Natural language processing
During the spring, a troublesome pattern plays out as marine birds along the California coast die from domoic acid poisoning, which is caused by harmful algal blooms. An early clue indicates when and where this problem starts spreading: rescued California brown pelicans, red-throated loons, and other species start turning up at wildlife rehabilitation centers with signs of neurological disease. Yet, though they pepper the state map, these centers are not interconnected enough to nip the issue in the bud. When staffers at one center diagnose a sick bird, others another 40 miles up the road might not be privy to that information.
So researchers at UC Davis recently tested an early detection system that uses artificial intelligence to classify admissions to rehabilitation centers, in the hope of sending wildlife agencies and researchers warnings about growing problems among marine birds and many other kinds of animals. Their system scans intake reports produced at 30 California centers, listing information like the animal’s species, age, reason for admission, and diagnosis. Then the AI uses natural language processing to categorize the reports, looking for patterns in the number of admissions related to certain illnesses and injuries.
The researchers used five years of data and more than 200,000 records to establish baselines for how frequently these conditions normally occur. When the system detects an anomaly—an unusually high number of cases in a given species—it automatically creates an alert, which is delivered to wildlife experts either via the system dashboard, an email, or text message. Because the system processes rehab center admission data in just a day or two, it can produce “prediagnostic” alerts, which are faster than waiting until diagnoses have been confirmed.
In July, the team published a paper describing a test of their system in the journal Proceedings of the Royal Society. “We wanted to use the data in an aggregate form to better help rehabbers to see the bigger picture, other than what they see at their individual centers,” says Devin Dombrowski, president of the Wild Neighbors Database Project and one of the authors of the paper.
During the one-year pilot study, the system identified several patterns that indicated larger problems. An influx of marine birds with neurological symptoms like head twitching and whole body tremors triggered an alert. Upon postmortem examination, these birds, including the black and white water bird species western grebes, were found to have domoic acid poisoning. A few months before, a high rate of clinic admissions in the San Francisco Bay Area for rock pigeons showing symptoms of neurological disease triggered another alert. Further investigation established the parasite Sarcocystis calchasi as the cause.
Study coauthor Terra Kelly, a veterinarian and epidemiologist at UC Davis, compares the system to syndromic surveillance for people, which uses electronic health records to monitor public health concerns, such as flu outbreaks, opioid overdoses, and the spread of Zika virus and Covid-19. She points out that an animal alert system could benefit human health, too. “Wild animals can serve as an early indicator” of diseases like West Nile virus, she says. The disease, which has killed more than 2,000 people since 1999, according to the Centers for Disease Control and Prevention, is often first noticed in sick birds before it’s diagnosed in domestic animals and people.
Additionally, Kelly says, “We could detect the first animal of invasive species that presents to a center in California.” For example, if the numbers of mourning doves admitted to wildlife centers suddenly changed, the system could create an alert that would signal to veterinarians that the Eurasian collared dove had arrived; they’re an invasive species that competes for food and can spread parasites to native doves.
Dombrowski and his wife, Rachel Avilla, are both wildlife rehabilitators who have been working to standardize record-keeping since 2010, when they founded the Wild Neighbors Database Project. They had become frustrated that staffers at rehab centers were using anything from paper records to isolated Excel spreadsheets to track medical records. “When it’s on paper, you’d have to go through boxes and boxes of paperwork to figure out what you gave that golden eagle 10 years ago,” Avilla says.
Together, they started the Wildlife Rehabilitation Medical Database, called WRMD (pronounced “wormed.”) It’s a free online tool that tracks and analyzes wildlife center information like prescriptions and test results. The database has participants at 974 wildlife centers in the US and in 23 countries; the contributors range from major academic institutions to individuals working out of their garages. “We wanted WRMD to be simple so everybody could use it,” Dombrowski says.
Guthrum Purdin, a veterinarian at the California Wildlife Center in Los Angeles County, whose patients range from small house finches all the way up to elephant seals, remembers what paper records were like before the database: “It would be really sketchy, like I can’t really read what someone wrote,” he recalls.
Purdin feels an early detection alert system would be useful. For example, his county recently sent out alerts about Rabbit Hemorrhagic Disease, which is found in both domestic and wild rabbits—but he thinks this kind of system would have let him know sooner. When a new kind of parasite started popping up in pigeons a few years ago, he says, “noticing that there was this unusual upward trend in trichomonas cases could be helpful in directing me toward a different diagnostic approach.”
The UC Davis AI-based alert system is actually built on the bones of this earlier database. (The new system’s name is the Wildlife Morbidity and Mortality Event Alert System.) In the pilot study, the system took in data from just 30 California centers that are already WRMD contributors, but in total, the WRMD database includes over 2 million archival records. It’s the reason why the study authors were able to access records going back five years.
By the end of 2021, the study authors plan to launch an improved version of the alert system, which the California Department of Fish and Wildlife will test. For now, they’re improving the machine learning model through a process called retraining, which runs new data through the existing pipeline to better inform the model. The authors are also working on developing new models that can predict multiple clinical classifications; for example, cases in which both neurological and ocular disease occur together.
Within the next couple of years, once everything is working well, the study authors want to expand beyond California, creating networks to assist wildlife agencies and veterinarians in other states. “The methodology is flexible enough to accommodate different regions and different kinds of animal taxa,” says Pranav Pandit, a postdoc at UC Davis who developed the study’s mathematical models.
Purdin, the Los Angeles County veterinarian, hopes the alert system will help vets head off outbreaks of emerging disease. “Being connected like that,” he says, “you can discover new problems.”