Algorithm that has learned to assess heart scans from human specialists to be tested in US hospital
Summary
The first ‘all-rounder’ artificial-intelligence algorithm has learned to assess heart scans from human experts. It could eventually help clinics without heart-imaging specialists to spot uncommon heart conditions. The algorithm will be trialled in a US hospital starting next year.
An ‘all-rounder’ artificial-intelligence algorithm could eventually help clinics without heart-imaging specialists to spot – and treat – rare heart conditions earlier. The AI system will be trialled in a US hospital next year to see whether its assessments match those from human experts.
How to train an all-rounder algorithm
What if you could train a system to analyse heart scans without explicitly telling it to do it?
Dr Rohan Shad, University of Pennsylvania, US
Medical algorithms are usually trained to do one specific task, such as looking for one disease. AI systems trained this way are usually very good at their job – but they easily get confused by our messy reality.
“When you train a machine-learning system to find only disease A or only disease B – well, sometimes, there’s both,” says heart specialist Rohan Shad from the University of Pennsylvania, US. “What if you could train a system to do things like [analysing heart scans] without explicitly telling it to do it?”
Shad and his team wanted their AI algorithm to be an all-rounder. They wanted it to learn how to assess heart scans the way human experts do, without any constraints around what to look for.
To do this, they showed the algorithm heart scans from 19,000 patients at US hospitals and asked it to match each scan with a corresponding medical report written by human specialists. “The matching system really was key because you could trick the system into learning many things at the same time,” Shad explains.
Foundation model
The algorithm’s first test was to analyse heart scans from nearly 46,000 UK Biobank participants
When tested on heart scans from UK Biobank participants, the AI system could spot 39 heart conditions, including uncommon inherited diseases such as hypertrophic cardiomyopathy and amyloidosis. “The system could also tell which gender the patient belonged to, without me ever asking it to learn this,” Shad says. “It could separate out young versus old patients, without me asking it to do that.”
Yundi Zhang, a healthcare AI specialist from the Technical University of Munich in Germany, is impressed by how well the algorithm spots uncommon diseases. For example, the system had only seen 92 cases of hypertrophic cardiomyopathy during its training. Yet its performance is like “something you can use clinically today”, Shad explains.
“We call it a foundation model,” says Shad. Just like the algorithm behind ChatGPT, these AI systems can adapt to new tasks with very little or no instructions.
First real-world test for new AI
We don’t have enough specialists compared to how many different cases we have every day.
Yundi Zhang, Technical University of Munich, Germany
Algorithms could eventually help doctors make better decisions, says Zhang: “Does this patient need to have an urgent appointment, or can they wait a little bit? We don’t have enough specialists compared to how many different cases we have every day.”
“A lot of diagnosis of [rarer] conditions relies on really, really expert level of care,” Shad adds. “That level of care may not be available in every single hospital. This is something that can bring expert-level diagnostics down to an algorithm.”
Shad’s team will start testing their system at the university’s hospital next year to see how it performs compared with human experts. For the time being, it will simply run in the background and won’t be used for diagnosis.
Still, Shad is excited to see how his algorithm fares in the real world. “I build these things in the hope that one day this helps patients,” Shad says. “The closer we get to that goal, the happier I sleep at night.”
