24 September 2012—A jump to the left and a step to the right are signs of healthy activity, as chicken farmers who stroll among their flocks already know. Now a team led by robotics engineer Stephen Roberts at the University of Oxford has found that patterns in the collective motion of a flock of chickens can help farmers predict disease weeks before onset. Call it a chicken time warp.
Roberts and animal-welfare researchers at Oxford first tested their pattern-detection system by asking it to warn farmers before a flock got “peckish.” That’s not a euphemism for “hungry.” Well-fed hens, it turns out, sometimes take out their worm-hunting instincts on one another. The system, which consisted of cameras recording a flock, followed by computer analysis of the footage, beat human experts at flagging the at-risk flocks before the madness took its toll [“Computer System Counters Hen Horrors,” September 2010].
“The key thing is, if you can predict at, say, one week [of age] which flocks are going to have trouble at six weeks, then you can intervene,” says veterinary scientist Kenton Morgan of the University of Liverpool, in England, who was not part of Roberts’s team. Morgan and his colleagues have also used machine learning—the artificial-intelligence approach underlying Roberts’s method—to monitor and predict health outcomes for broiler chickens.
Unlike Roberts’s camera-equipped team, Morgan’s group drew for its latest study on the equivalent of public-health data already collected by chicken farmers, such as how much water the chickens are drinking, their average weight, their growth rates, and other factors, to detect which birds were at risk of “hock burn.” Quality-control experts use signs of this condition, found on a portion of the legs of chicken exposed to chicken litter, as a sign of overall flock health. Farmers often record all those factors by hand—a labor-intensive process.
Keeping the chickens on camera, as Roberts’s team did, allowed for continuous monitoring. So after their success predicting the hen-peckers with their chicken-coop cameras, Roberts and colleagues turned to hock burn.
The team recorded 24 flocks of 34 000 broiler chickens each, raised in four henhouses. That gave them a larger data set than they had for their previous study, which monitored 18 smaller flocks of laying hens. They also updated their algorithm, which analyzes “optical flow”—the pixel-to-pixel changes from one scene to the next—to look for measures of flock movement that correlate with hock burn.
They then searched for differences in the optical flow of flocks at various dates throughout the chickens’ lifetime. At two weeks of age, the optical-flow analysis could already detect differences between healthy flocks and those that later developed hock burn. The results and timing were similar to the findings of Morgan’s group, made using the hand-collected health data. However, by aggregating day-to-day video records, the Oxford team made a dramatic improvement in their algorithm’s ability to predict disease. By aggregating data from the flocks’ first two days of life, the team predicted which flocks would show hock burn symptoms two weeks later.
“It’s a very cool study,” Morgan says, and it helps show the “tremendous potential for engineering solutions to animal health.” At the same time, he notes that the study was conducted in four chicken houses, so these results are specific to that site until researchers try the method elsewhere. Roberts agrees, noting that this study was of only 24 flocks and that “ideally, we’d have this data for thousands of flocks.” He also adds that the team had less success predicting other quality-of-life indicators.
Morgan says he imagines that future chicken farmers will feed their machine-learning algorithms data from a variety of sources, such as cameras, microphones, air-quality sensors, and moisture sensors. All that predictive power should buy farmers the time to improve conditions for their flocks and keep their chickens dancing.
About the Author