NEJM AI Grand Rounds Google’s Efforts to Build Patient-Facing AI: A Conversation with Drs. Alan Karthikesalingam and Anil Palepu
45 snips
Jun 18, 2025 
Dr. Alan Karthikesalingam and Dr. Anil Palepu from Google delve into the creation of AMIE, an AI designed for patient interactions. They discuss how AMIE was trained with synthetic conversations, proving to be more effective than real transcripts. The duo emphasizes the importance of empathy in AI and the challenges of incorporating machine learning in clinical workflows. They also touch on how AI is evolving in healthcare and its potential to transform patient care by providing quicker and more engaging interactions.
AI Snips
Chapters
Transcript
Episode notes
Synthetic Conversations for Training AMIE
- The AMIE system uses synthetic doctor-patient conversations generated by multi-agent LLMs to train AI for clinical history taking.
- This approach overcomes noisy, limited real-world transcripts by enabling tailored, diverse training reflecting the long tail of medical presentations.
Self-Play and Grounding Boost AMIE
- AMIE's training involved a self-play system with critic and moderator roles improving dialogue quality.
- Grounding case conditions via search during training enhanced the model's diagnostic conversation capabilities beyond naive prompting.
Synthetic Data Outperforms Human Transcripts
- Training on noisy human transcripts alone was insufficient for modeling clinical conversations.
- Synthetic data allowed targeted coverage of clinical conditions unlike incomplete transcripts from audio sources.