Amidst a global pandemic, innovators and designers are showing ever greater interest in using non-intrusive health-monitoring devices. In previous posts, I've explored the implementation considerations for optical biosensors and commented on the reliability of sensor data. Illustrating the concepts discussed previously, this post provides an account of our team's challenges and accomplishments in creating a longitudinal health monitor. Our subjects were young calves and we monitored them against the onset of bovine respiratory disease (BRD).
BRD is one of the costliest diseases to the livestock industry. It is also one of the most extensively studied, with efforts starting in the late 1800s and continuing today. Considering the U.S. feedlot industry alone, annual losses attributable to BRD add up to US$1 billion.1
BRD refers to a disease complex which results in respiratory infections in cattle. It is particularly difficult to control because it can be caused by a multitude of pathogens. Pathogenesis involves many factors including stress and possible viral or parasitic infections that suppress the host's immune system, allowing bacteria to rapidly reproduce in the upper respiratory tract.
Livestock operators suffer losses in multiple ways from BRD. First, there is the cost of treating sick animals, some of which require more than one course of medication. Second, animals in a feedlot are supposed to gain the maximum amount of weight possible, four pounds per day. When a sick animal fails to gain weight at that pace, that missed opportunity to gain weight is a permanent loss in productivity. Third, caring for sick animals increases labor costs. Lastly, some animals never recover from BRD. BRD is a major contributor to mortality among calves in U.S. feedlots.
Cattle in U.S. feedlots is a small fraction of the bovine population worldwide. American ranchers harvest nearly 300 million calves each year, all of which could potentially be victims of BRD. Because the causes for BRD onset is multifactorial, the cattle industry has not been able to reduce the damages caused by BRD despite decades of research and improvements in animal welfare.
To improve animal welfare and to reduce costs associated with animal mortality and morbidity, the industry has paid significant attention to improving early onset detection. Conventionally, the industry depends on experienced pen-riders who may notice changes in specific animals' behavior and, consequently, test their temperatures. If an animal's core temperature is above 40.5 °C (104.9°F), it is deemed sick with BRD. There is much room for improvement.
First, using a rigid temperature threshold to diagnose BRD ignores the reality that each animal has different nominal core temperatures. 40.5 °C is merely the average feverish temperature over a large herd of cattle. Some of them have core temperatures that are naturally lower than the average and could already be experiencing a fever even when their core temperature is below the threshold. This is the same principle that guides the understanding that longitudinal and personalized health monitoring results in better healthcare for human patients.
Second, although most feedlot operators are confident of their pen-riders' abilities, researchers report a majority of harvested cattle examined had lesions in their lungs, i.e. had suffered (and recovered) from serious respiratory illnesses, even though most were never treated for BRD. In truth, an experienced pen-rider can expect to take care of 8,000 to 12,000 heads of cattle per day. Consequently, each possibly sick animal could command fewer than two minutes of a rider's attention. Furthermore, because of the labor cost required to handle and test a possibly sick animal, riders are trained to err on false negatives. Making matters worse, the labor pool of pen-riders is aging with fewer experienced hands who could make accurate snap judgments.2 All these factors provide sound arguments for the creation of an automated early onset detection system for BRD.
Because a fixed core temperature threshold, 40.5°C, has been the gold standard for diagnosing BRD, there have been a lot of attempts to monitor temperature on a head of each cow. These attempts include measuring rumen telemetry temperature using an ingested bolus, measuring air temperature in the ear canal or the temperature of the tympanic membrane, measuring skin temperature, and applying thermal imaging to estimate ocular temperature. These approaches either suffer from very low data reliability or prove to be incompatible with workflow at the feedlot (more on this in the next section).
Veterinary medicine researchers investigating other onset symptoms of BRD incidentally reported on SpO2 measured from cattle at different stages of BRD. The data showed declining average value as the disease became more severe.3 Since BRD is a respiratory disease, it makes sense that SpO2 would drop as respiratory infection compromises oxygen intake. Indeed, experts in the field that we've consulted with agree that SpO2 should be a valuable biomarker. However, it also becomes clear that there hasn't been tracking of SpO2 values in a cow through a course of BRD infection to help us understand how that vital sign may change and whether it is a leading or lagging indicator.
Attaching health monitors to a herd of cattle must conform with the workflow in a feedlot. Working with industry experts, our team quickly learned the constraints.
- Ear tags are attached to cattle when they are put in a chute. Corralling animals to put them through a chute is labor intensive and is performed only a few times throughout the entire stay of an animal's time at a feedlot—notably during intake and when animals are sorted prior to harvest.
- Pen-riders check on the animals once a day on average. Animals could receive some attention during feeding, but the process is semi-automated in large feedlots.
- U.S. Food and Drug Administration (FDA) regulations require alien objects, such as a temperature-measuring bolus in the stomach, to be removed from the animal carcass so as not to accidentally contaminate our food supply. However, the head, including the ears, is not considered an edible part and is discarded after slaughter. Indeed, ear tags are already used for identifying and tracing feedlot animals. An animal is tagged as part of its intake procedure at the feedlot, and the tag remains on the animal until its harvest.
- BRD is most prevalent in young calves immediately after they are transported to the feedlot. This is so common that BRD is also called shipping fever. Once an animal is acclimated to its feedlot, the risk of BRD decreases significantly. However, the value of health monitoring does not completely go away as mortality cost of an animal increases over time. Most significantly, the industry is keenly interested not just in monitoring animals for BRD but also in monitoring their overall health and welfare.
Given the established workflow at feedlots, a smart livestock system (see Figure 1) could assist pen-riders by alerting them of animals whose oxygen reading has dropped below the longitudinal norm of that specific animal. The system would consist of an ear tag equipped with SpO2 and other sensors. Unlike wearable fitness devices, ear tags are permanently installed on cattle, so each would monitor approximately the same location on an animal and mitigate against data variations associated with how a device is worn from day to day. Even so, we speculate that data could vary as a calf grows and the length of its hair changes, but we surmise that such variations are slow and could be tracked with an intelligent baseline algorithm.
The ear tag (Figure 2) would collect sensor data and local processing would extract biomarkers including SpO2. A low-power radio, such as one based on LoRa, would transmit these biomarkers to a base station which could conveniently be housed in the office of a feedlot or installed on feed delivery trucks. Algorithms in the cloud could then process the biomarker data, maintain a longitudinal record for each animal and send an alert when the biomarkers deviate markedly from the norm. Biomarker data are only needed to generate at most one assessment per day, so the ear tag could run for a long time on a primary cell battery.
Incidentally, animal hair is not a significant problem for optical biosensing. Hair, like tissue, is a form of protein and is readily penetrated by LED lights (Figure 3). In fact, in young calves weighing at least 300 to 400 pounds, we obtained better perfusion indices in our PPG signals from a calf's ear tag than from fitness watches on smaller humans.
For our examination, we created an ear tag modelled after commercial tags used on cattle. The tag includes an optical PPG using a MAX86141 analog front-end with red and infrared LEDs. Because optical SpO2 measurements are vulnerable to motion artifacts, we included an accelerometer so we could detect motion and reject any PPG data captured when the tag is moving.
Since the goal of our feasibility study was to find a workable algorithm for cattle monitoring, we set our PPG to sample at 100sps. Bovine heart rate averages below 84bpm, which is not much different from the human heart rate. We selected a higher-than-necessary sampling rate to be sure that we captured the higher order harmonics of PPG waveforms. We stored all the raw data from the PPG and accelerometer on flash memory on the tag and recovered tags from our experiments for batch processing. This approach avoids any complexity of data loss from wireless transmission. Clearly, in any production deployment, we would have to run our algorithm on the ear tag and only transmit a small payload of biomarker data.
Trying to balance between operating life and the weight of the ear tag, we used a 1000mAh battery. To minimize power consumption, we did not include a green LED in the PPG since we could get heart rate information, though suboptimal, from red or infrared channels just as well. We were able to run the ear tag for about 10 days using our data acquisition scheme (see below).
It is important when creating a sensor to have a reference, or ground truth. However, getting ground truth of SpO2 on cattle is difficult. Pulse oximeters are designed to be used by stationary patients in clinical settings, and readings are generally taken as snapshots and not continuously. Undaunted, the team took pulse-oximeter measurements and drew blood from a handful of anesthetized heifers to correlate their blood-oxygen content with our sensor readings.
We took readings using a few different clinical-grade pulse oximeters with the goal of calibrating our ear tag sensors to a reference. While the oximeters' readings agreed with each other when oxygen concentration was high, when the blood-oxygen level dropped near around 90%, their results materially disagreed with each other. Even examining arterial blood for oxygen concentration was prone to human error, resulting in significant repeatability error. Our first attempt to establish ground truth was a bust.
Stepping back, we realized our objective was to detect the onset of BRD, while a drop of BRD was merely the symptom we were tracking. If we can control when an animal gets infected, using a controlled disease challenge study, there are established procedures to determine when a challenged animal becomes ill without having to know the animal's blood-oxygen level.
We worked closely with researchers at Texas A&M University AgriLife Extension on a series of experiments. Through a series of design changes, we experimented with different distances between the LEDs and the photodetector in our sensor design to improve perfusion indices for red and infrared lights. We also worked to improve the tissue-to-optical-sensor coupling by testing a series of mechanical ear tag designs, which we produced on a 3D printer. The 3D printed materials proved to be fragile, and our ear tag assemblies often did not stand up to the rough treatment of prolonged wear by our cattle. This last defect plagued our experiments, but we believe it could be easily remedied by moving away from using 3D printing. Nevertheless, 3D printing was the most efficient means by which we could change our mechanical design so we could learn and improve faster.
The AgriLife team led us to design a BRD challenge experiment following well-known industry protocol4. The protocol set up two groups of animals: one for control and the other to be challenged. The challenged animals were inoculated with Bovine Herpes Virus-1 (BHV-1) and three days later given Mannheimia haemolytica (MH) using broncho-selective endoscopy. The controlled animals were treated with phosphate-buffered saline solutions following otherwise identical procedures. Both groups were monitored on rumen temperature and feeding behavior and were visually evaluated for clinical signs of illness. These daily monitors, especially any rumen hyperthermia event (upward temperature spike), served to give us the ground truth of BRD onset. For good measure, we also took arterial and venal blood samples from the animals' ears a few times over the course of the experiment.
We monitored SpO2 levels compared to a baseline when the animal was known to be healthy. We could simply use R as our biomarker, where
SpO2 ≈ a0 + a1R + a2R2
In fact, for small deviations around a norm, the relationship between SpO2 and R is roughly linear: the constant a1 is a negative number. So, when SpO2 drops, R increases.
We took time to establish baselines for our references, rumen temperature, and feeding behavior for each animal. One day before we administered BHV-1, we installed our ear tags. The ear tags collected 40 seconds of PPG and accelerometer data every 40 minutes from 9p.m. to 6a.m. each day until they were removed or dislodged. The entire process was handled by the AgriLife team.
Many ear tags proved too fragile to withstand the rough treatment by the animals. Consequently, a significant portion of the sensors were destroyed and unable to provide useful data. Among the recovered data, both R (Figure 4a) and arterial blood-oxygen level (Figure 4b), SaO2, showed a small departure from the norm following the MH challenge. The magnitude of the change is about 50% larger than the normal daily fluctuation of the biomarker (control).
It is notable that over the experiment period, we experienced two significant thunderstorms and the average nightly temperature changed by around 15°C. The changes in environmental conditions did not affect our biomarker.
The experiment showed that R, and by inference SpO2, is a measurable but weak indicator for the onset of BRD. Physiologically, an animal will compensate for oxygen deficiency any way it can. So if respiratory function is affected during BRD onset, the animal would increase respiratory rate and heart rate to keep blood-oxygen levels from degrading badly. Only as the disease progresses, and its lung becomes further damaged, would these coping mechanisms fail to keep up and blood-oxygen level drop precipitously. As such, measuring rumen temperature via a bolus or monitoring feeding behavior with GrowSafe Systems provided a much stronger signal for the BRD challenge study. Unfortunately, neither monitoring system fits with the beef production workflow.
Attempts at creating an automated system to monitor animal wellness continue. There are some natural follow-ups to our feasibility study.
- Combine SpO2 (or R) with another biomarker. Using a single biomarker for BRD detection is vulnerable to false triggers. Although the signal from R is not pronounced, it may be less affected by environmental changes. So combining R with other biomarkers such as temperature and activity may provide more robust predictions.
- Qualifying SpO2 (or R) with resting heart rate. The optical sensors on our ear tags also provide heart rate of the animal. As discussed, resting heart rate should increase as an animal compensates for its compromised respiratory functions. We will have to modify our data collection protocol to capture accelerometer data long before turning on our optical sensors so we could identify resting heart rate.
- Improve motion artifact mitigation. Since a cow's ears move continuously, we can still improve our R biomarker detection algorithm against expected motion artifacts.
- Study over a much larger population. Each animal reacts to diseases differently. We need data from a much larger population of cattle: different breeds, in different climates, at different altitudes to validate any health-monitoring approaches to accommodate the variations.
- Natural BRD onset vs. challenge. BRD occurring in nature does not follow the pattern of a controlled disease challenge. Consequently, algorithms developed using a series of challenge studies may not be as effective in a production feedlot setting.
This study benefited from experts in a multitude of disciplines including animal science, veterinary medicine, bioengineering, optical design, data science, and algorithms, as well as the dedication of our applications staff. In an era that touts advances of using Big Data, we find that animal scientists and veterinarians often lack existing data pertaining to our use case. In our situation, the lack of data could be attributed to the fact that sensors that could collect longitudinal SpO2 measurements consistently on live animals did not exist. Such sensors were not developed because, until this diverse group of experts started to work together, engineers were not aware of any potential use cases.
Nevertheless, researchers remain interested in remote and noninvasive methods to assess animal health and welfare, working to detect not just BRD but also issues such as bloat and stress. Understanding how to monitor an individual animal's wellness helps us learn how personalized healthcare could be implemented reliably for human patients. Biosensors will continue to play expanded roles. Exciting times are ahead!
- 1Bacterial Pathogens of the Bovine Respiratory Disease Complex
- 3Blood Gases, Acid-Base Status and Plasma Lactate Concentrations in Calves with Respiratory Diseases
- 4Effect of Mannheimia Haemolytica Pneumonia on Behavior and Physiologic Responses of Calves During High Ambient Environmental Temperatures