I think we are actually no longer in the Big Data world. We’re in what we call a ‘broad data’ world. You may hear the term ‘deep data’ as well.
For us, broad data is about having myriad data types. In clinical you might have genomic data, phenotype data, medical scientific literature, electronic health record data, patient reported outcomes – there’s tonnes of data, and you need to be able to bring all that data together to drive insights.
That’s where we need to tap into AI, but it is not where most organisations are today from an AI perspective.
The goal of bringing all these data points together is being able to answer a lot of those questions faster and easier.
For example, we’re doing some work with the Broad Institute in the US around cardiovascular disease. By bringing together genomic data and electronic health records, we’re seeing if we can predict the likelihood of cardiovascular diseases earlier for a doctor and deliver a more robust genetic risk scoring.
That brings together very different types of data. In what I would call ‘narrow AI’, that would involve doing one thing with a slice of genomic data and doing something else with electronic health records. That’s valuable, you can get some power out of that, but you can’t scale it. Bringing those two data types together and being able to build artificial intelligence around extended data types allows you to reason and learn faster, which brings more advanced, in-depth insights to market.
The idea is that while you might have all this different data, you need to use it effectively – for example at the beginning of clinical trials. Protocol development is ripe for reinvention through data.
Definitely. People don’t really know if the protocol you’re creating is the right protocol, one that’s going to last for your six-year clinical trial.
Again, by using lots of different data points, you can change that. You can use data to understand your inclusion exclusion criteria better, what works and what doesn’t work. You can reduce the amount of endpoint modifications you have to make, and also get the right patients to your trial by targeting it accurately.
At IBM, we call AI ‘Augmented Intelligence’. Augmented intelligence is really about AI helping humans do things more effectively. It’s not about replacing humans, it’s about marrying the two together.
Technology is quite good at endless capacity. It’s good at looking at thousands of different endpoints. Humans are very good at learning and coming up with decisions, but sometimes they don’t have the capability to learn everything. Bringing those two together can be very powerful.
We can also look at bringing in different components to reduce human bias.
At IBM Research, we have something we call Project Debater. It wasn’t developed because we wanted to build a debater tool with AI. It was developed to prove that we can help humans reason and make well-informed decisions.
It is basically able to take a data-driven approach to generating a speech, and present that speech eloquently and with purpose. It can listen and comprehend another speech for a long period of time and rebut what it has heard. It’s able to think like a human does as far as dilemmas go – it can look at how a human reasons and comes up with arguments based on a situation.
We tested it in one debate on whether preschool should be subsidised in the US. Project Debater, who we refer to as a ‘she’, went against one of the top debaters in the world – Project Debater was for subsidising preschools, and he was against it. They had 15 minutes. The human had nothing in front of him, but Project Debater had about 10 billion text statements in her system that she could use. She wasn’t trained on the topic, but these 10 billion articles had some content related to it. In that 15 minutes, she was able to crawl through all that information, understand it and create a four-minute speech about why preschool should be subsidised.
Then her opponent came back and did a four-minute speech arguing the opposite. Project Debater’s speech was very factual because again it’s very evidence-based. His was more opinionated. Then Project Debater came back and rebutted that for four minutes.
You can watch the video to see who won, but it proved the concept that Project Debater was able to give more factual information in her speech. People said they learned a lot about subsidising preschool from her.
Imagine you’re in the emergency room, where doctors and nurses are always making last minute critical decisions. To be able to have a trusted system that you could have a dialogue with, that you could argue with, will help you make more informed decisions. It’s not going to make your decision for you, but it’s going to help you reason more effectively.
The reasoning side of AI is becoming increasingly important. When we brought Watson and other solutions to market, narrow AI was an emerging technology.
With narrow AI you can quickly get very good results from a thin slice of data, but narrow AI can be very complex as well. We did some work with Roche where we developed a model to predict the early risk of chronic kidney disease in diabetic patients leveraging real world data. It had 79% accuracy under the curve. That’s still narrow AI because we were able to use very good data – half a million patient records – to build an algorithm to predict the likelihood of an outcome.
But to take that and do it for something like the likelihood of a mental illness episode, you need to almost completely start from scratch. Your narrow AI is tied to your data – it’s very powerful but it’s not scalable. As we start to work with things like Project Debater we need to move away from that narrow set to a broad set, and actually have the system adapt itself when the data changes.
I don’t think we’re near the stage where we’re using technical systems completely stand-alone to diagnose at a global scale. It can provide a recommendation but then it’s still up to a human expert to make the diagnosis and decide how to go forward. Even if the technology has achieved this ability, I don’t think that humans and society are ready for it. You have to demystify AI first.
In sales and marketing, I still think people like to talk to each other to some extent. You’re still going to get that.
Nevertheless, do we need thousands of sales reps running around? No. We’ve done a lot of projects in the past around segmentation and targeting of doctors – based on how they prescribe and what they do. This is not necessarily using AI, but using advanced analytics. That’s also important to remember – not everything is AI nor does it need to be.
It’s been a journey through the evolution of the technology. I come from a technical background – I started programming when I was very young. I was one of those very rare female programming geeks in the ‘80s. To make a line across the screen was hundreds of lines of code. Nowadays you can build a whole app with 10-15 lines. Technology advances and that’s also the case with AI. We’re not done – there are a lot of ways it can still go.
We developed a solution with Medtronic called Sugar.IQ which is a good example of going from narrow AI to broad. It’s a personalised recommendation engine for someone with diabetes.
It’s taking in lots of different data types already. It’s taking in insulin information, the continuous glucose monitoring system, it’s taking in what you’re eating, how you’re sleeping, what sports you’re doing, where you’re living, what the weather is like. All those things are being brought in to give you information to help you decide on what you should eat, when you may want to do some exercise, etc.
It can predict with about 89% accuracy if you’re going to have a hypoglycemic event two to four hours in advance using algorithms that were developed by Watson Health. It’s quite powerful.
I do think that as more and more devices enter the market we will get more broad data, but a lot of what they’re bringing is noise. You have to get away from some of that noise.
Yes, and I think that’s a good thing actually.
People have asked the question, “Does IBM want to become a pharma company?” Actually, we are a healthcare IT company but we don’t want to become a pharma company – we want to work with our clients to bring them what we’re good at. We believe we’re good at applying disruptive technology, like AI or blockchain, and bringing clinical expertise with it.
Life science is prime for disruption. It has such complex data and you can do so much with it.
Christina Busmalis is director at IBM Watson Health, and global life sciences go-to-market leader Europe & Asia Pac. Prior to this role Christina was the global director of technology for the IBM account team, working with Novartis since 2006 where she focused on identifying and leveraging disruptive technologies to address complex business challenges across the life sciences value chain. She is also a member of IBM’s Academy of Technology. She has over 25 years of consulting, business development and project management expertise across Europe, US, and Asian markets and is based in Zurich, Switzerland.
Dr Paul Tunnah founded pharmaphorum in 2009, which is a content and communications company offering industry leading publications (www.pharmaphorum.com) and a strategic consultancy (www.pharmaphorumconnect.com). He is a recognised author, speaker and industry advisor on content marketing, communications and digital innovation, having worked with many of the world’s leading pharmaceutical companies and the broader ecosystem of healthcare organisations.