When Melissa Cantor was raising calves at the University of Kentucky as an undergrad, she realized there was a need to identify animals that were getting sick. Cantor also speculated that machine learning might be a tool to help detect illness prior to animals showing clinical signs.

Today, Cantor is assistant professor of precision dairy science at Penn State. She noted that with farms switching to social housing for calves and using robotic feeding, it’s sometimes more difficult to identify sick animals. Her question became “Are artificial intelligence (AI) algorithms or are humans finding sick calves?”

Interested in alternatives to antibiotics to treat animal disease, Cantor said AI provides a unique opportunity to act before animals are actually sick.

While many calves are raised individually, such calves are not as adaptive as those raised in groups. Individually raised calves are less competitive, more aggressive and more fearful in new experiences compared to calves raised with others. Numerous studies show that these lone calves have no performance benefits over calves raised in pairs.

This brings up the concept of social license to operate, an aspect of husbandry that’s often overlooked. “We need to make sure consumers feel good about buying dairy products, which means it has to be socially acceptable,” said Cantor. “We should always be thinking about whether the consumer would accept our practices.”

Dairy producers are aware that milk intake is related to disease, but as calf housing shifts to pairs or groups, early disease detection becomes more difficult. Diarrhea is a major cause of calf death, and one in three calves will experience diarrhea according to producer-reported data to USDA.

Pneumonia is more difficult to identify. It’s easy to train someone to look for diarrhea, but more difficult to teach the early signs of pneumonia. Healthy calves can easily start with a viral respiratory infection, which shifts to a bacterial pneumonia due to changes in weather or diet, movement to a different pen or farm or other stressors.

Robotic milk feeders, which collect information from individual calves via electronic identification (EID), are a solid start to using technology to detect illness. “Instead of paying someone to feed calves in the middle of the day or increase bottle feedings, the robot can take over,” said Cantor. “When we slowly remove milk from a young calf, her response is improved and she transitions to grain better, which is important for rumen development and minimizes stress which is correlated with pneumonia.”

A prediction model can determine which calves will show signs of diarrhea within 24 hours. This is especially useful for farmers who can’t observe calves daily or do rectal exams for diarrhea.

Calves that become ill with diarrhea have lower milk intake and are less active. Researchers created an algorithm to identify potentially sick calves and found the best indicator was calves’ activity 48 hours prior to onset of clinical disease. While this predictive tool isn’t ready for farm use, ongoing trials are underway to improve its effectiveness.

Pedometers are also used on calves. In a study where calves were housed individually, researchers collected information including steps, lying time and activity index. The data measured how quickly calves moved their legs and how many steps they took each day.

“Play behavior is associated with positive welfare and health,” Cantor said. “We know calves that move more probably aren’t going to get sick.”

Early indicators that signal impending illness can significantly reduce antibiotic use.

Calves also stop drinking and eating prior to the onset of pneumonia. Cantor said research shows that a week before calves are clinically ill, they’re starting to show decline.

“We knew the correlations exist, so we tried to take every technological variable and call it the automatic data set,” said Cantor. “Those features were automatic coming from technology versus a veterinarian.”

Cantor said precision technology is more accurate than humans. “Pedometers and robots can classify a calf as clinically ill at 96% accuracy,” she said. “Human information doesn’t change the performance at all.”

The question becomes whether using technology is a cost-saver. Farms can’t afford to use every available technology, so AI must be economical.

“We constrained what the algorithm was allowed to use based on budget,” said Cantor. “With a startup budget of $7,000, the accuracy was about 70%. A $2.50/day budget is realistic for manually feeding calves and health scoring.”

However, a startup budget of $15,000 would allow either a robot or a pedometer, and accuracy for predicting illness jumps dramatically.

“This isn’t just a unicorn reality,” said Cantor. “It’s something we can do. If a farm can only choose one [technology], there’s a benefit for disease detection. The challenge is we have yet to develop the algorithm to clinically find pneumonia in calves. There’s no technology on the market yet that can screen calves with disease.”

The limitations to machine learning are clear: 80% of the farmer’s time is spent managing data. “The idea is taking a complex sensor and being able to find calves when they get sick,” said Cantor. “We use signs like head tilt, abnormal lying, coughing. These are all things technology has never been able to capture because everyone uses only an accelerometer.”

An accelerometer works for lying time and cud-chewing behavior because it’s based on up-and-down motion. Detecting illness as complex as pneumonia requires more creativity. “We’re looking at things we use as ‘cow sense’ to find a sick calf,” said Cantor, noting the symptoms mentioned above. “These are much harder behaviors to automate, so we’re looking at gyroscopes, accelerometers and magnetometers that measure tilts and shaking.”

Research for more complex tools is still in the exploratory stage. Researchers are collaborating with start-up companies that can process information and fuse it into software so researchers can code what calves are doing and “talk” to the sensor data.

“Play behavior looks very different than lying or standing,” said Cantor. “We can label this data to train an algorithm to understand what these look like in the data to eventually automate everything.”

With 18 million data points per tag coming in daily, it’s impossible to perform calculations without the algorithm.

by Sally Colby