Google's new AI chatbot AMIE to converse with patients, make diagnoses like doctors

Tech giant Google has unveiled a revolutionary chatbot named the Articulate Medical Intelligence Explorer (AMIE), designed to engage in diagnostic conversations with patients, demonstrating a level of reasoning comparable to human doctors.

In a blog post, Google Research's Alan Karthikesalingam and Vivek Natarajan, who led the research, highlighted the significance of physician-patient communication in medicine and the potential of AI systems to enhance accessibility and quality of care. However, they acknowledged the considerable challenge of approximating clinicians' expertise.

AMIE was trained on real-world datasets, including medical reasoning, summarisation, and clinical conversations. The team created a unique self-play-based simulated diagnostic dialogue environment to train the chatbot, incorporating automated feedback mechanisms in a virtual care setting. An inference time chain-of-reasoning strategy was also employed to enhance AMIE's diagnostic accuracy and conversation quality.

The chatbot's performance was assessed through consultations with simulated patients, played by trained actors, and compared to those conducted by 20 board-certified primary care physicians (PCPs). The study, involving 149 case scenarios from Objective Structured Clinical Examination (OSCE) providers in Canada, the UK, and India, revealed promising results.

AMIE demonstrated diagnostic conversations at least as effective as PCPs across multiple clinically meaningful aspects of consultation quality, including history-taking, diagnostic accuracy, clinical management, communication skills, relationship-building, and empathy. The chatbot exhibited greater diagnostic accuracy and superior performance in various pointers, showcasing its potential in the field of conversational diagnostic AI.

Despite the promising outcomes, the researchers emphasised the study's limitations, particularly the use of unfamiliar synchronous text chat, calling for further research before AMIE can be implemented in real-world settings.