Presagen Webinar Series: The dilemma making AI globally accessible & affordable

In Feb 2023 Dr Jonathan Hall presented the webinar “The dilemma making AI globally accessible & affordable”. Presagen’s webinar series presents new technologies and challenges related to AI in healthcare, women’s health and fertility. Below is the transcript of the presentation, the webinar video, and the presentation slides.

Slide 1

Hi everyone, Thanks for joining me. I’m Jonathan from Presagen.

Today, I will be talking about how AI can become accessible and affordable to billions of people globally.

Slide 2

In healthcare, many places in the world suffer from unequal access.

Without the same accessibility to vital healthcare services, people can struggle to get the care, information, and treatments they need.

The affordability of healthcare can also be a huge discrepancy from place to place.

In some cases, the most technologically advanced treatments are either unaffordable, or not present altogether.

Minorities, and people in remote areas suffer the most. They simply can’t get access to the healthcare they need.

Slide 3

Artificial Intelligence, or AI, is a powerful tool in healthcare that is becoming more available every day.

But is it available to people worldwide?

For AI to be successful in healthcare, it needs to be Accurate, that is, it needs to have good, consistent performance.

It also needs to be Accessible. That is, it is not exclusive to one hospital but can be used by all hospitals, in an unbiased way, so everyone can access it.

Finally, it needs to be Affordable, that is, it is not a premium service, but it can be offered at low cost.

Slide 4

There is a dilemma in creating AI for healthcare.

It is really difficult to achieve all 3 of those items!

Let’s focus on the AI being Accurate, and making sure it has the best performance.

Normally, Medical AI Companies will create an AI that works really well in a single hospital. It is trained on the data in that hospital, and calibrated to work well in just this one setting.

While this AI has the potential to exhibit the best performance in this setting, has the AI really captured the knowledge it needs to work worldwide?

It’s not widely Accessible. It can’t be transported to another hospital and still work effectively. It is definitely biased on the hospital’s key demographic. It is not ‘for all’.

Also, such a custom development is not Affordable – it becomes a premium-use product in a single hospital rather than being widely used and accessing the benefits of scale.

It is highly cost-inefficient to make this AI, and so people can’t get the good healthcare results they deserve.

Slide 5

Instead what if the AI was reporting less Accuracy? But actually, it might be the AI is really ‘more’ accurate.

If it could report good, consistent performance, but also work in multiple hospitals, this AI might be better suited at solving a global problem.

If the AI was able to scale to different hospitals, and work in an unbiased way on all these different demographics and patient groups, a single AI can become a powerfully Accessible tool to people around the world, including places that never have trained an AI. Consider rural communities, and clinics that are too small to create their own expensive AI.

Now, since the AI is useable in different settings, it’s now much more cost-effective to produce. That means that it becomes more Affordable to people around the world. It can benefit from the economy of scale.

Slide 6

In order to go forward with a practical way of training scalable AI in healthcare, we need a new paradigm in order to maximize the performance of the AI, while maintaining Accessibility and Affordability.

Slide 7

First, we need to have a globally diverse data set.

This means that, before we do any AI training, we need to be able to have many demographics and clinical settings around the world represented.

This ensures that an AI will be trained to be unbiased, and take in the context of many different settings.

To make sure we can use AI on data around the world without seeing or moving private data, we created a Decentralized Federated AI Learning algorithm that can use global datasets and not the private information.

Slide 8

Next, we need to ensure the AI is able to work effectively without being tricked or harmed by bad data.

Data around the world is very messy, and even a small amount of errors can hamper the AI’s abilities.

Furthermore, medical data can be inherently poor quality, because it is often subjective, based on diagnosis without a firm ground truth, and is subject to uncertainty.

To solve this problem, we created an Automated error detection system that can find and remove data so that it doesn’t confuse the AI. It is able to identify errors even experts can’t identify.

Slide 9

Thirdly, we need ways of training the AI to be Reliable and Robust.

We found that gaming ‘Accuracy’ can lead to a downfall, where the best AIs simply won’t scale to a brand new hospital.

Instead, we use new Confidence-based metrics, that ensure that the AI accuracy is good in a brand new hospital it has never seen before.

This helps create Reliable AI that is Robust and capable of being used in the real world.

Slide 10

Lastly, we need to use Knowledge to guide AI to make sure it solves the right problem we intend it to.

We can’t just “throw data at AI” and expect it to yield for us a good result.

By using expert knowledge of the problem domain, such as is gained by working with top healthcare professionals who understand the problem, we can guide the AI to solve the most crucial problems.

It also means the AI is not trying to learn everything all by itself. Instead, we can be more targeted, by reducing the problem to a simpler, smaller problem. We can remove irrelevant patterns that are complex to learn, using ‘masking’ and it allows it to work across multiple demographics much more easily.

Slide 11

By using this new paradigm, we are able to follow the process that ultimately ends up with an AI that can work on multiple clinics it has never seen before.

We train the AI on a diverse dataset, then we validate the AI by removing errors, and selecting the highest performing AI based on confidence, rather than accuracy.

Then we can test this Robust AI on a Blind Dataset – a real world dataset.

Finally, we can show the AI is Reliable by testing it on a Double Blind dataset – real world clinics which it has never seen before.

Slide 12

All of these methods together show a practical way that AI can scale reliably to many different hospitals, making AI Accessible and Affordable.

It makes sense commercially, in that it is cost-effective, and socially, in that it is unbiased and accessible.

To carry this through, it needs this new paradigm of a ‘global first approach’ for creating and testing the AI.

It means AI can be used to solve all kinds of problems, it is not only the ‘famous’ problems that AI is only cost-effective to solve.

Thus it opens up whole new areas, such as women’s health, that now become feasible.

Slide 13

I would like to thank you for joining this webinar.

This presentation is available on our web site at Presagen.com.

Thank you.