Smart Ear Tags and Machine Learning Catch Calf Pneumonia Early, New Study
Bovine respiratory disease (BRD) is still one of the most harmful health problems affecting young cattle around the world. It affects an estimated 30 to 40 percent of calves in the United Kingdom and about 30 percent in the United States. A research team led by Johanna M. C. Brans at Scotland’s Rural College has now shown that combining information from several farm technologies with machine learning can detect the disease in dairy-bred calves that have not yet been weaned. The system reached an accuracy of nearly 70 percent without the need for farmers to check each calf’s health by hand every day. The study, published in Frontiers in Animal Science, shows how smart technology could change the way farmers monitor calf health.
The financial costs of the disease are high. In the UK, BRD costs an estimated £30 to £80 for each affected animal. Lasting lung damage in a young female dairy cow can cause losses of up to £722 per case. Researchers in the United States have estimated short-term costs of about $41.54 to $42.15 per calf, while studies in Australia have reported losses ranging from AUD$67.10 to AUD$384.97. The harm goes beyond these immediate costs. The disease can affect how efficiently animals use feed, how quickly they grow, their ability to reproduce, the quality of their meat, and even how much milk they produce later in life.
BRD is difficult to diagnose because several different germs can cause it, and its symptoms can vary. There is also no single standard test for the disease. Farmers commonly use the Wisconsin score, which gives points for body temperature, discharge from the nose and eyes, coughing, and drooping ears. But checking every calf every day takes a lot of time and is difficult on large farms. The score also shows only the animal’s condition at that moment. Research suggests that the method identifies only about 30 percent of sick animals, even though its specificity is high, at roughly 90 percent. Lung ultrasound scans can give more accurate results, but they require trained workers and repeated checks, which add cost and effort.
To address these problems, the Scottish team monitored 118 calves of mixed breeds at SRUC’s Crichton Royal Farm between August 2020 and March 2021. A total of 100 calves were included in the final analysis. The calves joined the study at around nine days old and were weaned at about 54 days. Each calf wore a newly developed ear tag designed specifically for calf management, called WellCalf and made by Smartbell. The tag used a three-axis accelerometer, a three-axis gyroscope, and an infrared temperature sensor to collect 86 different measurements. The calves also wore IceQube step counters on their back left legs. These recorded five daily measures of movement, including how long the calves stood and lay down, how many steps they took, and a measure of their movement.
The researchers collected two more types of data. An automatic milk-feeding system made by BioControl recorded when each calf visited to drink milk or water, how long each visit lasted, and how much the calf drank. This produced 28 feeding measurements, including the amount consumed and drinking speed. A weather station also recorded daily environmental conditions, providing 24 more measurements related to temperature, humidity, and other factors. After cleaning the data and removing measurements that were too similar to one another, the researchers had 152 variables. They reduced this number to 100 for their models. In total, the analysis covered 223 disease events and 223 healthy events across 98 calves. Two calves stayed healthy throughout the study.
The team used a random forest algorithm, a type of supervised machine learning that combines results from many decision trees, to build 40 different models. Each model used a different combination of data. The simplest models used only details about each animal, such as sex, breed, and birth weight. The most complex models combined daily information from all four technologies and used 29 variables. The researchers also built models that averaged data from the previous one, two, or three days to see whether looking at behavior over a longer period improved predictions. They adjusted the models’ settings using random search, then checked each model with five-fold cross-validation before testing it on data in which healthy and diseased cases were not evenly balanced.
The researchers compared all the models with daily Wisconsin scores recorded by trained farm workers. The level of agreement between different workers was measured using Fleiss’ kappa, which gave a score of 0.78. A calf was classified as sick if its total Wisconsin score was at least 5 and its temperature was above 39.4 degrees Celsius. The researchers trained the models on balanced data using a method called SMOTEENN. This method adds examples of the less common disease class and removes noisy data points. They also tested the models using two measures designed to reflect practical use: how often alerts from the models were correct, and how accurately the models identified complete disease events.
The results showed some important trade-offs. The best overall balance came from a model called Bm1, which used only automatic feeder data averaged over the previous day. It correctly identified 56.0 percent of its alerts and detected 69.1 percent of disease events. It also produced some of the highest numbers of correctly identified sick and healthy cases. The highest accuracy for alerts, at 67.5 percent, came from model FTIEB, which combined daily data from all the technologies. The highest accuracy for detecting disease events, at 69.9 percent, came from model Bm2, which averaged feeder data from the previous two days. Combining several technologies did not always improve disease detection compared with using data from just one source, but it consistently improved alert accuracy.
The findings support earlier research showing that sick calves often change their behavior before clear signs of illness appear. Previous studies have found that calves with BRD move less, spend more time lying down, and drink less milk in the days around diagnosis. Some machine learning systems have predicted illness up to six days before a formal diagnosis. In the Scottish study, models using data from previous days did not always perform best on standard testing measures. However, they were among the strongest when used to generate alerts. This suggests that looking at average behavior over several days could be useful for farmers.
The authors point out several limits to the study. These include the relatively small number of calves, missing data, and the fact that the Wisconsin score is less specific than diagnostic methods such as lung ultrasound. All the calves were kept in one environment, and most of the data came from Holstein animals. This may limit how well the results apply to other farms and breeds. Even so, the researchers say their findings show that BRD can be detected early without manually scoring every calf each day. This could reduce the amount of work required and make health monitoring more consistent. They recommend further testing with independent data from commercial farms. Reliable early-warning systems could help veterinarians target treatment more carefully, reduce the spread of disease, and limit antimicrobial resistance by avoiding unnecessary treatments.
Subject of Research: Using machine learning and combined sensor data to detect bovine respiratory disease in dairy calves that have not yet been weaned
Article Title: Integrated data from different technologies coupled with machine learning can detect bovine respiratory disease in pre-weaned dairy-bred calves
Article References: Brans, J. M. C., Miller, G. A., Chitty, J., Mason, C., Adityan, V., Haskell, M. J., & Duthie, C.-A. (2026). Integrated data from different technologies coupled with machine learning can detect bovine respiratory disease in pre-weaned dairy-bred calves. Frontiers in Animal Science, 7, Article 1957311. https://doi.org/10.3389/fanim.2026.1957311
Try Science Kits
Image Credits: AI-generated
DOI: 10.3389/fanim.2026.1957311
Keywords: bovine respiratory disease, dairy calves, machine learning, random forest, precision livestock farming, ear tag sensors, automatic milk feeder, pedometer, Wisconsin health score, early disease detection, animal welfare, livestock technology