Cases that would have sat in a queue for twenty hours were now surfaced immediately
In the winter of 2020, as the pandemic's paperwork threatened to overwhelm its medicine, researchers at Stanford University turned to artificial intelligence not to cure the virus but to rescue the hours lost between a positive test and a human response. Working with a local health department drowning in a thousand handwritten faxes a day, they built a system that could read, classify, and prioritize case reports in minutes — collapsing an eleven-day delay into something closer to the speed of contagion itself. It was a reminder that in a crisis, the unglamorous work of moving information quickly can be as consequential as any treatment.
- American public health departments were processing COVID-19 faxes so slowly that the average gap between a positive test and official tracking stretched to eleven days — long after the window for intervention had closed.
- In Austin alone, a thousand faxes arrived daily, many handwritten and incomplete, creating a backlog that left patients unisolated and contacts untraced while the virus moved freely.
- Stanford researchers deployed a two-stage neural network that scanned incoming faxes, recognized standard case report forms, extracted symptom and risk data, and generated priority scores — all in real time.
- Tested over fourteen days on a live fax line, the system processed over 1,200 documents with 98% precision in identifying case forms and flagged urgent cases more than twenty hours faster than manual review.
- Early pilot results confirmed the system could surface high-acuity cases that would otherwise have languished in the queue, giving public health workers a genuine chance to act before transmission chains extended further.
By late 2020, with COVID-19 cases surpassing 76 million worldwide, the United States faced a crisis within the crisis: the paperwork couldn't keep up with the virus. In Austin, Texas, the local COVID-19 office was receiving roughly a thousand faxes a day — case reports, test results, contact information, much of it handwritten, duplicated, or incomplete. On average, it took eleven days for a single result to move from fax machine to official tracking system. By then, the chance to isolate a patient or trace their contacts had largely passed.
Researchers at Stanford University and Contra Costa Health Services built an AI system to address this directly. It worked in two stages: a neural network first identified which incoming faxes contained standard health department case report forms, then extracted checkbox data on symptoms and risk factors to generate a priority score, surfacing the most urgent cases immediately.
Tested over fourteen days on a real fax line, the system processed 1,224 documents and performed with striking accuracy — 98% precision in identifying true case reports, and 87% specificity in flagging high-priority cases, catching 49 of 59 that genuinely warranted urgent attention. Cases that would have waited more than twenty hours in a manual queue were now surfaced in minutes.
The researchers were candid about the work's early stage: the system had been trained on only five report templates, and broader deployment would require recognizing new form types and improving data extraction. But the essential proof was there. The technology wasn't meant to replace public health workers — it was meant to give them back the hours they needed to do what only humans can do: call patients, trace contacts, coordinate care. In a pandemic, those hours are everything.
By late 2020, as COVID-19 cases climbed past 76 million worldwide, public health departments across the United States faced a grinding logistical crisis. The virus itself moved fast. The paperwork did not. In Austin, Texas alone, the COVID-19 office was receiving roughly 1,000 faxes per day—case reports, test results, contact information, all arriving on paper, many handwritten, many duplicated or incomplete. It took eleven days on average for a single test result to make its way from a fax machine into the official tracking system. By then, the window for isolating a patient and tracing their contacts had largely closed.
This was the problem that researchers at Stanford University and Contra Costa Health Services set out to solve. They built an artificial intelligence system designed to do something that sounds mundane but was, in the context of a pandemic, genuinely urgent: read handwritten faxes in real time, identify which ones were COVID-19 case reports, and flag the most serious cases for immediate attention.
The system worked in two stages. First, a neural network scanned incoming faxes and classified each page, determining whether it contained one of the standard Confidential Morbidity Report forms used by health departments. If it found a match, the system registered the page against known templates and extracted the checkbox data—which symptoms were marked, which risk factors were noted. From this metadata, the system generated a priority score and flagged high-acuity cases for the front of the queue.
The researchers tested the system on a real fax line at a local health department over fourteen days. The system processed 1,224 incoming faxes and proved remarkably accurate. It achieved 98 percent precision in identifying actual case report forms—meaning almost none of the documents it flagged as COVID reports were false alarms. When it came to spotting high-priority cases, it maintained 87 percent specificity, catching 49 out of 59 cases that truly warranted urgent attention. More importantly, it did this in minutes rather than days. Cases that would have sat in a manual review queue for more than twenty hours were now surfaced immediately, giving public health staff a genuine chance to act.
The team began a pilot deployment at a single health department, and the initial results were striking. The system successfully alerted staff to several high-acuity cases that would have otherwise languished in the backlog. The researchers acknowledged the work was preliminary—their system had been trained on five different report templates, and they knew that scaling it would require the ability to recognize new forms and extract data more reliably. But the core insight was proven: automation could collapse the delay between a positive test and a public health response from days to hours.
What made this work matter was not the technology itself but what the technology enabled. Every hour a case report sat unprocessed was an hour a patient remained untested, unaware, potentially spreading the virus. Every day a contact went untraced was another day the virus had to move. The researchers understood this. Their system was not built to replace public health workers but to give those workers back the time they needed to do their actual job—to call patients, to trace contacts, to coordinate isolation. In a pandemic, that time is the scarcest resource of all.
Notable Quotes
The system successfully alerted the team to several high-acuity cases, which would not have otherwise been processed until over 20 hours later— Stanford research team
The Hearth Conversation Another angle on the story
Why faxes? It's 2020. Why are health departments still using fax machines to report COVID cases?
Because that's what the system was built on. Fax is old, but it's reliable, it's secure, and it's what many local health departments have been using for decades. You can't just flip a switch and move everyone to a new system in the middle of a pandemic.
So the bottleneck isn't the fax itself—it's the manual work of reading and sorting them?
Exactly. A thousand faxes a day, many handwritten, many incomplete. Someone has to look at each one, figure out if it's a COVID report, extract the relevant information, decide how urgent it is. That takes time. And time is what you don't have when cases are doubling.
The system caught 49 out of 59 high-priority cases. What about the ten it missed?
That's the real question. Those ten cases would have been flagged eventually, but not for hours or days. The system isn't perfect, but it's fast enough that it changes the outcome. It's the difference between isolation happening today and isolation happening next week.
What happens to the cases the system flags? Does a human still review them?
Yes. The system doesn't make the final call. It just moves the urgent cases to the front of the line so a human can act on them immediately. It's triage, not diagnosis.
And the false positives—cases it flagged as high-priority that weren't?
The system had limited false positives, which matters because false alarms waste staff time. But the real measure is what it caught and how fast. If you're going to err, you err on the side of caution in a pandemic.