In our last Blink we talked about a potential harm of AI in our practices, hallucinations, the fabrications AI can make. Today we're celebrating AI at its best, and the positive outcomes it can have for a patient's vision. Let's look at a recent example in the news relevant to eye care.
Our story begins with Rhys, a 48 year old ambulance volunteer from Bedfordshire in England. His world was turned upside down in late 2024 when he collapsed and suffered a seizure. Investigations revealed a pituitary gland tumour. As time went on he suffered peripheral vision loss from optic chiasm compression.
Rhys opted for surgery, and was offered the chance to take part in the world's first AI-assisted pituitary tumour removal.
During the surgery the AI analysed the live camera feed rather than relying on pre-surgery scans. This provided the surgical team real time support in their decision making. It worked frame by frame on the endoscopic video, highlighting critical structures at the base of the brain as the surgeon moved. A millimetre of error can be critical here, leading to complete vision loss or even death. The tumour was removed without incident.
Following surgery Rhys was amazed to see his vision had immediately improved. After a week he was back walking independently without sticks. He continues to recover at home, increasing the length of his daily walks.
So how does an AI learn to do this?
The support system learned the same way the wine expert in our last post learns. It was exposed to a vast number of pituitary operations during training, learning which patterns indicate danger and which lead to success. It's possible during training for such a system to be exposed to more surgeries than a real surgeon sees in an entire career. It's worth noting this is one patient in an ongoing clinical trial, but the result is promising.
What was particularly clever about this AI was that it was built to communicate its own confidence, so the surgeon knows when it's unsure. This touches on what's known as the black box problem, where we often can't tell what an AI is actually looking at to reach its answer. Communicating confidence isn't a solution to that, but it helps.
This is what "human in the loop" means. Like it or not, AI has strengths we don't have. Used correctly it can augment us rather than replace us.
I think the future is bright if humans and AI combine their strengths, as long as the right safeguards are applied. After all, it is often a human who detects this problem in the first place, by making the decision to perform a visual field. The technology is remarkable, but our part in this patient's care still matters.
One final thought. If an AI steers a surgeon wrong and a patient is harmed, who is liable? The surgeon? The hospital? The AI developer? That's a question the profession hasn't answered, and one we'll come back to.