Google eyes AI-based diagnostic dialogue tool as part of medical interview system

 A team of AI researchers at Google Research and Google DeepMind has developed the rudiments of an AI-based diagnostic dialogue system to conduct medical interviews. The group has published a paper describing their research on the arXiv preprint server.

When a doctor interviews a patient to determine what medical situation needs to be addressed, the research team noted, they come armed with a deep background in medical education and practical experience. Though it might seem at times to a patient that such interviews are cursory at best, most are both efficient and result in accurate results.

But there is also room for improvement, the researchers note. One of the areas where many doctors fall short is in their bedside manner. Due to personality quirks, a tight and busy work schedule, or dealing with patients who are not always kind and polite, some medical interviewers may be seen as stiff or stilted, creating the perception that the doctor does not care much for the patient's welfare.

In this new effort, the team at Google has noted that many LLMs, such as ChatGPT, often come across as quite empathetic and more than eager to help—traits sometimes lacking in doctor interviews. That gave them the impetus to begin working on an LLM that might one day evolve into a real-world diagnostic dialogue system.

The new system is called the Articulate Medical Intelligence Explorer, or AMIE. The team at Google is quick to point out that it is still purely experimental. They also note that building their system has been both unique and difficult due to the dearth of doctor/patient medical interviews available for use as training material.

That led them to try a novel approach to teaching their system how to query a patient. First, they trained it on the limited amount of publicly available data. They then attempted to coax the system into training itself by prompting it to play the part of a person with a specific illness. Next, they asked the system to play the part of a critic who has witnessed multiple interviews that the system conducted.

The system then interviewed 20 volunteers trained to pretend to be patients. The results were rated by medical professionals to determine accuracy. The volunteers who had played the patients evaluated the system's bedside manner.

The researchers found AMIE to be as accurate in diagnosing the patients as trained doctors. They also found that it scored better with its bedside manner. The team at Google plans to improve the system's capabilities when tested in more realistic real-word conditions and to improve its bedside manner further.

More information: Tao Tu et al, Towards Conversational Diagnostic AI, arXiv (2024). DOI: 10.48550/arxiv.2401.05654

Google Research blog: blog.research.google/2024/01/a … r-diagnostic_12.html

https://medicalxpress.com/news/2024-01-google-ai-based-diagnostic-dialogue.html