AO Foundation: AI enables patient privacy breakthrough in clinical applications

High-quality? Highly sensitive

For over 60 years, the AO Foundation has been pushing the boundaries of medical research. With more than 4,500 patients enrolled in studies, and a global community of some 460,000 professionals, AO is at the very forefront of the intersection between healthcare and technology.

And that pioneering spirit made Zühlke a natural fit when the AO Foundation was looking for a technical partner on a new project: exploratory research into how AI modelling can improve spinal CT scans.

The challenge? To combine and use typically disparate datasets – from a raft of clinics and hospitals – without compromising patient data.

‘High-quality patient data is the basis for training artificial intelligence for any medical application’, explains AO’s Head of Technology Transfer, Roland Herzog. ‘That data is usually available in a given hospital’s local files, but a number of laws and data privacy regulations have to be followed if you want to access or use it. And hospitals are generally very reluctant to grant that access’.

A complex web of patient consent, clinical trial contracts, and regulatory red tape often makes taking datasets out of local repositories an impossible task. But a lack of usable data creates a barrier to advancements in AI modelling that could help automate some of the heavy lifting in CT scan diagnoses.

So, using a combination of next-generation machine learning techniques and a better-connected data ecosystem, that was exactly the barrier we at Zühlke set out to tackle.

Decentralising the data

In order to really understand what this project has made possible, we need to break down a pretty complex term from the AI space: federated learning.

In essence, federated learning is the process of training AI models on datasets that exist in separate siloes by sharing learnings without sharing the data that generated them.

Imagine it like this: you and a group of fellow chefs are trying to perfect a recipe, but you’re separated by geography – meaning you can’t taste each other’s food. Instead, you each make the dish, taste it, and share your tips for improvement. Each chef incorporates everyone else’s tips and repeats the process until, ultimately, the dish is the best it can be.

That’s federated learning in a nutshell.

‘If you want to train an artificial intelligence model’, explains Roland, ‘you might need data from 1,000 patients data to get good results. But you might not have 1,000 patients in one single hospital, or the 1,000 patients in this hospital may not have the desired diversity’.

'You might need data from 1,000 patients data in order to get good results. But you might not have 1,000 patients in one single hospital'.

‘In that case’, he adds, ‘you’d need to have data from different hospitals, so you’d either export all these different datasets to one local database or – as with federated learning – you leave that data in each hospital and bring your AI models to them’.

The process is then one of multiple ‘loops’ where each hospital’s local AI model shares parameters. Those shared parameters are incorporated into each model, the training runs again, the findings are shared again, and each time the model improves.

And, crucially, that decentralisation of data means that no patient records have to be shared externally.

With spinal CT scans chosen as our testbed case, Zühlke and the AO foundation were able to move from preliminary workshops to a full, working proof-of-concept in just a few short months, opening the door to game-changing AI implementations in healthcare and beyond…

A big breakthrough, no bones about it

Zühlke’s work with the AO Foundation was designed to test the federated learning model on spinal CT scans and imaging data, with a view to improving how well AI can tell bones apart from tissue, muscles, and organs. Roland explains:

‘Imaging data like this delivers information in three dimensions, and if you want to use that for planning surgeries then you need this 3D data to accurately represent the human body. That means determining the borders between bones and organs. These borders can be blurred, though, making it less easy to determine exactly what’s what’.

The typical answer here is to manually assess a scan, but this takes time and is subject to human error. ‘The more you can optimise this process’, Roland says, ‘the faster and better things become. So AI in this instance is an attempt to automate the process of defining a human body out of a cloud of three-dimensional data’.

'The future for this concept lies wherever large datasets are needed to train AI, but where data privacy is a big issue'.

As demonstrated at the Intelligent Health Summit in 2020, our proof-of-concept showed that using federated learning is a win-win in this kind of situation; it can dramatically improve those AI models without needing to centralise sensitive data.

And that’s exciting because it could enable much more widespread use cases.

As a starting point, stakeholders at our partner clinics helped establish the foundation for future implementation across clinics. But the technology could benefit much more than just the healthcare industry.

Say, for example, that a country’s tax data is separated into different provinces or states. You might have an AI model that needs to be trained on all this tax data, but you might not be able to get that data out of each state for privacy reasons. With federated learning, you could train that model without having to move or share any data.

The future for this concept could be in medicine, it could be in banking, or it could be governmental. Ultimately, it’s wherever large datasets are needed to train AI, but where data privacy is of the utmost importance.