In 1982, Joy Milne detected her husband’s Parkinson’s disease with her heightened sense of smell. She wouldn’t realize the source of the scent until after her husband was diagnosed with Parkinson’s over a decade later. The couple attended a support group, and Milne smelled the disease on almost every person there.
The Milnes’ case was the first indication that Parkinson’s had a distinct smell. Milne could smell the disease years before doctors could make a conclusive diagnosis. For diseases like Parkinson’s, where early detection and treatment are vital, this was an incredible discovery.
Diseases like Parkinson’s, Alzheimer’s and cancer change a person’s body chemistry. They cause different chemicals to be produced, which can often be detected in sweat, blood or urine. These chemicals have a different smell that most humans cannot detect. However, there have been dogs trained to smell diseases from cancer to COVID.
Vasant Dhar researched how AI could be trained to smell diseases. Credit: Dmitri Petrov.
Now, scientists like Vasant Dhar, a professor of data science at NYU, are investigating the possibility of creating an AI model that could use smell to diagnose diseases like Parkinson’s.
How to smell with no nose
The idea is this: a doctor would collect a sample – blood, urine, etc. – from a patient and send it to the lab. At the lab, scientists would give the sample to a mouse to smell, recording the neural response in the mouse’s olfactory bulb, the part of the brain that processes the sense of smell. That neural response would then be given to the AI, which would compare it to a database and determine whether it matched any of the diseases in the database.
“The first order of business is to create a shared database such that any researcher from anywhere in the world has a benchmark against which they can run their algorithm,” Dhar said.
The database would be similar to ImageNet, the iconic image database comprising more than 14 million images that helped set the stage for the current AI wave and has been used to train self-driving cars. It contains millions of images, including some of road scenes and things like stop signs and crosswalks. Each image is labeled so a system can be trained on the database to recognize items on the road.
“With smell, it’s harder to do that, because you have a molecule. You can describe it in terms of language, but it’s very inexact,” Dhar said.
Smell can be highly individualized; two people could describe the same smell entirely differently. This is why the database needs something Dhar calls “the truth” of each smell. In this case, the neural signal from the mouse. Just like the ‘truth’ of an image of a stop sign would be the label ‘stop sign’, the truth of the smell of Parkinson’s would be the neural signal it excites in the mouse.
Dhar says science is still “a few years away from conducting research where we can build good predictive models that have a low false positive rate and a low false negative rate.” It can be tricky to optimize both precision (a measure of false positives) and recall (a measure of false negatives). Improving one tends to lower the other.
“There will always be errors. No system will be free of errors. But the question is, is it good enough to put into practice because its false positive rate and false negative rate are acceptably low,” he added.
Considering ethical issues
Integrating AI into healthcare will require careful consideration of ethical concerns. “The kinds of ethical issues that we need to be concerned about are around the unethical uses of data,” Dhar said. There needs to be people monitoring how the platforms and data are being used, he added.
He also mentioned the importance of the prompt given to the AI. A simple task like “Help your patients and establish a close relationship with them,” to an AI acting as a therapist could have dire consequences. “Maybe the AI realizes that the best way to establish a close connection with their patients is to disarm them, or to flatter them or to do something so that the human begins to trust the machine, and the machine is just being manipulative,” Dhar said.
Already, there have been cases of extreme delusions and even instances of AI chatbots encouraging or leading young people to commit suicide. There is even a word for it; AI psychosis “describes how interactions with artificial intelligence can trigger or worsen delusional thinking, paranoia and anxiety in vulnerable individuals.”
“One of the biggest challenges is in AI for mental health, because that’s an area where the machine can be incredibly useful, but it can also be incredibly dangerous,” Dhar said. Mental health patients may be especially susceptible to manipulation. Interacting with an AI could worsen their illness, as it has been known to support delusional thinking and encourage harmful actions.
At its core, though, the issue of AI in healthcare doesn’t change what healthcare professionals have always done. “The ethical issues are no different in medicine than they’ve always been: looking after the health of the patient and doing the best possible job you can,” Dhar said.
