The untapped potential of AI

Sekou Remy believes the power of AI lies not in its potential to replace human decision-making, but in its ability to enable “superhuman” decision-making.

Given a simple choice between, say, Raisin Bran and Cocoa Puffs, making a decision is relatively easy. You might weigh the taste of one against the other, compare the nutritional value and price, and make your decision. No advanced technology needed.

But more complex decisions — like determining the best course of action to take when trying to rein in a COVID outbreak — involve weighing the pros and cons of multiple micro-decisions, using data obtained from a wide range of sources. Each of those sources has to be evaluated for veracity, reliability and relevance, with the data from each source weighted appropriately.

Portrait of Sekoou Remy

Sekou Remy, research scientist and technical lead at the IBM Research Lab in Kenya

Information overload

“There are so many challenges in evidence-based decision-making,” says Remy, research scientist and technical lead at the IBM Research Lab in Kenya. “Today, very few decisions are made with complete information. Even the simplest decisions often have innumerable potential outcomes, and for high-impact decision-making, such as in epidemic planning, it is not possible for a person to evaluate all possible options.”

With that in mind, Remy and his team are working to create systems that can help scientists and policy-makers use all the data available to them to make better decisions.

“I think of it as enabling superhuman decision making,” he says. “Today, only in the movies can superheroes do things like figure out that — of all of the millions of possible scenarios — only one would lead to success. Our approach leverages algorithms and scalable compute to provide this capability to any user. And we do so using models and data that they can recognize as trustworthy, either because they specified the models used, or because they trust the experts who contributed their models to the platform.”

“The team I lead works on getting AI, cloud and models out of the lab, and synergizing them to support decision-makers,” he says. Their current focus is on helping researchers and governments make better decisions regarding malaria and COVID-19.

A researcher writes on a glass wall

The decision-making process

Remy explains there are seven basic steps that go into making a decision. First, identify exactly what it is you’re trying to make a decision about. Then gather information and tools that can inform the decision-making process.

“The third step, which is where we begin to come in, is when you try to understand what alternatives could possibly exist,” he says. “So you don’t yet know which is the best solution, but you come up with a list of things that may work.

“The fourth step, then, is the process of evaluating those things, and the fifth step is where you actually need to choose something. After you’ve chosen, then you act, and then you review.

“Those middle three steps, of generating alternatives, evaluating those alternatives and choosing one — for certain users in certain situations — are very hard to do. Sometimes, it’s because there really are too many alternatives.

“Let’s go back to the cereal aisle. You have hundreds of options to choose from. Which one are you going to pick? Well, you’ll probably just pick the one that you’re comfortable with, the one you think tastes the best. You never even evaluate something that may be better than what you’ve had before.

“In that ‘gather’ stage, let’s say we create a simple AI model that ‘thinks’ like you as far as the way things taste. So if you give that model Raisin Bran, and you love Raisin Bran, that model is going to say, ‘this is amazing.’ But if you give it a cereal you hate, it’s going to say ‘this is terrible.’

“What we’re doing, from an AI perspective, is allowing algorithms to provide recommendations for things that might delight you. We understand what you like, because we’ve asked you to come up with some things that check the boxes for you. Then we create an algorithm that can interact with that ‘digital twin’ of you, so we don’t need to ask you as many questions to come up with alternatives you might like even better than Raisin Bran.”

Steps in the process, text and icons

Improving the COVID response

While choosing between breakfast cereals isn’t a matter of life and death, making choices that impact public health often are.

With COVID, for example, Remy explains, “you have a case where people have models and data, and they are in desperate need to better understand both what has really happened, and what they can possibly do going forward.” His team is working with leaders in Uganda, Senegal, the Democratic Republic of Congo and Nigeria to improve their response to COVID.

“That’s where we’re able to provide innovation. A computer model can predict what might happen to COVID transmission rates if you did a lockdown for two days, versus two weeks, versus if you did a lockdown for two months. But that's not enough to make a decision.”

No human being could parse all the relevant data as effectively and efficiently as a well-designed AI system can.

“The point of this research is that it can apply to any complex question,” says Remy. “For COVID, the question is, which combination of interventions makes the most sense for a specific context? For example, let’s say that because of population density, social distancing is something you can tell people to do, but the reality is that it’s not going to happen. So we look at what could occur, depending on what course of action is taken.”

Applying AI to malaria

While COVID poses a new threat to sub-Saharan Africa, the region has been waging a long-term fight against malaria, with about 80% of global cases, leading to an average of 200,000 deaths a year. With funding from the Bill and Melinda Gates Foundation, Remy’s team is working with key stakeholders in the malaria ecosystem within East Africa to help determine the most effective, efficient way to get the incidence of malaria in the populace down from 20 percent to less than five percent.

“There are many things you could possibly do,” he explains. “And when you do those things matters as well. For example, if you do an education program for kids, but you do it in December, right before the Christmas holidays, that program may not be as effective as if you had done it right at the beginning of the school year. Or if you give people bed nets, it might be more useful to give them the bed nets right before the rainy season, when there are a lot of mosquitos around, as opposed to in the dry season.

“And then there’s another element to consider — maybe instead of giving people bed nets, you give them money to improve their homes to keep mosquitos out or eliminate places where mosquitos can breed. Using the malaria models tuned specifically for the places you manage, you can get recommendations for which combination of interventions could work for you. These recommendations can even factor in concepts like the cost-effectiveness of interventions, or the quality of life improvements that can result in a particular location from particular choices.

“The thing is, malaria is totally preventable, so nobody has to die from it. If AI can help us better address the problem and target the funds more efficiently, the benefit could be enormous.”

IBM Research building in Africa

Mitigating bias

One question that frequently arises when AI is used to help make decisions is whether that AI is truly fair in its data assessments. Going back to the cereal analogy, if the model being used to pick your breakfast has been programmed to see raisins as repellent, then it’s not going to choose Raisin Bran, even if that is the best match for your tastes.

Remy says that while it’s impossible to ever be sure that bias is completely eliminated, there are ways to mitigate it. 

“In our case, the integrity of the computer model is the most important factor. There's actually a really robust modeling community that builds those things. In the case of malaria, for example, there's a malaria modeling community, and they are experts in that field. They are epidemiologists and bioinformaticians, and they know their stuff.”

While incorporating data from multiple models into an AI program can sometimes lead to conflicting conclusions, “from our perspective, that’s a good thing,” Remy says. “There’s value in understanding those disagreements. That basically tells us that we need more data there, or we need to ask a different question.”

Defining success

A perfect outcome of the team’s work would be to eliminate malaria or COVID, but Remy and his team recognize that those are longer term goals — toward that end, the immediate focus in on proving to researchers in the field that AI is something that can actually be used to help them make better decisions.

“What we bring to the picture is to say, ‘There are too many choices for you, as a human being, to search through and evaluate effectively. Let AI help you by searching through this very large space of possible actions to generate alternatives, evaluate those alternatives, and provide a recommendation.’

“It’s taken us two years to get in the room with some of these decision-makers,” he says. “While we may paint a very good picture, there’s an established ecosystem around decision-making that we are interacting with. Our job is to prove to that ecosystem that what we are proposing — while it may sound like science fiction — is something that is real, and that can help them in their work.”

Three team members sit casually by a laptop and IBM logo