From Hackaday:
Researchers at UC Santa Cruz have developed a proof-of-concept device called a-Heal, intended for use inside existing commercial bandages for colostomy use. The device is fitted with a small camera, which images the wound site every two hours. The images are then uploaded via a wireless connection, and processed with a machine learning model that has been trained to make suggestions on how to better stimulate the healing process based on the image input. The device can then follow these recommendations, either using electrical stimulation to reduce inflammation in the wound, or supplying fluoxetine to stimulate the growth of healthy tissue. In testing, the device was able to improve the rate of skin coverage over an existing wound compared to a control.
Yes, the intent is good but has this been tested with patients of all body types and complexions because we know how machine learning models don’t work very well with dark skin tones in medical settings in particular. This is from the article that Hackaday quoted and it’s quite telling:
The AI model used for this system, which was led by Gomez, uses a reinforcement learning approach, described in a study in the journal Bioengineering, to mimic the diagnostic approach used by physicians.
Why are we mimicking the work of a diagnostician? While there are cases where a physician can’t necessarily meet a patient, why can’t we address that part and improve it? Also, the idea of a faceless mechanism playing the role of a qualified doctor feels very off to me.
I’m so tired of ML being used in places where there’s a case to improve a human process or connection. And before you call me a Luddite or something else that’s inaccurate, I just wrote about a good ML use case yesterday regarding seismology. I just think ML in medicine and public health isn’t where people think it is and it can be very dangerous.
Filed under: bioengineering Hackaday public health reinforcement learning