Following the earthquakes that struck Turkey and Syria in 2023, the countries’ health systems experienced widespread structural damage. Hospitals saw whole departments collapse or become unstable due to aftershocks. Health centers and clinics were forced to close at least temporarily due to fractures in roofs and walls.
Mobile health services had to be established throughout the country to treat urgent cases without access to this physical infrastructure in many cases. Assessments were launched relatively quickly to determine the scale of impact and to determine a plan for repairs, but many lower-level facilities, such as pharmacies and dialysis centers, were forced into substantial waits before their places in the priority queue. Delays throughout this process limited the specific understanding of the need for replacement capacity in the health system.
One potential answer to this problem came from above the planet.
Satellite images paired with artificial intelligence models supported humanitarian health assessments in significant new ways. According to findings from a new paper by researchers from Direct Relief and Stanford University published in PLOS Digital Health, these new approaches hold significant promise for rapid and remotely sensed understanding of the impact of disasters on health systems, but many questions still need to be answered in order to make them accurate and fully trusted by response agencies.
Two of the most prominent models in use during the 2023 earthquake response came from Microsoft AI for Good and from Google. Researchers compared how these different AI models determined what buildings were affected, the extent to which rooftop damage could determine the overall damage to the structure, and the quantification of damage using different methods.
While the models agreed in several areas, they also disagreed in significant ways. Particularly when overlaid with the locations of hospitals and health facilities outside the most heavily damaged zones, the models often produced varying estimations of the degree to which facilities may have suffered structural impacts from the earthquakes. Errors in the locations of facilities themselves added an additional layer of complexity.
Andrew Schroeder, Vice President of Research and Analysis at Direct Relief and one of the authors of the paper, said the models differ in how damage is defined. However, the model’s potential is promising, he said, as it could identify population access to emergency medical supplies like pharmacies and doctor offices.
“It’s very impressive in terms of speed and scale. The ability to run a full assessment of impacts on a health system in the wake of a disaster like this might be one of the most important breakthroughs needed in humanitarian health response.”
Schroeder stressed, however, that more work is needed to make these models into core elements of humanitarian analytics. “In the future,” he said, “if we’re able to control for a number of the sources of error and conflicting interpretation that we identified in this research, for instance by creating a common or interoperable calculation for damage extent, we could aim to have a full-scale assessments of the change in the healthcare capacity of an affected area potentially within 42 to 78 hours after an event, without having to put assessors on the ground.” He stressed that error correction is key to building trust in models, without which decision-making in crisis situations cannot use these types of novel approaches.
Speed after disasters saves lives. With some of the improvements outlined in this new research, AI-based health systems damage assessments might dramatically enhance the ability of Direct Relief and many others to minimize the costs to people of the damage to our built environment.
