Dr. Sema Sgaier: ‘We see tremendous potential for applying machine learning to the development sector’
Dr. Sema Sgaier is Co-founder and Executive Director of Surgo Foundation, a privately funded ‘action tank’ focused on behavioral drivers impacting health and development. She works at the intersection of behavior, data, and technology to solve complex global development problems. She was a neuroscience and genomics scholar at Harvard Medical School and a fellow with the Centre for Global Health Research. She received her PhD in cellular and molecular biology from New York University and her MA in neuroscience from Brown University. Previously at the Bill & Melinda Gates Foundation, she led large-scale health programs in India and Africa. She is faculty at the Harvard T.H. Chan School of Public Health. She was selected as a Rising Talent by the Women’s Forum for the Economy and Society.
What do you believe the major limitations of human development work are today?
Billions of dollars are spent each year on improving the lives of people with limited resources (i.e. those who are often referred to as ‘living in poverty’) but that money doesn’t always achieve the maximum impact. A crucial piece of information is needed to drive results — deep and comprehensive understanding of what drives people’s behavior and the surrounding ecosystem that influences them. Starting with this deep understanding can transform how development programs are designed and delivered.
Another important limitation that we are seeing is that development programs tend to jump to solutions and focus on scaling specific interventions, yet little is known about the causal chain of factors that influence specific outcomes, including how and why people make decisions about the health and wellbeing of their families.
How is Surgo Foundation trying to change the way that development challenges are approached?
We believe that people with limited resources have choices and should be treated as active customers instead of as passive beneficiaries of development programs. This reframing is incredibly important because it reflects choices that all people have, regardless of where they live or their income. People also live in complex dynamic systems where they are interconnected with others and within a broader ecosystem. Our mission is to catalyze the global development sector by developing solutions that put customers at the center of their programs and address why people behave the way they do accounting for both the human and the systems lens.
We believe that new tools and approaches are needed to do this, and we are investing in developing a toolbox for the sector that collects and analyzes data holistically. This includes combing the best approaches and methods used by other fields and adapting them to the development sector. As part of this toolbox, we are testing various machine learning approaches.
What potential do you see for the application of machine learning to the development sector?
In the private sector, organizations are racing to adopt machine learning to help improve everything from sales to operations. We see tremendous potential for applying machine learning to the development sector, but it will take time — we need to be practical and we really need to test these new approaches to the problems at hand.
Machine learning has the potential to be transformative in both in its ability to help us uncover patterns in the data and develop better predictive models of the world. There are many different types of machine learning and ways that it could be used. For example, causal machine learning models may help us better understand the upstream chain of events that lead to a particular outcome, such as maternal mortality. Predictive machine learning models can help us make better predictions such which women are more at risk for developing complications during delivery. With such information, we can better design and target interventions to reduce the number of women dying at or soon after birth. Classification algorithms can be used to segment into groups based on their behaviors and drivers of their behaviors, which is very useful for more effective targeting of health services. These are just a few broad set of approaches that we are testing; the machine learning field has many more categories and specific tools to offer.
We don’t think that machine learning will completely take over other methodologies such as traditional statistical analysis. In fact, they complement each other. Think of it as adding another lens to our existing toolbox. Different shots and different lens all contribute to the story of the subject. However, there are some very real challenges around applying machine learning to international development data.
Why hasn’t machine learning been adopted before?
Machine learning is a very broad field and encompasses numerous approaches. As a discipline, it has been around for decades, with some approaches (such as neural networks) gaining more traction and attention recently. Broadly, two recent developments have made it possible to apply it in the international development sector. First, high-speed computational power has only recently become affordable at this scale. Second, the vast investments in data by donor organizations and governments and their willingness to openly share their raw data.
What have you learned from your machine learning initiatives thus far?
We are testing a wide range of machine learning tools in our work. As part of this broad exploration, we are specifically testing the application of causal machine learning models through Machine Learning Initiative for Precision Public Health (ML4PxP). In this initiative, we’ve brought together a diverse group of stakeholders to test this platform, including experts in machine learning, such as the University of Sussex and GNS Healthcare, international development implementers, such as the University of Manitoba, and funders, such as the Bill & Melinda Gates Foundation. ML4PxP is still in its early stages, but we have already learned a lot just from bringing together these diverse viewpoints.
First, we learned that you really need to invest time and effort into bringing together the international development experts and machine learning experts. When we first started, no one was speaking the same language. We had to help both sides understand each other in order to collaborate. Second, we learned that data quality matters. Machine learning, while a promising tool, will not be able to provide meaningful insights if you don’t have quality data. Third, it takes time and further innovations in the methodologies test and adapt these methods to development problems and data sets.
What factors could hinder the application of causal machine learning models for precision public health?
The biggest challenge is demonstrating value. Right now, for example, we don’t know if causal machine learning approaches will give us valuable insights over what we have learned from traditional statistics. In order for causal machine learning approaches to be used in the sector, they need to provide valuable new insights to tough problems.
Even if we are able to demonstrate value, there is another challenge of bringing together two separate communities — the international development community and the machine learning community — to stimulate adoption.
How are you planning to address these issues?
We are putting our own dollars forward to test the approach with the goal of being very open with what we find. Once we have a few use cases that show value in applying causal machine learning approaches, we will move from testing to scaling. With our deep experience in international development and our immersion into causal machine learning through this pilot, we are in a unique position to bring together the two communities to stimulate adoption.
What do you hope to have achieved in five years’ time with your machine learning investments?
With technology advancing as rapidly as it is, it is tough to say exactly. If we can demonstrate the value of applying causal machine learning models to international development data, we’d like to serve as a conduit to stimulate adoption. In five years, we would hope to see causal machine learning being scaled across the international development sector to help inform the design and delivery of more effective programs.
