'AI will be a catalytic force to reducing inequalities, including in global health': Gates Foundation's global health tech chief

Dr. Zameer Brey, who leads technology diffusion for the Gates Foundation, has always focused on systems and scale. After obtaining his medical degree in South Africa, he decided not to work as a clinician because he "realized that having one more doctor would make no difference in a broken system."

Instead, he started to work on transforming systems: "My colleagues often refer to me as the confused doctor. I don't do many things well, but the one thing I instinctively do well is connect dots. As a result, I've been able to support system process transformations, organizational transformations, and even a few nationally scaled innovations, simply by being able to connect dots."

Dr. Brey is now one of the leading figures working at the intersection of AI and global health and development, which he believes will lead to improvements in health outcomes and equity on a scale never seen before. "Our hypothesis is that AI will be a catalytic force to reducing inequalities across multiple programs, including in global health," he says.

Exemplars News spoke with Dr. Brey, who also leads a Gates Foundation internal AI Task Force that is exploring ways the technology can be applied to all the organizations focus areas and oversees the foundation's AI-related Grand Challenges grants, about how AI could transform global health – as well as the need to remain vigilant to ensure the technology is distributed equitably and accounts for the needs of the Global South.

Could you tell us about your personal journey – how did you come to work at the intersection of AI and global health and development?

Dr. Brey: After I qualified as a clinician in South Africa, I realized that having one more doctor would make no difference in a broken system. I realized we needed to work on the system so that committed health care workers could deliver the care they wanted to deliver without the system holding them back. This is even more so with marginalized and poor patients who didn't get the care they deserve. I wanted to augment my skill set so went to business school – my intention was to have both a clinical background and a deep systems orientation to work at that intersection.

Since then, I've been working on organizational improvements and turnarounds in the health care space. Basically, I've been trying to understand the best way to introduce health innovations in lower-middle-income countries in a way that drives impact. My colleagues often refer to me as the confused doctor. I don't do many things well, but the one thing I instinctively do well is connect dots. I can look at a problem, understand it deeply, and then connect it to innovative technological capabilities that could resolve that particular problem. As a result, I've been able to support system process transformations, organizational transformations, and even a few nationally scaled innovations, simply by being able to connect dots.

How is the Gates Foundation integrating AI into its global health initiatives?

Dr. Brey: For context, I lead the Gates Foundation's task force on AI. It's been set up as a cross-foundation initiative, supporting global health, development, etc. It really serves as a coordinating function – a lot of the actual work [on AI] is being done at the program level. As an organization, our hypothesis is that AI will be a catalytic force in reducing inequalities across multiple programs, including in global health.

Honestly, when we think specifically about health, there are a few things we've been struggling with for a long time. At the top of the list is human resources. How do we crack that nut? Are we really going to say we need another 100 years to have enough doctors, nurses, and community health care workers? According to WHO standards, Africa currently has a deficit of 4.2 million health workers. That's not a problem we can solve quickly. Even if we try and solve it quickly, there's a fiscal implication to adding 4.2 million health care workers. It's not surprising that you actually have doctors and nurses sitting at home because there isn't the fiscal capacity to actually deliver the services.

So, we really have to think creatively about what AI offers – not only at the base of the pyramid such as community health care workers, nurses, and doctors, but even at the specialist level. I think it's often lost on folks in the Global North that there are countries in Africa that have fewer specialists than one district in the US. Malawi has probably five or six radiologists. That's it. So, in that particular context, most of the population will never actually be able to access the care they need if things continue along the current rails.

How do you think AI can specifically advance global health outcomes?

Dr. Brey: For us, in terms of how AI can support global health priorities, there are essentially four buckets. The first is around the limited number of health care workers and clinical decision support. How do we support the very constrained, limited number of health care workers and allow them to deliver care with high quality, speed, and reliability? I want to caveat right now that we don't see any AI solution replacing health care workers. We're always thinking about how to strengthen their roles and positions in health care systems. The role of AI is closing some of the gaps in terms of clinical decision support – helping health workers, whether it's community health workers, nurses, or clinicians, make better and faster decisions at the point of care by being able to synthesize the stream of outputs and information.

For example, with large language models and AI, could we actually start creating electronic health records (EHR) using voice-based technology that could assist with point of care decision making? One of our grantees in Pakistan is proving this right now. She's basically putting cell phones on desks – obviously under institutional review board and ethics approvals – and recording doctor-patient consultations and translating them from Urdu into English using automatic speech recognition. Then the English version goes into a large language model in a format that she's already set up and comes back as an EHR in real time.

The magic is that it's an EHR with some risk stratification layered on top of it. Let me give you a real-life example. A pregnant woman came into a clinic and the nurse recorded that she had an elevated blood pressure of 150 over 100, which the nurse entered into the EHR using the voice-based tool. The doctor then had a consult with the woman and said, 'Congratulations, your pregnancy seems to be going fine. We don't need to do anything else. Here are your iron tablets. We'll see you in three months.' But because the nurse had entered certain risk factors, the phone immediately started buzzing and basically said, 'Hang on, we haven't addressed the high blood pressure.' The doctor picked up the phone and he said, 'Oh, I didn't realize her blood pressure was 150 over 100.' The woman was saved from preeclampsia.

The second bucket is direct-to-consumer health applications. As health care professionals, we're not always trained to engage with people and understand their preferences, risks and fears. Often that particular task is given to the least qualified person. It's often seen as a nice-to-have, which I think is why we sometimes don't get the results we hope for with our biomedical tools. In many low and middle-income countries, we've also taken the concept of health care and we've stuck it into the four walls of a clinic and these clinics and health care systems are busy. But wellness isn't just in the four walls of the clinic. Before someone consumes two liters of soda, we need to be able to engage them and ask them, 'Is this actually a good idea?'

I'm going to share with you one from South Africa because it's really interesting. An HIV chatbot is being used in an area of Johannesburg that has among the highest HIV prevalence in the world, particularly among adolescent girls and young women. The idea was, 'Can you use AI to develop tailored communications to provide safe, reliable, non-judgmental, quick information to young girls who are at risk of HIV acquisition?' Our pilot showed us that, yes, absolutely, you can.

Even though it's a small sample size and we want to be cautious about not generalizing it, we've discovered a few surprising things. The first is that the women who engaged with the bot really appreciated it and wanted to continue engaging [with it] outside the study period. They indicated that they trusted the bot. They disclosed things to the bot they had never disclosed to any human before. They didn't feel the risk of stigma, and because of their disclosures, we were able to get a better assessment of their situations. We're seeing more disclosures from people who may not have disclosed some things at a clinic out of fear the nurse or other person would tell their aunt or their mother, etc.

The pilot participants actually developed a feeling of agency and a better sense of their real risk. Some of them actually asked for self-tests for HIV PrEP and where to get HIV services and treatment. They were saying, 'Oh wow, I'm at high risk. Actually, I need a test today.' Or, 'I'm going to continue to be at risk. I need HIV PrEP.' Now, we don't want to say we've solved this challenge, but any inkling of closing this gap between knowledge and action in health care is worth gold.

I have another example out in Nigeria where another of our grantees is working on a phenomenal wellness chat bot. I decided to mess around with it – I often mess around with these chatbots to try and break them and see how much I can stretch them and so on. I was blown away because I said to it, 'I'm male, this is my weight. My BMI is 37. I consume three liters of soda and two loaves of white bread [a week]. My father died at 48 with type 2 diabetes. At the end of 15 minutes, it had me hooked by the way it set up questions. It said, 'I understand your concern. How would you like me to help you? Do you want to understand your risks? Do you want diet guidance? Do you want somebody to speak to?' I found it so intuitive. By the end of it, I had a workout regime and a diet plan that was tailored for Nigeria's diet. It didn't give me crazy, weird, super expensive things that I wouldn't find anywhere in Nigeria.

The third bucket is how we shorten the arc we discussed earlier, which is the timeline from evidence and new tools to implementation and impact. There are often delays in evidence getting synthesized from normative guidance to country- level policies, and from those policies being turned into SOPs [standard operating procedures] that lead to implementation and actual impact. That arc can take a decade. Then donors ask, 'Why did it take 11 years for that super cool vaccine to get into arms?'

One pilot in this area is trying to help regulators optimize their internal processes. That is going to be super interesting. We have five regulators that have already raised their hands and said, 'We want to use AI to optimize how we do our work because 70% of what we do is formulaic and algorithmic so we're going to start with that.' One concrete thing this pilot has already delivered is taking the massive amounts of evidence that gets generated during normative guidance and developing country-level policies. Just to give you a flavor of that, the WHO releases 15 guidelines a week. Even if you're a super smart and super committed clinician, there's no way you're keeping up with that. That's just at the WHO level, but it all needs to be translated.

One of the groups we've funded, The Health Foundation, has basically demonstrated that you can compile clinical guidelines 25 times faster than humans with a high level of accuracy, and with the ability to see incongruous recommendations across multiple areas. For example, they found that up to 30% of guidelines actually were incongruent. One says, "Give paracetamol." But the other says, "Give an antibiotic." Same country trying to deal with the same clinical presentation but giving very different advice. What we want to do is build out that capability to be able to produce those guidelines because we also know that clinical editorial expertise is really hard to come by. Which is part of the reason there are big delays in getting policies set up.

The fourth bucket – and I will acknowledge that we haven't done enough work on it – is workflow optimization. We know that in even very resource-constrained environments, there are massive amounts of inefficiencies. It's nobody's fault, but when you look at the supply and demand sides, there are major mismatches we could optimize by optimizing workflows.

I'll give you a concrete example. If we look at radiological equipment, it's not easily available in many contexts. Yet the actual utility and uptime of those kinds of instruments is likely to be downwards of 60% in many contexts. Are there ways to rethink that? Yes, absolutely. What we find is that many workflows in health care are 75% waste, and 15% to 20% are value-add. The rest is necessary, but not value-add, like taking your name or other administrative details. Now, with AI, can we start to change those percentages by looking at the process and making recommendations on how to rearrange resources. The fundamental issue is, how do we do more with less? It's a question that many countries, governments, and hospitals are going to face. I think AI offers the potential to synthesize and analyze that information at speed and provide recommendations on how to optimize.

How did the COVID-19 pandemic influence or accelerate your work?

Dr. Brey: Obviously, COVID caused a lot of suffering and had a very negative impact on the world. But from a health system innovation perspective, it gave us an unprecedented opportunity to not only move fast, but scale fast. Most of the innovations I got to work on scaled up and continue to endure and grow.

There was a confluence, I would say, of three things. There was a burning platform – governments wanted to do something to protect their citizens and populations. The second thing was that innovators and technologists wanted to make a difference and were trying all sorts of things. The third is that the evidence-policy-implementation arc that's often tenuous and protracted and usually takes 10 to 12 years, was fundamentally clipped. I had the personal experience of going from crazy idea on Friday afternoon to implementation with 100 sites by Wednesday afternoon and 1,000 sites by the end of the month. The confluence of those three things created an opportunity to drive innovation.

Let me give you an example. I was having a discussion with a very senior policymaker late on a Tuesday night about using data to identify individuals at the highest risk of death from COVID and then proactively reaching out to them and actually bringing them in for care. We knew, for instance that, based on the data, people with diabetes, the elderly and those with comorbidities were dying at a rate of five to 10 times more than the rest of the population. We were just having a discussion about the data. The next morning, the premier [of the country] made an announcement that they were rolling out this initiative and would scale it up over the next week. The health system never does that kind of thing, right? We wait for people to get super sick and then try to fix things. I later said to the policymaker, 'I don't understand what happened here. Usually, you guys go through 16 steps. He said that because the most senior political person had made the announcement, everything would fall into place. People who might usually say, 'We can't, we can't.' didn't have that option anymore. We never have this opportunity, but when we do we have to grab it.