Google Research Research Leads Alan Karthikesalingam and Vivek Natarajan highlights the importance of physician-patient conversations in medicine. They indicate the potential for AI to improve diagnostic conversations and increase care availability, accessibility, quality, and consistency. However, approximating clinicians’ expertise is a significant challenge. They created Articulate Medical Intelligence Explorer (AMIE), an AI system designed to enhance diagnostic conversations while balancing empathy and fostering relationships.
Development of AMIE
To train AMIE, the team used real-world datasets of medical reasoning and conversations. They created a simulated learning environment to scale AMIE’s knowledge across various medical conditions and contexts and fine-tuned it using an evolving set of simulated dialogues. AMIE then participated in a continuous learning cycle. The team also implemented an inference time chain-of-reasoning strategy to refine AMIE’s response.
Performance and Evaluation
AMIE underwent performance evaluations and simulated diagnostic conversations with trained actors. In a randomized study, AMIE performed at least as well as primary care physicians (PCPs) on several axes of consultation quality. It had greater diagnostic accuracy and outperformed PCPs on multiple evaluation axes.
Limitations
The team stresses that their research is the first step in a long journey. They note that their evaluation method did not represent real-world clinical practice and that transitioning AMIE to a reliable and safe tool will require additional research to ensure safety and reliability.
Overall, Google Research’s AMIE shows promising potential to improve clinical dialogue and raise diagnostic accuracy through AI technology.