In regions upended by conflict and climate threats, how do you devise health systems and strategies for groups that don’t access formal health systems or show up in existing data?
It’s a challenge that Sonia Navani, an implementation scientist and Blum Senior Fellow, has been working on for decades with refugee and displaced populations.
Through roles with the International Rescue Committee, UNICEF, UNFPA, and Columbia University School of Public Health, Navani spent years building health data systems in conflict and disaster settings. She describes a pattern that kept surfacing across the more than 15 countries she’s worked in: The groups facing the steepest health risks were also the ones least visible in the data systems.
That invisibility wasn’t random. It largely tracked with what she calls “more granular gradients of poverty — who could reach a clinic early, who showed up late, and who never showed up at all.” As digital health and AI-enabled tools started showing up in some of these settings, Navani says many of them inherited those same blind spots.
But what shifted her thinking wasn’t just who was missing from the data. It was what she was seeing climate change do. In the communities she works in, floods and extreme heat aren’t just making food harder to get — they’re changing what people eat, when they can get it, and how often those disruptions hit.
Navani describes this as going beyond the familiar picture of food insecurity. It’s actively reshaping dietary patterns in communities already stuck in poverty, and driving chronic disease risks that didn’t exist in the same form a generation ago. The health tools being extended to these settings, she says, were designed for more stable situations where diets and environments don’t shift seasonally.
What struck her wasn’t just that the problems were getting worse. It’s that these disruptions are forcing changes in how communities live, eat, and cope — and those changes are visible and trackable, and in some cases connected to things that are low-cost and doable without relying heavily on outside aid.
Navani’s current work focuses on one area where she’s building this out — food-based nutrition for women in early pregnancy — but she says the problem is much bigger than nutrition.
“Communities in poverty are generating observable signals, particularly related to long-term chronic disease risks such as diabetes and asthma, but most existing tools are a mismatch — preconditioned on stability,” she said. “The signals don’t translate into contexts that are anything but stable, and where different aspects of that poverty contribute to whether disease risks compound or not.”
Her response has been to try to build something that goes beyond fixing data shortages. She wants to connect what’s coming out of climate science and biology so that the tools are looking for earlier signs of disease risk — but designed for unstable conditions instead of stable ones. She says this is getting more urgent in communities where health systems aren’t set up for chronic disease prevention, let alone treatment.
A framework she has developed for this purpose she calls Precision Community Health. She describes it as “a translational approach that applies representational learning to identify when deprivation shapes disease risk at a community scale — providing pattern-level profiles for chronic disease mitigation where epigenetic mechanisms are a major influence on disease risk — alongside individual genetic risks.”
So what does that actually look like on the ground? For Navani, putting this into practice meant building a whole process — not just the AI part, but how you get useful evidence out of communities that formal research doesn’t reach, how you connect that to science that works under unstable conditions, and how the results get back to health workers in a form they can actually use.
She founded DHDI, a nonprofit research group, to explore these areas with local civil society partners, from generating data ethically with communities, to figuring out where the real traction points are for preventing disease under climate stress and poverty, to getting support back to women and health workers in ways they can act on.
For Navani and her team, the community comes before any data generation or code. They developed Climate Health Photovox, a 26-session curriculum running with local partners in eastern DR Congo. It’s a digital storytelling program where women learn not just how but why climate events are reshaping local agriculture and food, and what that means for family health.
Using shared smartphones, women document those shifts themselves: photographing local foods, tagging them with local names and seasonal details. Navani describes it as building a visual record of what’s happening through the eyes of the very groups that existing systems tend to miss.
She notes that women who may not own phones but regularly use them within households become participants in shaping both the questions and the evidence — “not as data subjects, but as co-investigators,” as she puts it. The curriculum was designed to stand on its own. Women come away with food and nutrition strategies, smartphone skills, and connections to other communities dealing with similar climate challenges. But the images they share also feed into the AI work, and Navani says that dual purpose is built in from the start.
Of the different health problems where this kind of community-generated data could be put to use, Navani started with nutrition for women in preconception and early pregnancy. She calls it the area where the signals are most clear, the interventions are affordable, and there’s a specific biological window where acting early matters most.
The food images and the image meta-data feed into EpiNu, a proof-of-concept AI system that Navani says was built for the conditions where most tools fall apart — offline, low connectivity, no text input, no self-reporting. The system runs on a micronutrient database her team built specifically for preconception and pregnancy.
Navani explains that standard nutrition data often doesn’t reflect what people actually absorb from food — not just what’s in it. She gives the example of iron in beans: Guidelines list the iron content, but compounds in the food can block most of it from being absorbed.
EpiNu tries to account for that gap. The system is aimed at community health workers, giving them guidance based on what’s locally available rather than general recommendations that Navani says rarely match reality for women dealing with both poverty and climate disruptions.
The focus on early pregnancy, she says, is where the science and the urgency come together. There’s a narrow window early on where what a woman eats can have lasting effects.
“Many women don’t access clinical care until well after those early windows have closed,” Navani said. “In DR Congo, this is showing up in the data — excessive neural tube defects, a largely preventable problem where folate is a major factor. Supplements remain central, but many women receive them too late, if at all. The question is what can be done with food already in the household to unlock more of what these women need.”
The urgency behind this work isn’t abstract for Navani. Darfur remains the clearest example.
In 2004, she helped set up data systems in displacement camps in South Darfur and returned over the following years to support local teams, only to find the same women in the same makeshift shelters under the same conditions.
“They would remember me, which was both amazing and horrifying,” she recalled. “Horrifying because it meant they were still there — the same people, in the same tents made of plastic sheeting. The only thing that changed was the international actors around the table.”
That lesson has only gotten sharper. Darfur is in extreme crisis again.
“If the architectures of aid systems in place for decades have reached their limits,” she said, “then the question is not how to rebuild them in their original form for a world that doesn’t exist anymore, but how to rethink them entirely — and whether the computational tools now available might, for the first time, make that possible.”
At the Blum Center, Navani works with students in computer science, data science, computational social science, developmental engineering, and other fields. She says the projects push students to deal with constraints they don’t usually face.
“Most computational models are built in and for stable, well-resourced environments, then treated as universally applicable,” she said. “Working on EpiNu pushes students to rethink those assumptions — what counts as sufficient data, what ethical data collection looks like under conditions of partial access and shared devices, and what scalability actually means when infrastructure cannot be taken for granted.”
Looking ahead, Navani and her team are expanding their work in eastern DR Congo with larger networks of local partners. The bet, she says, is that building for the hardest conditions first actually makes scaling easier, not harder.
“If you design for eastern DRC and it works, moving into higher-infrastructure settings for communities experiencing poverty is a much shorter leap than trying to go the other direction,” she said. “Most people assume you build for stability and adapt down. We think it’s the other way around.”
