Mar 5, 2020|
Schistosomiasis is one of the most important of the so-called neglected tropical diseases (or NTDs), meaning it draws relatively few first-world researcher dollars. That’s partly because it is battled largely on the ground in remote, poverty-stricken locations of sub-Saharan Africa. Being able to address such a disease remotely and at scale could be a big boon to public health officials. With a seed grant from the Stanford Institute for Human-Centered Artificial Intelligence (HAI), De Leo, Suzanne Sokolow, executive director of the Stanford Program for Disease Ecology, Health and the Environment, and an interdisciplinary team of Stanford researchers has now offered a proof of concept for doing just that.
Traditional efforts to understand and predict the location of schistosomiasis snail populations involve taking samples at villagers’ water access points. But sometimes researchers find no snails even near villages with 99% infection rates. To better understand why, De Leo asked Andy Chamberlin, a member of his lab, to gather drone images for a several square kilometer area. “It was like opening a new window and showing this incredibly complex landscape of vegetation in the water,” De Leo says.
From the drone footage, the team could see large patches of floating vegetation where snails live just beyond the cattails and other tall, emergent vegetation near the shore where scientists typically did their sampling. And when they used drones to map that floating aquatic vegetation, De Leo’s team found a clear statistical association between the area and percent of the landscape covered by the vegetation and the presence of disease in the local community. “We can now say with reasonable confidence that when we see this type of vegetation, we will very likely find the snails that amplify schistosomiasis,” De Leo says.
This conclusion pointed De Leo’s team toward doing larger-scale mapping of aquatic vegetation using not only drone images, but satellite images as well, which they obtained from DigitalGlobe.
Because Andy Chamberlin, a member of De Leo’s lab, had spent so much time both sampling snails in the water and looking at drone imagery he had become quite good at differentiating snail habitat in drone and satellite images. “But we can’t ask Andy to spend his life classifying imagery on the screen,” De Leo says. So instead, Chamberlin and a Stanford undergraduate hand labeled drone imagery and satellite images collected for a 16 square kilometer area during the same period of time. They then used those images to train an AI image recognition tool to replicate Andy’s hard-earned wisdom. The result: The algorithm identified the floating vegetation with an accuracy of about 85 percent, says Zac Yung-Chun Liu, the research technician in De Leo’s lab who led the preliminary machine-learning analyses.
To take this work a step further and identify disease hotspots as they develop, De Leo, Sokolow and John Bauer, an HAI fellow at Stanford, still need to overlay its maps with information about human settlements and activity, since both snails and humans must be present for the parasite to survive.
The team sees these results as a successful proof of concept. “Once we’ve trained the AI algorithm with further imagery and have ground truthed machine learning predictions to assess to what extent we can trust them, we think we can be ready to deploy it at a large scale,” De Leo says. Then, he says, “instead of sampling every site locally, we will be able to estimate the extent of snail habitat in any given area and any given year.”
In the long run, he hopes the maps could help eradicate the disease by guiding the appropriate use of environmental controls such as clearing the vegetation that serves as snail habitat. If the approach proves scalable, it could benefit millions of people worldwide.