Skip to main content

An AI model shows health centers in California aren’t prepared for power outages

An AI model shows health centers in California aren’t prepared for power outages


Most don’t have backup power sources

Share this story

Sylmar Converter Station, a piece of power grid infrastructure operated by LADWP, for a story about how an Oregon wildfire almost caused the lights to go out in California on Friday night. Sylmar Converter Station is the terminus of the Pacific DC Intertie
Wildfires in the West can lead to power outages.

There are around 20 federally qualified health centers in the areas touched by the Lava, Tennant, and Salt fires, which started in Northern California at the end of June. The centers are part of the state’s healthcare safety net and offer primary care to low-income patients in the area. Despite the fire risk to the facilities, none of those 20 centers had backup generators, says Andrew Schroeder, who runs analytics programs at the humanitarian aid organization Direct Relief. 

That’s key information for officials and aid groups: if the power went out because of the fires, those centers might not be able to keep medications and vaccines refrigerated or turn on computer systems with patient medical records. But before May, that’s not information Schroeder or anyone else working on response efforts would be able to look up. Despite the threat of outages and shutoffs during extreme heat and fire season, states and industry associations often don’t know which facilities have backups. 

“We had a lack of information about where health centers have generators, where they have solar panels, have people installed battery backups for refrigerators, and things like that,” Schroeder says.

Now, thanks to a new AI model finished at the end of May, they have a better picture of the backup plans at these health centers just in time for what’s projected to be a devastating fire season. It shows a sparse landscape: only around 30 percent of facilities have backup power and access to refrigeration for medical supplies, says Jeni Stockman, senior program manager at artificial intelligence company Macro-Eyes, which built the tool. “It suggests there’s a huge gap in the resiliency of sites,” she says

“There’s a huge gap in the resiliency of sites”

The information is acutely important for the response to the COVID-19 pandemic because the vaccines for the disease have to be stored at cold and ultracold temperatures. Sending shots to a place that couldn’t keep them cold during a power disruption risked losing the doses. So Macro-Eyes worked with the California Primary Care Association (CPCA) on a project funded by Direct Relief to build a tool that would map out the resources at each center. The groups soon realized it could have other applications outside of COVID-19, like where to direct aid in other emergencies, says CPCA Acting President Robert Beaudry.

“Because of the fires, and because of [power company] PG&E’s ongoing problems, we have power outages in California. We wanted to know which of our health centers could continue to operate,” Beaudry says. “The more information that we had around these types of issues, the more we could inform emergency preparedness in California.” 

The project was modeled off of work Macro-Eyes did in Sierra Leone mapping out what resources health facilities had at any given time. The government didn’t have good information about which health centers had access to electricity and which didn’t, says Benjamin Fels, the founder and CEO of Macro-Eyes. “In the context of COVID-19, when you get these really scarce vaccines, you don’t want to send them to a point on the map where they cannot be kept cold — because that basically means you’re throwing them away,” he says. 

“When you get these really scarce vaccines, you don’t want to send them to a point on the map where they cannot be kept cold”

In conversations with Schroeder and Direct Relief, the company realized people in California didn’t have that information, either. Hospitals in the US that get federal funding are required to have backup generators. But outpatient and community care clinics, including federally qualified health centers, often don’t. 

The facilities can charge Medicare and Medicaid for services, so they have to report certain information to the federal government — demographic information about the patients they serve, insurance status of those patients, and so on. It’s a very extensive form, Schroeder says. “What it doesn’t contain is much of any information about the actual infrastructure of the places,” he says. 

The CPCA has surveyed centers to find out what resources and power sources they had available, but those haven’t dredged up much information. Beaudry says the association had a sense of the power resources for around 8 percent of the sites in the state. 

Macro-Eyes’ model helped fill in the rest of the picture. It pulls in data from a range of publicly available sources — like lists of places that had permits for generators and satellite images for solar panels. From there, it’s able to infer which similar types of health sites might have the same features. If, for example, the model already has information from survey data on three health centers out of 10 that have similar locations and serve a similar number of people, it could project that the remaining seven would have the same resources as the first three. Health centers can then confirm whether the model got it right, which helps the AI continue to learn. 

Mapping out the data helped CPCA see where there are limitations in access to power. Officials and aid groups could use that information to decide where they should send supplies. That could include anything from COVID-19 vaccines (which need to be frozen) to insulin, which has to stay refrigerated, Schroeder says. “We often support needs for insulin after disasters,” he says. “We need to know where we’d feel confident sending things that require cold storage, without also sending some kind of ability to back up their refrigeration capacity.”

“We need to know where we’d feel confident sending things that require cold storage”

The data could also help organizations figure out which places in the state could use extra generators. “We can see where there’d be islands of rural health populations that would be potentially disconnected,” says DeeAnne McCallin, the director of health information technology at CPCA. She says the California Department of Public Health recently told CPCA that they may have extra freezers or refrigerators available and asked if there were any health centers that could use them. That’s the kind of inquiry the data could help answer.   

It’s also valuable for the health centers themselves because they’re able to see during an emergency if there’s a nearby facility with access to a generator or a refrigerator that they could direct patients to or use to store supplies. 

The Macro-Eyes modeling tool can be used in places other than just California. The company and Direct Relief think that the Gulf Coast, which has been battered by increasingly devastating hurricanes, could be a good next target. 

Major disasters — like those Gulf Coast storms, Hurricane Maria, and the California fires — have made it clear that having a backup power plan is a key piece of preparedness. That’s especially true for healthcare, which is vital during disaster response. “We’ve designed a health care system which is really in some ways very sophisticated, and very unequal, but really fragile when it comes to access to power,” Schroeder says.