90 F1 (CPT) on real-world data
• Reinforcement learning with sequential stages outperforms single-stage optimization and SFT alone
• Patient engagement agents: 92.3% authentication success, 88.5% insurance verification, 78.2% appointment scheduling
• All outputs include evidence mapping, confidence scores, and comprehensive guardrails for safety
• Clinical specialists validate each model release against real-world workflows and guidelines

Amazon Connect Health combines domain-specific fine-tuning, reinforcement learning, and rigorous clinical evaluation to deliver trustworthy AI that reduces administrative burden while maintaining patient safety and clinical accuracy.