Natural language processing
Alison Stribling has learned a lot about infectious disease since she transferred onto Covid-19 response at the health department in Contra Costa County near San Francisco. One of her discoveries: How vital fax machines are to US pandemic response.
Across the country, labs and health providers report new Covid-19 cases to local health departments. At Contra Costa Health Services, officials use the data to start contact tracing or send extra help in certain cases, such as at a care home or to an infected health care worker.
On a typical day in Contra Costa, only around half of those reports arrive electronically; the rest, as many as hundreds, flow in via the fax line, creating a Sisyphean reading list. “It can be a very long day, especially during surges,” says Stribling, a public health program specialist. “It’s that feeling of ‘I can never get on top of this.’”
Read all of our coronavirus coverage here.
Now, Contra Costa’s fax first responders have a little high-tech help. Just before Thanksgiving, the department booted up software called Covid Fast Fax, developed in hurried collaboration with Stanford University researchers. It flags the most urgent new faxes using machine-learning algorithms. When Stribling and others in the fellowship of the fax returned to work after the holiday, they had a backlog of hundreds of faxes to read—but at least knew where to start. “That was great timing,” Stribling says.
Like much else about US pandemic response, the project highlights the creakiness of the country’s health system. It’s also another example of creative minds patching it up with hasty innovation, after skilled auto workers made face shields, or homemade hand sanitizer. In 2020, such projects can be lifesavers. Contra Costa’s Stanford collaborators have now released their code and methodology for other researchers or health departments to use.
Contra Costa got its AI helper after Amit Kaushal, a Stanford professor and practicing physician who works on integrating machine learning into health care, offered the department his skills this spring. Kaushal suggested collaborating on a grant he’d received to work on curbing the virus’s spread with a contact-tracing app using Bluetooth signals. Officials got more excited when he tossed out the notion of an AI-enhanced fax line.
Health officials in Contra Costa were struggling with more than just the volume of incoming case reports. Faxes appear as PDFs on a server, not piles of paper—this is after all the 21st century. But it’s tricky to spot and assess a Covid-19 case at a glance. Cases can be reported on several different forms, which are also used for other diseases, are often scrawled by hand not typed, and sometimes arrive in a jumble of other messages or records. On a usual day, two public health specialists would be assigned to read through and prioritize incoming faxes. “Very few faxes are the same, and it takes a lot of attention to detail and training to know what you’re looking at,” says Stribling, who for a time led the team handling incoming case data. “That can be hard to do for eight hours or more.”
Kaushal and fellow Stanford researchers aimed to tame the problem using machine-learning software that analyzes images—technology more commonly aimed by medical researchers at tumors, not faxes.
To avoid handling sensitive medical data, Kaushal recruited some fellow physicians to fill out disease reporting forms with randomly generated patient data in authentically doctorly scrawls. The phony forms were sent to a fax line to create authentic-looking sample data. Grad student Adam Lavertu used that data to train software to classify whether a page of an incoming fax contains a new Covid-19 case report or is something else, like a medical record or report of tuberculosis.
Results were good. But the team hit a snag when it tried to build a second AI model to automatically transcribe all the data from those pages. “We were pretty good but not good enough,” Kaushal says. “Doctors’ handwriting broke our AI system.”
The Stanford researchers refocused on a humbler task: deciding whether a new Covid-19 case report needs urgent attention. It looks at check boxes on forms identified as reporting a Covid-19 case to determine whether the case appears to demand an urgent response, such as someone in a care home or a health worker. In a test using about 1,000 real faxes sent to Contra Costa, the model correctly identified high-priority cases 83 percent of the time, a level considered good enough for a real-world test.
The researchers worked closely with Stribling and other health staff to integrate the software into their workflow but didn’t have time to design a pretty user interface. Instead the software renames each PDF file with one of five priority tags—such as “01_hcw” for health care worker—allowing county staff to see the most urgent cases quickly by sorting the list.
The health department first put Covid Fast Fax to work the day before Thanksgiving—the first day since the pandemic began that staff took a break from monitoring new case reports. They returned two days later to more than 400 faxes waiting to be read, more than a day’s work. “It was a little scary to see,” says Stribling, who recently transferred to work on Contra Costa’s vaccine deployment. “But it was very helpful to have something in place to focus staff’s time on the most important things in that folder.” Staff still process every fax received.
Here's all the WIRED coverage in one place, from how to keep your children entertained to how this outbreak is affecting the economy.
By Eve Sneider
Kaushal and his fellow researchers are pleased with their pandemic creation, even though it’s more hacky than the usual Stanford AI project. “If we were not in a pandemic, no one in their right mind would say let’s figure out some artificial intelligence to extract information from faxes,” Kaushal says. Lavertu, who put aside work using AI to spot online discussions of drug side effects, says the project is a reminder that you don’t need new, breakthrough algorithms to improve health care. “We spend a lot of time focused on how we can make AI better and better,” he says. “This project demonstrated how AI can already help in its current form.”
Although conceived for 2020’s singular circumstances, Covid Fast Fax or something like it might one day find uses beyond the pandemic. Fax machines are expected to have a central place in the US health care system for years to come. “It’s the ultimate flexible tool,” says Julia Adler-Milstein, director of University of California San Francisco’s Center for Clinical Informatics and Improvement Research.
Despite billions spent by the federal government to encourage health providers to process more data electronically, moving medical info between institutions still often falls back to fax. Incentives for digitizing have been mostly aimed at hospitals and doctors, not the public health departments and long-term care facilities now on the Covid-19 front lines. “We’ve paved certain roads, but when patients spend the rest of their time on dirt roads, that limits the value of the roads you have improved,” Adler-Milstein says.
The Trump administration took a late swing at the fax machine earlier this month. The Centers for Medicare and Medicaid Services proposed requiring state Medicaid plans and some private insurers to create APIs to move patient data faster, and launched a consultation asking for feedback on what would happen if faxes were eliminated. “It’s not acceptable that our health information capabilities remain mired in the Stone Age,” CMS administrator Seema Verma wrote in a blog post.
That proposal would take effect in 2023. For now, fax machines are unlikely to be demoted from their central role in the response to the pandemic. Cases in Contra Costa County have hit new highs in recent weeks; staff will likely have more faxes than ever to process on Saturday, after taking half of Christmas Eve and all of Christmas Day off. But they’ll know which ones to read first.