NeuroVoices: Michael Sughrue, MD, on Using Machine Learning to Build a Better Brain Map

The founder and chief medical officer of Omniscient Neurotechnology discussed how the use of machine learning and big data will be critical in revolutionizing our understanding of the brain and mental illnesses.

This is a 2-part conversation. Click here for Part 1.

In 1909, using the anatomical and cellular structure of the brain’s surface, German neurologist Korbinian Bordmann created one of the first early brain maps, dividing the human brain into 47 parts. Although revolutionary, the Brodmann model was limited in explaining more complex neurological functions—things that still frustrate clinicians in research labs today. More than a century later, clinicians across the world have seen little changes to their tool kits to navigate and understand the brain.

Because of the complexity of the brain, it might be time to lean on more advanced measures, according to Michael Sughrue, MD. Sughrue, the founder and chief medical officer of Omniscient Neurotechnology, believes that the use of approaches such as machine learning and artificial intelligence (AI) are the ways of the future, as these techniques can parse down and translate millions of data points. Omniscient has adopted this mindset too, by utilizing connnectomics, another wide lens, big data look at the brain’s functional and structural connections.

In the second part of a conversation with NeurologyLive®, Sughrue provided insight on the use of connectomics and how it is used in conjunction with other advanced technologies. He also stressed the importance of relying on these approaches to better understand complex aspects of the brain that are perhaps unattainable by humans.

NeurologyLive®: Can you provide an overview of what connectomics is?

Michael Sughrue, MD: Connectomics is an “omics” just like genomic or proteomics or any other omics. It’s a big data approach for biological data focused on brain connectivity or brain connections. That means specifically how parts of the brain are wired together and how they talk to each other. The key issue with connectomics is there’s a lot of ways to do it. You can talk about microconnectomics, how neurons are connected, but that’s dauntingly difficult. The human brain is extraordinarily complex—almost beyond comprehension. We’re not attempting to tackle it at that scale, I don’t know if technology is at that point where we can do that. What we do know is that at the macro scale, it’s still fairly complex, but manageable with existing computing speeds. What we can start to do is look and see how different parts of the brain are talking to each other. What makes it most excising is that it lets us look into the black box a little deeper about what’s happening in people’s brains when things aren’t working so well. Things like mental illness. To summarize, connectomics is a method for using things like AI and machine learning to make the brain make sense at its connection level.

How can AI help paint a better picture of what's going on in the brain?

We have to take a step back and say, where is the revolution in artificial intelligence happening? It’s not sentient computers, its machine learning. Machine learning are mechanisms for how you teach a computer to learn relationships from data. There’s a lot of different approaches in that, but the one everyone gets excited about are things like convolutional neural nets, deep neural nets, or what’s called deep learning. A lot of people have used medical AI or machine learning approaches to replicate a human’s performance. You get a bunch of radiology scans, you teach the machine to pick up a stroke, and then it flags it for a human. That’s all the utility in that.

What makes machine learning more interesting though, is that machine learning is very good at picking out a small amount of signal in noisy, complex data sets. Instead of this being five variables, it’s millions of variables. When machine learning picks up a stroke, it’s finding the 17 voxels that tell you there was a stroke happening. But you can turn it loose on other things. What we’ve approached—and what most people who do the connectome approach—is intelligence augmentation, or IA. This is how we go in and find something that may not be obvious to the naked eye that is useful. For example, if I have a patient with depression, how do I figure out why they have suicidal thoughts? Where in the brain is that happening? It has immense potential because we’re not trying to replicate humans, we’re trying to reach humans things they couldn’t do on their own. That’s where connectomics moves the goalposts: when it starts to show us things we didn’t know.

In the future, how much of this approach will be on the shoulders of humans vs computers?

The whole point is this: you shouldn’t use machines to replicate humans at things humans are bad at. If you talk about catching a pneumothorax on a chest x-ray, we can replicate that. Humans are good at that. Any things that humans miss are probably not clinically relevant. That’s a whole different thing than something that we’re profoundly bad at, which is understanding the human brain. If the human brain was easy to understand, we would be too simple to understand the brain. As a result, there needs to be a device that’s more sophisticated than the device that’s dying and that can really throw firepower at it. We can start thinking about automating humans out of it when humans become good at it. But for now, we need substantial help just to stay afloat. There may come a point where eventually the machine says, “Hey, do this,” but right now we’re drowning. We have 60,000 suicides in the US each year. If we were doing a phenomenal job with this, we should just automate this for efficiency, but you can’t make something efficient that you suck at. We need to stop sucking.

Transcript was edited for clarity. Click here for more iterations of NeuroVoices.