New AI platform led by Fred Hutch aims to accelerate cancer breakthroughs

A coalition of cancer centers led by Fred Hutch on Wednesday unveiled their plan to use artificial intelligence in the fight against one of the world’s most intractable diseases.

The Cancer AI Alliance (CAIA) is led by Fred Hutch, and includes the Dana-Farber, Johns Hopkins, and Memorial Sloan-Kettering cancer centers.

For years, AI boosters have claimed it will find a cure for cancer, often as a reason to continue developing the technology despite concerns about safety and the environmental impact of training large language models. But using AI in healthcare is difficult, because medical data is highly sensitive and heavily regulated.

The new AI gets around that through a technology called “federated learning.” Using it, the CAIA platform is able to “travel” to each cancer center, learn from the locally stored data, and then bring that knowledge back to a "central orchestration node."

“ You send code to the data instead of data to the code,” said Jeff Leek, Chief Data Officer at Fred Hutch and head of the initiative.

Fred Hutch says that the CAIA platform is the first of its kind. Computing power and technical expertise were donated by Microsoft, Amazon, the Allen Institute for Artificial Intelligence, and a handful of other companies.

“This is a way to reason across the data of all the different cancer centers very rapidly, and to put tools in the hands of physicians that enable them to ask questions in plain language and then calculate that across a million patients,” said Leek.

The platform is launching with eight initial projects, including predicting how patients will respond to immunotherapies and identifying rare cancer trends.

Leek also demoed a project that allows doctors to ask questions like they would using any other AI chatbot. He entered a sample question about lung cancer, “written out in plain language, just like you would type into a ChatGPT interface.”

The AI instantly started spitting out code that — in a real use case — would be checked for accuracy by a biostatistician at Fred Hutch, then sent out to the four cancer centers.

“This is a way to reason across the data of all the different cancer centers very rapidly, and to put tools in the hands of physicians that enable them to ask questions in plain language and then calculate that across a million patients,” Leek said.

One of the challenges of large language models and AI systems is they’re only as good as the data they’re trained on. That can build biases into the models if there are gaps in the available data, if for example, a community without access to healthcare isn’t reflected in the clinical record.

Leek says the CAIA team is aware of that challenge, and hopes to address it by widening the range of participating institutions.

"One of the benefits of federation is that it's quite scalable," he said, "so one of the things that we're actively looking at is for additional partners and cancer centers to come on board such that that representation and diversity can be further enriched over time."