How Microsoft runs its $40M ‘AI for Health’ initiative

Last week, Microsoft announced the latest news in its ongoing “AI for Good” program: a $40M effort to apply data science and AI to difficult and comparatively under-studied conditions like tuberculosis, SIDS and leprosy. How does one responsibly parachute into such complex ecosystems as a tech provider, and what is the process for vetting recipients of the company’s funds and services?

Tasked with administrating this philanthropic endeavor is John Kahan, chief data analytics officer and AI lead in the AI for Good program. I spoke with him shortly after the announcement to better understand his and Microsoft’s approach to entering areas where they have never tread as a company and where the opportunity lies for both them and their new partners.

Kahan, a Microsoft veteran of many years, is helping to define the program as it develops, he explained at the start of our interview.

John Kahan: About a year ago, they announced my role in conjunction with expanding AI for Good from being really a grants-oriented program, where we gave money away, to a program where we use data science to help literally infuse AI and data to drive change around the world. It is 100% philanthropic — we don’t do anything that’s commercial-related.

TechCrunch: This kind of research is still a very broad field, though. How do you decide what constitutes a worthwhile investment of resources?

AI for Health is a prime example of how we can create change where people aren’t investing, or they can’t invest. I say aren’t because there’s no ROI. That’s why you see us working on things like leprosy and SIDS, because there are areas that, you know, frankly, the world has forgotten.

So we will invest in areas where there are gaps in investment around the world. We look for areas where we believe we can have impact in the world. And my mission, by the way, expands beyond AI for good, like we work on sustainability, we work on affordable housing, broadband, areas where that we believe as a company, we have an ability to have an impact in the world.

You say that there’s no ROI, and of course there isn’t one short-term for something like leprosy research, but at the same time you have to have a metric to determine how to dedicate your time. It’s still results-based. What’s the process there?

It’s a funnel, like anything else. We have a set strategy, particularly in AI for Health, where we work in three main areas. Quest for discovery, which is accelerating research in prevention and diagnosis of diseases. Global health, which is about ensuring that we have a shared understanding of mortality and longevity for the purposes of preventing the next global health crisis. Third is equity, ensuring that we help serve underserved populations.

So stage one is, does it fit within the strategy? Because there’s many, many things, not just in health, but in other areas that are outside the strategy. We went through a lot of soul searching of where we think we can make a difference, and those were the three key areas.

But within that, there has to be some decision-making to make sure we don’t spend our resources in the wrong way, because there’s way more demand than we could ever fulfill.

The second bar is we have to have data at scale, otherwise AI is worthless, you know? We look for areas where they have data available for us to work on. Sometimes we’ll take projects that don’t, if we think we have a reasonable time frame to collaborate and collect that data. Leprosy is a prime example; we had to adjust the protocols, but the Novartis Foundation, our partner there, had the ability to collect data at scale.

The third thing is the partnerships must have experts. We are not by any means experts in these medical fields. I have people on my team that are physicians and also data scientists. I have people [who are] world-class data scientists in genomics, they have PhDs in that — but that’s not our role, we just happen to have that extra expertise. We are not the medical experts. The medical experts have to be on the other end.

The fourth bar is, can the operation scale on its own when it’s done? Because we can get it to production, but we don’t own the production. We work towards nonprofits as much as we can. We will partner with academia, but only if we know that eventually there’s a place for it to land in the market. The beauty of working with Path and Novartis Foundation is they have people on the ground they can deploy, you know, they can make change directly. You can publish papers, but at the end of the day, my job is literally to make changes in the world.

The last piece of the equation is that they need to be self-sustainable when we’re done. I have a set of program managers that are technical experts on this process. And they’re sorting through every one of these initiatives.

Seems like it can be difficult to judge. But it’s evidence-based in the end.

The thing we don’t do, which I want to make clear, is we don’t — like, I think the Gates Foundation does this, they assign a financial measure to the thing. Maybe someday we’ll get more sophisticated at that. At the end of the day, you can can boil things down to numbers.

I remember, well before I got into any of this, many many years ago, I had a breakfast meeting with Bill Gates and he explained how he does things. It’s very number-driven, like I put one dollar in here and I can save five lives there. I don’t know whether his philosophy has changed over 17 years, but it was very eye-opening.

Like, how do you invest in charity yourself? You do it from your heart, probably, which is what I do. But as a businessperson, you can’t do that. You have to have some kind of criteria on the way you do things. So we have to be selective in what we do. We have to choose partners that are sustainable and can really make an impact in the world.

In your post, you mentioned how most of the AI professionals in the world are in the tech sector proper and only a small fraction are in the healthcare and nonprofit world.

Some of these organizations, it’s very hard for them to hire and retain at scale. The beauty of what we do for a living is that of course, there will be skills transfer. That isn’t our objective, but it happens naturally.

At Seattle Children’s, where we started several years ago, they didn’t speak our language, we didn’t speak their language. But we all sat down and realized the value we were each providing. Over time, they started to build the skills within their own environments. And vice versa — of course, we get more subject matter knowledge as well.

Is helping them retain talent part of the goal of the program?

That’s not really the goal. So with Children’s, our data scientists were in volunteer mode, they went over there and they realized that there was data available, the researchers knew it was there but had no knowledge of how to use it. This was large-scale data, a 24-page document explaining it, you had to write code to get access to it. From a researcher perspective, they knew it was there but they couldn’t access it.

Now that data is in the cloud, a Bayesian network is on top of it, our AI is on top of that to get meaning out of it, and now there are published papers from it.

Now let’s assume we walked away and we’re completely done — though we rarely get done, we try to keep these going over long periods of time. But we walk away, at this point they can run the models, they can enhance the models, they can do it on their own and use it in the regular research they do. They’re continuing to do their day jobs. The data just exists in ways that everyone can learn from.

Microsoft has the privilege to be able to dedicate $40M to these causes, but I imagine smaller companies with less cash on hand would like to help along these lines independently. What can startups and small businesses do to make AI more accessible and applicable to global health and social good goals?

There are several things that any company can do — first they’ve got to recognize that open sourcing of data and models, the right way of course, privacy compliant and all that, is good for the world. That’s why we use GitHub for everything. You have to recognize that for anyone to succeed here it’s about sharing and connecting things together. Siloed data and models will not solve these problems, it’ll make it worse. If you start protecting the pieces, you won’t learn from each other.

We want to open source as much as possible, we don’t want to own anything. And we only publish when the medical experts say so. Every agreement we do with a partner, we will make these models available for others to use — they understand that going in. There’s no proprietary IP; we want the world to iterate on it, because we’re not the only ones getting better at this.

It seems like this is the time to make moves, since AI tools are expanding in so many other ways, it would be a shame for these areas to get left behind.

It’s a personal mission. I wouldn’t be doing this if it wasn’t for my son. No one should ever have to suffer the loss of a child again. So unless they fire me I’m going to be here until the day I’m gone. By getting awareness out on these issues things happen much faster — you see donors and resources come to the table. I see rapid change occurring not just in SIDS, but in affordable housing, and sustainability, and broadband, on and on, by using a small number of data scientists and the platform of Microsoft to make change in the world. We’re living in a time when we can do that.

So, no one should be afraid to share capital and resources for the good of mankind. There’s no secret sauce, but I think there’s a recipe we can all follow.