Improving Behavior Monitoring of Free-Moving Dairy Cows Using Noninvasive Wireless EEG Approach and Digital Signal Processing Techniques

Featured Application

A method that allows brain signals to be monitored in freely moving bovines is described. The method uses noninvasive electrodes to minimize stress during EEG monitoring and allows bovines to behave normally during the process. The method can be applied to investigate changes in brain electrical activity under normal handling conditions.

Abstract

Electroencephalography (EEG) is the most common method to access brain information. Techniques to monitor and extract brain signal characteristics in farm animals are not as developed as those for humans and laboratory animals. The objective of this study was to develop a noninvasive method for monitoring brain signals in cattle, allowing the animals to move freely, and to characterize these signals. Brain signals from six Holstein heifers that could move freely in a paddock compartment were acquired. The control group consisted of the same number of bovines, contained in a climatic chamber (restrained group). In the second step, the signals were characterized by Power Spectral Density, Short-Time Fourier Transform, and Lempel–Ziv complexity. The preliminary results revealed an optimal electrode position, referred to as POS2, which is located at the center of the frontal region of the animal’s head. This positioning allowed for attaching the electrodes to the front of the bovine’s head, resulting in the acquisition of longer artifact-free signal sections. The signals showed typical EEG frequency bands, like the bands found in humans. The Lempel–Ziv complexity values indicated that the bovine brain signals contained random and chaotic components. As expected, the signals acquired from the retained bovine group displayed sections with a larger number of artifacts due to the hot 32 degree C temperature in the climatic chamber. We present a method that helps to monitor and extract brain signal features in unrestrained bovines. The method could be applied to investigate changes in brain electrical activity during animal farming, to monitor brain pathologies, and to other situations related to animal behavior.

1. Introduction

Many biological systems emit signals. If these signals are interpreted correctly, they can provide information about the processes that occur in the system, adding to knowledge about biological systems. Electroencephalography (EEG) is the most common method to access brain information. During EEG, brain electrical activity is recorded by placing electrodes on the scalp. This method has high temporal resolution and is safe, easy to use, and affordable [1,2]. In the last few decades, great effort has been devoted to understanding the human brain. This effort has allowed neurological disorders such as epilepsy [3,4,5] and Parkinson’s disease [6,7,8] to be diagnosed through the brain–computer interface [6,9,10].

Several DSP techniques have already been applied to humans and laboratory animals, e.g., FFT (Fast Fourier Transform) [11], wavelets [12], AGR (Adaptive Gaussian Representation) [13], ANN (Artificial Neural Network) [14,15], Lempel–Ziv complexity [16], and entropy [17], among others [18,19]. This kind of technology has also been employed to investigate some diseases in domestic pets [20,21,22]. Concerning farmed animals, most of the effort that has been put in understanding brain processes has focused on pre-slaughter sensitization [23,24]. The exception is a recent study on sheep [25], which involved a method that did not use restraint and employed implanted electrodes to analyze animal brain signals instead.

Within the Brazilian agribusiness, meat production is one of the items whose contribution to the trade balance has grown the most, thereby making bovines an important field of study [26]. In this scenario, understanding the brain processes in bovines will allow some situations such as heat stress [27,28] and pain [29,30], among others [24], to be controlled.

Acquiring brain signals in animals that can move freely presents various challenges, and the number of research studies dealing with bovine brain electrical activity is limited. Most of these studies have been conducted in laboratory conditions [31] and used subdermal electrodes [31,32] in animals that were subjected to some type of containment [24,32]. These conditions are unrealistic when we consider farmed animals, mainly large ones [25].

Telemetric recording systems can monitor physiological parameters in freely moving animals without any restrictions in their exploratory behavior [33]. In almost five decades, different systems have been developed, from large systems [34] to transmitters that can be implanted beneath the skin of small animals (e.g., rats) [35].

Another challenge associated with electroencephalography involves the presence of artifacts within the signal. Given the nature of the acquisition process, the EEG signal can mix with other biological potentials. Broadly speaking, artifacts can originate from internal (physiological activities of the subject) and external sources (environmental interference, equipment, electrode pop-ups, and cable movement) and interfere with recordings in both temporal and spectral domains [2]. Where signals from animals are concerned, the presence of artifacts is hardly avoided. In the case of bovines specifically, there is the additional challenge of prolonged rumination throughout the day.

Given the number of research studies on bovine brain signal analysis and the kind of signal processing techniques they apply [23,24,31,32,34,36,37,38], a good strategy to characterize these signals is to assume that they are nonstationary, as in humans, and to employ time–frequency techniques to estimate the spectra.

EEG is typically described in terms of rhythms and transients. The rhythmic activity of EEG is divided into frequency bands [1,2]. This characteristic makes traditional analysis rely mainly on the detection of spectral power changes [39]. The Power Spectral Density (PSD) is calculated by Fourier Transform, the estimated autocorrelation sequence that is found by nonparametric methods. One of these methods is Welch’s method [40]. Because the EEG signal is nonstationary, the most suitable way to extract features from raw data is to use time–frequency domain methods, such as STFT (Short-Time Fourier Transform) [41,42,43]. STFT is a time-dependent Fourier Transform that may be calculated by sliding a window over the signal to compute the FFT [41]. It is represented by a two-dimensional time–frequency plot, called a spectrogram. The major drawback of STFT is the compromise between time and frequency resolution. If a short window is applied, information derived from each FFT will be well-localized in time, but the frequency resolution will be poor.

Feature extraction is an important task in signal processing. DFT (Discrete Fourier Transform) can convert an equally sampled signal into a list of coefficients. These coefficients comprise a finite combination of complex sinusoids ordered by frequencies; in other words, DFT converts a digital time-domain signal to a frequency domain [44]. To enhance DFT estimation, the FFT algorithm may be applied. As the name implies, FFT is a fast version of DFT [45]. Fourier Transform is one of the nonparametric feature extraction techniques for EEG signals [46,47]. In addition, the electrical activity of the brain (EEG) exhibits significant complex behavior with strong nonlinear and dynamic properties [48,49]. Nonlinearity in the brain is introduced even at the cellular level because the dynamic behavior of individual neurons is governed by threshold and saturation phenomena [50]. This implies that nonlinear methods can be applied to investigate EEG dynamics [49].The Lempel–Ziv complexity (LZC) [51] uses symbolic techniques to map a time series into a sequence that retains its dynamics. LZC has been widely used in biomedical applications to estimate the complexity of discrete-time signals [52]. Normalized complexity can be used to quantify nonlinear and nondeterministic data [53].

This study aimed to develop a method to monitor bovine brain electrical activity by noninvasive techniques in animals in an experimental paddock. Another objective was to apply techniques such as PSD (Power Spectral Density), STFT, and LZC to characterize the obtained signals, a necessary step to expand this field of study.

2. Materials and Methods

Twelve cows were included in the experiments, which were performed according to the Institutional Animal Care and Use Committee Guidelines of the College of Animal Science and Food Engineering of the University of São Paulo.

This section is divided into three subsections: (1) developing a portable and low-cost wireless electroencephalograph; (2) developing a method to monitor bovine brain electrical activity; and (3) extracting the features of these signals.

2.1. Wireless Electroencephalograph

The EEG acquisition system consisted of two modules: The first module, which was embedded in the animal, concerned the recording, conditioning, and wireless transmission of the EEG signals. The second module, which was called the base module, received data from the embedded module and connected the system to a computer [54].

2.1.1. Animal Module

This module could be attached to the animal’s neck, and it performed three tasks: (1) signal amplification and conditioning, (2) analog-to-digital conversion, and (3) wireless transmission of the digital data. A microcontroller sampled the EEG analog signal, converted the analog signal to a digital signal, and transmitted the recorded data [54]. Each of these parts is detailed below.

The electronic amplification and conditioning circuit consisted of (i) an amplifier that pre-amplified the EEG signal and (ii) three analog active filters: a 0.05 Hz high-pass filter, a 1.5 kHz low-pass filter for antialiasing, and a band-pass filter to remove 60 Hz electromagnetic interference. The input impedance was on the order of Megaohms. The total gain of the circuit was 10 V/mV.

The A/D converter stage of the circuit used a 10-bit analog-to-digital converter of a Microchip^TM PIC16F877A microprocessor [55]. Through its USART interface, the microprocessor also controlled telemetric transmission. The circuit used the BIM2-433-160 transceiver [56], which transmitted and received the data at 433 MHz with low energy consumption in a range of up to 200 m. This setup allowed for the real-time transmission of up to 120 data points per second. Therefore, a safe sampling frequency of 100 Hz was chosen.

2.1.2. Base Module

The base module received and stored data arriving from the embedded modules. This module consisted of one 433 MHz channel that used a BIM2 transceiver [56]. Dedicated software version 1 [54] in the computer stored the data in a database and showed the EEG in real-time on the display monitor.

2.1.3. Equipment Validation

To validate the test equipment (TE), the TE’s ability to amplify signals of low amplitude and frequency was evaluated and compared to that of the control equipment (CE), which consisted of a 32-channel digital electroencephalograph from the company EMSA Medical Equipment S/A, model BRAINNET BNT-EEG. The following signals were evaluated: 3.90-Hz and 30.20-Hz sine waves generated by a sine generator. Stretches of 5 s were collected at a sampling frequency of 100 Hz.

In the frequency domain, the PSD was estimated by the Welch method for each signal. In the time domain, the comparison was made by considering the signal-to-error Ratio (SER) (Equation (1)). The method considered the developed equipment as the output of a nonlinear predictor of the EEG signal acquired by the control equipment [57].

$$SER={10log}_{10}\left[\frac{\sum _{i=1}^N{\left(x\left(i\right)\right)}^2}{\sum _{i=1}^N{(x\left(i\right)-p\left(i\right))}^2}\right]dB$$

(1)

2.2. Signal Monitoring

The experiment was carried out in the city of Pirassununga, state of São Paulo, Brazil (latitude 21°59′46″ south and longitude 47°25′33″ west, altitude of 627 m). The signals were acquired from Holstein heifers.

To develop a method to monitor bovine brain electrical activity, it was necessary to define the sample frequency (Fs), determine where to place the electrodes (positioning and fixation) and equipment, and deal with artifacts and other acquisition issues. These conditions were established by conducting measurements with two animals.

Considering the frequency range of bovine brain signals reported in the literature [31,37] and the real-time transmission limitations of the equipment, the Fs was initially set to 100 Hz. This value ensured that frequencies up to 50 Hz were mapped without aliasing. This 50 Hz limit expanded the frequency spectrum observed by the mentioned authors and allowed for the presence of frequency components above 40 Hz, associated with pain and stress [58]. The signals were digitized in the embedded module.

The preliminary study for electrode positioning test was conducted on animals sedated with 1 mL/kg of Xylazine (Rompun from Bayer^TM), totaling 6 mL. For each position, signal collections were performed with durations ranging from 20 to 70 s, and from these collections, artifact-free segments were selected for analysis.

This study was conducted using noninvasive electrodes consisting of 1.0 cm gold discs. Electrode fixation was attempted with either strong adhesive tapes or cyanoacrylate, also known as super glue (Super bonder—Loctite). First, the area at the forehead of the cows where the electrodes would be fixed was shaved. Then, the electrodes were filled with conductive gel for the EEG. Next, the animal module was placed in a small bag around the animal’s neck. The module remained fixed for a month to facilitate the animal’s adaptation to the equipment.

The equipment comprised a single-channel electroencephalograph (bipolar measurement), so the sensor consisted of two electrodes to acquire the biopotential signal acquisition as well as a grounding electrode. The grounding electrode was positioned at the back of the animal’s neck, whereas the other two electrodes were tested at three different positions (POS1, POS2, and POS3) of the animal’s forehead. These positions were selected on the basis of the bovine head anatomy. Anatomical head sections were analyzed to investigate positions for electrode fixation. The images of heads that were used for this analysis were obtained at a local slaughterhouse, so they are not images of heads of the animals that were used in the experiments.

After the establishment of acquisition conditions, a new experimental stage was initiated, and the signals were acquired from awake animals (free from sedation) in two situations: (1) from animals contained in a climatic chamber (cage containment) and (2) from animals in pasture (animals with freedom of movement).

The main experiment involved twelve Holstein heifers, which were divided into two groups of six. In the first situation, the animals were contained in 1.0 × 2.0 m individual cages arranged as a 2 × 3 matrix inside a climatic chamber. In the second scenario, the animals were situated in enclosed compartments within an experimental paddock. These compartments measured 4.0 × 3.0 m and were arranged side by side in a 1 × 6 matrix. The cows had unrestricted access to food and water throughout the experiment.

During the experiment, the temperature and relative air humidity were collected and used to determine the enthalpy of the place, as proposed by Albright [59]. Enthalpy is considered an index of thermal comfort and reflects the environmental conditions related to the thermal stress experienced by the animals [60,61]. This stage of data acquisition lasted three full days.

2.3. Signal Characterization

For both groups of animals, all the stages of brain signal characterization were performed with the Signal Processing Toolbox^TM from MATLAB^® software releases R2015a and 2019a, developed by The Mathworks, Inc.

In pre-processing stages, different methods can be employed to handle artifacts. One of them rejects the epoch or segment of EEG data that is labelled as artifactual [19]. Sauter et al. [62] proposed a sequential method where spectral analysis is performed after corrupted segments are extracted. For this purpose, artifacts are detected by visual inspection and then rejected. The data considered as artifacts were those that exceeded the amplifier’s scale and neighboring segments. The signals used for analysis were labeled as “artifact-free” epochs. Welch’s method was applied to artifact-free signal segments to estimate the signal PSD and to verify whether the spectral content is due to brain signals.

PSD makes the frequency spectrum smoother than the raw FFT output, and the Welch algorithm [63] is a nonparametric method to estimate PSD. Let us consider that the FFT is calculated over the entire duration of the signal. In the Welch algorithm, instead of processing the FFT over the entire time domain, the signal is separated in windows with the same size. Window size affects the clarity of the result by cutting frequencies with periods larger than the window. Windowing is taking a sample of a larger dataset and tapering the signal at the edges of each interval. This makes the signal smoother without sharp transitions that can disturb the frequency spectrum representation [64]. Welch’s method was applied to artifact-free signal segments to estimate the signal PSD.

Considering that STFT is a sequence of Fourier Transforms of a windowed signal, it can be represented as in Equation (2) [65].

$$X_{STFT\left[m,n\right]}={\sum }_{k=0}^{L-1}x\left[k\right]g\left[k-m\right]e^{-j2\pi nk/L}$$

(2)

where x[k] denotes the signal, and g[k] denotes an L-point window function.

STFT was also employed to artifact-free signal segments so that data could be mapped in the time–frequency domain.

For LZC calculation, each signal segment is converted into a binary sequence s(n), as follows [52,66]:

$$s\left(n\right)=\left\{\begin{array}{c}1, if x\left(n\right)>m\\ 0, if x\left(n\right)<m\end{array}\right.$$

(3)

where x(n) is the signal segment, and n is the segment sample index from 1 to N (segment size). The segment length N was chosen for at least 400 samples, a value for which normalized LZC is stabilized [52]; m is the signal mean value. Thereafter, the resulting binary sequence s(n) is scanned from left to right, and the number of different patterns is counted. The complexity value c(n) is increased every time a new pattern is encountered [51,52,66]. Lempel and Ziv [51] showed that

$$\underset{n\to \infty }{\lim }c\left(n\right)=b\left(n\right)=H\frac{n}{{\log }_{k}n}$$

(4)

where n is the segment length, and k is the number of different symbols in the sequence (in the binary case, k = 2, and H represents entropy (Equation (5)).

$$H=-\frac{1}{\ln \left(k\right)}\sum _{i=1}^k{p}_i\log {(p}_i)$$

(5)

where $p_i$ indicates the probability that a state i is obtained by counting the occurrences of each symbol, divided by the total number of symbols in the sequence.

To avoid the variations due to the segment length, normalized LZC values are calculated as follows:

$$C\left(N\right)=\frac{c\left(N\right)}{b\left(N\right)}$$

(6)

The more random the sequence, the greater the number of different patterns present in it; that is, its normalized complexity C(N) approaches one. Meanwhile, C(N) values that approach zero are associated with deterministic sequences. Normalized LZV–C(N) was applied to artifact-free signal segments to investigate oscillatory, chaotic, and random components in bovine brain signals. To illustrate the relationship between C(N) and signal characteristics, low- and high-complexity known sequences were calculated.

3. Results and Discussion

3.1. Equipment Validation

Figure 1 presents the graph of a 3.90 Hz sine wave sampled at 100 Hz by the test equipment (TE) and control equipment (CE). In this case, SER was 39.35 dB.

Figure 2 shows the graph of a 30.2 Hz sine wave sampled at 100 Hz by the test equipment (TE) and control equipment (CE). In this case, SER was 38.36 dB.

Figure 3 compares the PSD of both sine waves obtained with the TE and CE.

In the time domain, the signals were only slightly significantly different as compared to the signals obtained with the method proposed by Kavitha and Narayana Dutt [57]. The main difference between the signals was due to the equipment resolution, which was 10 and 12 bits for the TE and CE, respectively. The PSD indicated the same frequencies for both the generated sines.

The first telemetric equipment developed to record EEG [34] weighed approximately 0.450 kg, transmitted data up to 91.44 m, and was completely analog. The equipment cost about USD 1500, including the receiver. The literature has no other report on the development and use of other EEG equipment for use in cattle.

The equipment developed herein weighs 180 g and transmits data up to 200 m. In addition, it has a microcontrolled processing system that allows digital data transmission and costs approximately USD 100.

3.2. Signal Monitoring

Figure 4a shows a head section of an adult bovine. This figure illustrates the distance from the brain to the frontal region of the bovine head. Finding the best position for signal acquisition involves locating the region where this distance is as short as possible. Figure 4b shows the layout of the monitoring electrodes at the three evaluated positions.

Figure 5 depicts a disc electrode and electrode fixation with adhesive tape.

The animal module consisted of a box measuring 2.5 × 7.0 × 11.0 cm³ and weighing 180 g. To make it less uncomfortable for animals, the box was placed in a fabric bag, which was attached to the neck with a soft ribbon. Figure 6a presents the animal module placed at the animal’s neck, whilst Figure 6b,c shows the halter that was used to protect the electrodes and the electrodes fixed with super glue.

The adhesive tape failed to keep the electrodes rigidly attached to the animal, so they often came loose. The glue (Figure 6b,c) kept the electrodes in the right position for a whole day of acquisition. At the end of the day, the electrodes and the animal’s forehead had to be cleaned for a new acquisition. Compared to the use of implanted electrodes, which can go on for a year [25], this was a short period. It is important to highlight the possibility of noninvasive monitoring here, though.

Figure 7 contains the signals acquired at the three selected positions for one of the two animals in this stage of the experiment, while Figure 8 corresponds to the artifact-free signal sections in Figure 7.

Figure 9 contains the PSD for the signal sections in Figure 10.

In Figure 7a, the signal at POS1 presented many artifacts, most of which were due to ear movement. POS2 provided longer artifact-free sections than POS1 and POS3 as a result of two main facts. First, POS3 was largely influenced by eye movement, whereas POS2 is the region where paranasal sinuses are smaller, and hence, the surface is closer to the cerebral cortex. This result was expected, given the analyzed anatomical head section (e.g., Figure 4). The three positions had similar PSD profiles, and the observed frequencies (2–10 Hz; 13–27 Hz) agreed with those obtained by Suzuki et al. [31]. Therefore, the metric to choose the best electrode position was the length of the obtained artifact-free section, and POS2 was selected as the best one to acquire brain signals in cattle.

Most studies in this area have been conducted in laboratory conditions [31] and used subdermal electrodes [31,32] in animals that were subjected to some type of containment [24,32]. The method presented herein allows brain signals to be recorded with surface electrodes in freely moving animals.

Figure 10 shows the arrangement of the awake animals within the cages inside the climatic chamber, while Figure 11 illustrates their positioning within enclosed compartments within the experimental paddock.

Figure 12 displays the enthalpy values for the animals in the climatic chamber or pasture during the first of the three days of experiment.

In the climatic chamber, between 20:00 h and approximately 22:30 h as well as after 03:00 h, the animals experienced a heat stress situation, but most of the time, they were in a comfortable situation (30–70 J/kg of dry air).

The stress condition was due to animal agitation during periods of enthalpy above 70 J/g, resulting from panting and increased water vapor. This led to a rise in internal relative humidity, progressively hindering heat loss through breathing and the skin surface. Under these conditions, the temperature and relative air humidity were above 32 °C and 60%, respectively. After a period of reduced animal agitation and air renewal in the chamber, the condition of thermal comfort was established, with average values of 25 °C and 60%. In the pasture, there were no conditions of thermal stress. In fact, during the analysis period, the enthalpy decreased as the incidence of solar radiation decreased, which modified the temperature and relative air humidity [60].

3.3. Signal Characterization

Figure 13 exemplifies a pre-processing stage when artifacts were removed from the signal for one of the animals in the climatic chamber (Figure 13a) and one unrestrained animal (Figure 13b).

The proportion of artifact-free data in relation to the original data seen in Figure 13 was observed for most of the database. In all the acquisitions, at least one artifact-free section could be selected.

Figure 14a contains the PSD for the artifact-free signal section in Figure 13a, while Figure 14b shows the PSD for the second artifact-free signal segment in Figure 13b.

We were not able to establish a direct relationship between enthalpy and any specific behavior observed during the experiments (eating, drinking water, chewing, standing, lying down, or walking). The discomfort experienced by animals in the cage overlapped with any other behavior. This reflected even on the duration of artifact-free epochs in the signal (less than 4 s for all the animals).

Figure 15 illustrates STFT for the second artifact-free signal segment in Figure 13b.

We were not able to associate the observed frequencies in the group with any specific animal behavior, such as eating, drinking water, chewing, standing, lying down, or walking, mainly because of the amount of data that was acquired. However, the spectrogram clearly showed the predominant frequencies in cattle signals (Figure 15). As expected, the delta wave exhibited high amplitudes between 1 and 3 Hz and was particularly prominent between 1 and 2 Hz. These frequency values might potentially correspond to the delta wave or arise as a result of the rumination process [67]. The electromyographic signal has amplitudes in the range of mV, while the brain signal presents amplitudes in the range of μV. Considering this difference, in Figure 13, the muscular movements might be related to segments of “scale clipping”. These segments were not used for processing, but it is still possible that there could be some interference in the selected segments for processing, which could contribute to the presence of frequencies in the range of 1–3 Hz observed in Figure 15.

There were also well-delimited regions between 4 and 7 Hz (Theta), 7 and 12 Hz, (Alpha), 12 and 30 Hz (Beta), and above 30 Hz (Gamma). The observed frequencies still agreed with the observations of Suzuki et al. [31] and Takeuchi et al. [37].

This methodology allowed us to acquire brain signals from animals that could move freely while minimizing the stress generated by confinement. It also allowed the acquisition to occur noninvasively. However, the acquisition conducted in this way produced data with a greater number of artifacts, but it had the advantage of reducing the discomfort caused by using subdermal electrodes. Another important point to consider is the challenge imposed by the size of the animals. Considering the larger amount of data, this methodology could be applied to investigate changes in brain electrical activity during animal farming, to monitor brain pathologies, and to other situations.

Table 1. Lempel and Ziv normalized complexity C(N) for different time series.

Sequence	C(N)
S1 = [0 1 0 1 0 1 0 1 0 …1 0 1 0 1 0 1 0 1 0 1]	0.0202
S2 = sin(2π30t)	0.0559
S3 = sin(2π0.5t) + sin(2π10t) + sin(2π18t) + sin(2π30t) + sin(2π30t)	0.0129
S4 = sin(2π0.5t) + sin(2π10t) + sin(2π18t) + sin(2π30t) + sin(2π30t) + (array of random numbers)	1.0000
S5 = white noise	0.9957
S6 = array of random numbers	1.0870
S7 = logistic map in chaos threshold value [53]	1.0000

lists the normalized complexity for seven known sequences.

Table 2. Normalized complexity C(N) for artifact-free signal sections.

Sequence	C(N)
Climatic chamber 1st section (size: 4 s)	0.9894
Climatic chamber 2nd section (size: 4 s)	0.6596
Climatic chamber 3rd section (size: 4 s)	0.7483
Climatic chamber 4th section (size: 4 s)	0.7475
Climatic chamber 5th section (size: 9 s)	0.5857
Pasture 1st section (size: 240 s)	0.6078
Pasture 2nd section (size: 15 s)	0.6163
Pasture 3rd section (size: 3 s)	0.8656

contains the C(N) for the artifact-free signal sections in Figure 13.

We obtained values similar to the values in Table 2 for the other five animals we evaluated. High normalized LZC C(N) values can be associated with the presence of noise, random numbers, or the predominance of one of these events (Table 1). On the basis of the values we observed for bovine signals (Table 2), we can consider that the artifacts were not completely eliminated and that cattle brain signals have some random characteristics. Costa, Tech, and Silva [53] used the logistic map in the chaotic behavior threshold to obtain the entropic q-index for different values of initial conditions and found that C(N) always presented values around one. Thus, complexity values next to one might indicate that bovine brain signals also have chaotic components.

The presented results demonstrate that it is possible to noninvasively acquire brain signals from animals that are free from sedation and have the freedom to move.

Future improvements to this method could encompass augmenting the number of animals and enhancing signal acquisition during distinct behaviors (eating, drinking water, chewing, standing, lying down, or walking) to facilitate the association of these behaviors with signal features. Moreover, it holds significance to meticulously characterize the activities undertaken by the animal, particularly during the prolonged process of rumination, during which the animal’s extended engagement can potentially impact muscular movements, consequently influencing the recorded signal. The integration of imaging techniques to monitor this activity would offer a valuable opportunity for further insight. Another point for the future is to handle artifacts automatically, given that this is a time-consuming stage of signal processing.

In terms of equipment and signal acquisition, future enhancements should consider expanding the acquisition range to enable data collection in larger areas and improving the battery autonomy for extended experiment durations. However, the central focus remains on refining the electrode fixation techniques, even accounting for complete hair removal, area cleaning, and the use of EEG conductive paste, as the utilization of surface electrodes in these longer experiments remains a challenge.

4. Conclusions

EEG signals can be acquired from unrestrained cattle by using surface (noninvasive) electrodes and telemetric equipment. There is an ideal position for electrode attachment at the front of the head so that longer artifact-free signal sections can be acquired. The signals present typical EEG frequency bands, such as the bands found in humans and other animals. The signals also display random and chaotic characteristics. As anticipated, restrained animals exhibited signals with more artifacts, as higher temperatures, such as 32 °C, lead to increased discomfort, thereby causing a rise in motion-induced interference. Additionally, the equipment yielded valuable data from cows in a paddock under moderate temperature conditions. The method developed herein can be applied to understand several situations involving bovine (or other bigger farm animals) brain signals ranging from normality to some brain pathologies.