GabyPeople already turn to “Dr. Google” for medical answers. Now Google has developed an AI that answers medical questions with 92.6% accuracy, performing as well as doctors on disease-related queries.
Artificial intelligence as a medical consultant
The AI’s accuracy matched the responses of nine doctors from the UK, the United States, and India, who answered the same 80 questions.
Google researchers say this technology does not threaten the jobs of general practitioners. However, AI provides detailed and accurate answers to questions such as “Can urinary incontinence be cured?” or “What foods should be avoided with rosacea?” Researchers suggest that in the future, this consulting robot could be used for medical helplines.
The senior author of the AI program research called Med-PaLM, Dr. Vivek Natarajan, notes that their goal is to make the system trustworthy for both doctors and patients. When people turn to the internet for medical information, they encounter an overwhelming amount of content. As a result, people may latch onto the worst-case diagnosis among many possibilities, creating unnecessary anxiety. Dr. Natarajan says the latest language model will provide a brief expert assessment that is unbiased, cites sources, and states any uncertainties. It can be used to triage and determine the urgency of a person’s condition and what medical care they need. The research leader believes this will be very helpful in situations where there is a shortage of experienced doctors.
Well-trained medical consultant robot
As noted by the publication Daily Mail, the AI program Med-PaLM was adapted from PaLM, a language model that had not undergone specialized health training. Researchers carefully retrained the model using higher-quality medical information.
The medical robot was also taught how to report uncertainty when it has gaps in knowledge. Previously, it was trained on doctors’ responses to questions so it would reason correctly and avoid providing information that could harm a patient.
The created medical consultant robot had to pass a benchmark test that combines six datasets of questions on medical topics, scientific research, and inquiries from medical consumers. Additionally, it received a dataset of 3,173 medical questions that people were searching for online.
According to the research, Med-PaLM provides potentially harmful answers to patients only 5.8% of the time, compared with 6.5% for the surveyed doctors.
It still has shortcomings and is not a substitute for real doctors.
There is still a risk of “hallucinations” in AI. This means the model can generate responses without data to support them. Engineers do not fully understand why this happens, and the technology is still being tested.
AI expert Anthony Kohn says “hallucinations” will always be a problem for large language models because of their statistical nature. That is why they should be treated as assistants, not decision-makers.
Dr. Natarajan says the robot can answer the kinds of questions doctors face on medical exams and in real practice, and he calls that exciting. He adds that doctors don’t need to worry that AI will take their jobs; it will free up time for patient communication.
However, Professor of Information Technology James Davenport says that while AI can answer medical questions, consultation is not the same as practicing medicine. If medical practice were only about answering questions, we wouldn’t need teaching hospitals or years of post-training practice. For that reason, he says it’s too early to conclude that AI is encroaching on the jobs of medical professionals.
