AI is a useful tool for handling many tasks in private as well as professional settings, but medical advice, especially diagnosis, is one area where I generally draw the line. General-purpose chatbots tend to sound confident and convincing while giving medical advice, even when they are wrong, as studies have previously found.
That being said, Google’s new research points to a much more convincing use for AI in healthcare. Its Articulate Medical Intelligence Explorer chatbot, or AMIE for short, could help prepare both sides before the appointment even begins.
AMIE did the questioning before the doctor walked in
Researchers from Google and Beth Israel Deaconess Medical Center tested AMIE with adults seeking care for new, non-emergency health complaints. Patients chatted with the AI up to five days before their appointment through a secure web link. AMIE asked about their symptoms, gathered relevant history, and produced possible diagnoses for discussion with their physician.
The conversations were actively monitored by physicians, and none had to be stopped under the study’s predefined safety rules. That being said, supervisors did catch one hallucination and added missing clinical information in five interactions. Of the 98 patients who completed both stages, doctors reviewed AMIE’s transcript before 44 visits. They said it helped them prepare in 33 cases and could have changed their approach in 25. AMIE’s list of possible diagnoses included the eventual diagnosis in 90% of cases. Its top three possibilities had the correct answer 75% of the time.
The future I can get behind
What truly intrigues me is what happens as systems like AMIE mature. I would be far more comfortable with AI acting as a second pair of eyes than trying to make the final call itself. A doctor could compare their own assessment against what the system picked up, revisit a symptom that may deserve more attention, or check whether another diagnosis is worth considering. The study did not test AMIE in exactly that role, but it is a logical extension of what the technology is already capable of doing.
Healthcare is already experimenting with a similar approach with medical coding tools that can comb through doctors’ notes, lab results, and patient records for secondary diagnoses that may otherwise go unrecorded. Those systems are not free of criticisms of their own, but the underlying use of AI makes sense to me. Let it look for something a busy human may have missed, then leave the actual judgment to the doctor.