Diagnostic tool AMIE: AI beats family doctors in clinical study

The medical history between doctor and patient sets the tone for the correct diagnosis, treatment, and trust in the therapist. The AI system AMIE, developed by Google, has now performed impressively well in a study.

In a study, an AI system made more accurate diagnoses and also performed better than primary care physicians in terms of empathetic conversation skills. However, the conversation was a chat and therefore not a realistic scenario for a primary care practice. / © Adobe Stock/chinnarach

Artificial intelligence (AI) systems are becoming increasingly more powerful and reliable and could improve the accessibility and quality of care in the future. AMIE (Articulate Medical Intelligence Explorer) is one such AI system, which researchers led by Dr. Tao Tu, Dr. Mike Schaekermann, and Dr. Anil Palepu from Google Research in Mountain View, USA, are now presenting in the journal "Nature." The AI system has been optimized for diagnostic dialogues in a clinical context to enable natural, empathetic, and medically sound interaction between AI and patients, particularly in text-based telemedicine scenarios.

AMIE is based on a large-scale language model (LLM) fine-tuned using multiple data sources, including the MedQA (Medical Question-Answering) dataset, which contains multiple-choice questions used for the United States Medical Licensing Examinations (USMLE), long-form physician question-answering systems that provide detailed explanations and reflect the complexity of real-world clinical scenarios, summaries of clinical notes, and transcribed real-world medical conversations.

A special feature of AMIE is the so-called two-stage self-play process, in which the AI assumes the role of both doctor and patient in simulated dialogues. These dialogues are evaluated by an additional so-called critic AI, which provides feedback on diagnostic quality, empathy, and professionalism. This creates an iterative self-improvement system.

To increase the relevance of clinical decisions, AMIE uses a so-called chain-of-reasoning strategy during the dialogue, in which symptoms, differential diagnoses and necessary additional information are analyzed in several steps before an answer is formulated.