In June, I had a fall that resulted in a toe fracture. It coincidentally happened to be the toe that I used to have gout symptoms in before I was on the right medicine to curb the gout. Nineteen days after my injury, I had some pain in the toe. I thought it might be due to the bone healing. I thought it might be due to something I had done to aggravate the toe the day before.

When I went to bed that night, the pain increased. I had my first gout attack in a number of years. That felt odd. The following morning I went online to research gout and my healing toe injury. The AI (artificial intelligence) search engine explained that a healing toe injury can produce a gout attack. I had figured that might be the case. AI confirmed it.

This is the proper use of AI for medical circumstances. It should not be to type in symptoms and see what kind of diagnosis AI can suggest. Unfortunately, that is what many people do. It means overdiagnosis in some situations. It means an incorrect diagnosis in others.

There was an article in the October issue of Pediatrics, the journal of the American Academy of Pediatrics. Its title was “AI at the bedside: An argument for a Training Wheels approach”. The implication was that we need to be careful about overusing AI in the medical field.

One of the topics covered in the article is related to consent for AI use. Some offices use AI-generated notes based upon the history provided by the patient. The article asked the question about whether all patients agree with this or not. There are currently no guidelines. Patients may prefer a non-AI note in the medical record. This is an area that will need to be worked out in the future.

Guidelines will need to address the risk of error. They will need to address the risk of privacy. They will need to address the risk of bias. They will need to address how the information is stored. They will need to address how long it remains in storage. They need to address who has access to the information.

Another issue pointed out is that AI generators are programmed to focus on medical details. There are items in the patient visit that they may miss. An example is related to the article I wrote a few weeks ago about the importance of body language in doctor visits. The article refers to a doctor noting poor eye contact from the patient. In that case presented, that was important to the diagnosis.

Another example is the important information that the AI generator sees as non-medical in nature. The article refers to a child developing an “intense interest in trains”. The AI generator ignored that info. However, the obsessive compulsive behavior was critical to the ultimate diagnosis.

A danger that was raised was related to the fact that the more physicians got used to using AI generated notes, the less likely they were to look at them critically.

The dangers to physicians were described in a series of words. One was deskilling. That refers to stopping to use a previously learned skill. Another was never-skilling. That refers to not learning a needed skill in the first place. The third word was mis-skilling. That meant learning the wrong behavior.

Overall the use of AI in medicine is here to stay. However, like any new medical technique, it needs to be looked at critically. Only in that way can we make sure we are getting the best result possible.

If I had just looked up pain in my toe without the history of the broken toe almost three weeks earlier, the AI would have called it gout attack. I might have thought my medication was not working and needed to be changed. Fortunately, I have gone four months without any further gout attacks.