triact package for R: analyzing the lying behavior of cows from accelerometer data

Importing data

The Triact R6 class encapsulates both data and all core functionality (Fig. 3). After creating a new instance of the Triact class, acceleration data of one or several cows is imported into the object with a $load_… method. Calling $load_files() processes the raw accelerometer files and imports them into the R6 object. Given the method’s arguments are appropriately specified, delimiter-separated text files with different characteristics can be imported, e.g., different separators, file header length, data-time formats, and more. Alternatively, users can process raw data on their own and import the concatenated data as a data.frame-like table using the $load_table() method. However, this requires the user to exactly prepare the data as requested (column names, data types, etc.). Imported raw data and added analyses (see next paragraph) can be accessed between any step of the workflow via the $data field.

Detecting lying and standing posture

Calling $add_… methods triggers analyses of lying behavior and the calculation of proxies for the level of physical activity. These analyses are obtained for each time point of the acceleration data and added in a new column to the tabular data in the Triact object.

The $add_lying() method performs the classification into lying and standing. Standing is defined as “not lying” and includes walking and other activities in upright posture. The two postures lying and standing are thus mutually exclusive and collectively exhaustive. The simple rule-based algorithm is composed of three steps and uses only information from the upward axis (Fig. 4). In the first step, the upward acceleration is filtered to obtain the gravity component of the signal, by default by using a moving median filter with a window size of 10 s (Fig 4B). In the second step, a threshold (crit_lie; Fig. 4B) is used to classify the filtered acceleration values into standing (>0.5 g) and lying (<0.5 g). The default threshold of 0.5 g corresponds to a sensor tilt or angulation of the leg of about arccos(0.5/1 g) ≙ 60° (Fig. 5). Finally, in the third step, lying bouts shorter than 30 s are “removed” by reclassifying them as standing. This step aims to remove falsely detected short lying bouts (false positives).

The first two steps of the algorithm are sufficient to reliably detect lying bouts of ~2 min and longer. In fact, virtually perfect accuracy in detecting these bouts is expected, with uncertainties similar to direct observation by a human observer (a few seconds at most). The total daily lying duration can therefore be expected to be highly accurate, as the contribution of the rare and less reliably detectable very short lying bouts (a minutes or less) to the total lying duration is typically negligible (Tucker et al., 2021). Detecting such short lying bouts with limited reliability can however significantly compromise accuracy of the daily number of lying bouts (and consequently also the number of standing bouts). Challenging is the avoidance of short false positive lying periods corresponding to the cow standing but lifting the hind leg on which the accelerometer is attached toward the horizontal position for several seconds to up to half a minute or more (Fig. S1; Videos S1 and S2). Typically, this occurs in self-grooming behaviors. Because true short lying periods of less than a minute are uncommon when the cow is not forced to do so, it is reasonable to set a minimal duration for a lying period and correct accordingly (third step in our algorithm). The IceTag sensors with proprietary algorithm apparently suffer from the same challenges, do not (sufficiently) address this issue, and manual exclusion of short lying bouts post hoc is common when results of these sensors are used in scientific context (e.g., Hendriks et al., 2020b; Tolkamp et al., 2010). In a study with 28 lactating dairy cows, Kok et al. (2015) reported the shortest true lying bout to be 33 s long and that correcting the IceTag sensor output using 33 s as a minimum duration maximized accuracy. As this agrees well with our experience that true lying bouts shorter than 30 s and lifting the hind leg in standing posture longer than 30 s are both very rare (not systematically studied), this seems to be a threshold that adequately balances the tradeoff between true positive and false negative lying bouts. However, although very rare, lying bouts can be much shorter than the threshold—on the pasture we observed lying bouts as short as ~8 s (Fig. S2; Video S4). Very short standing bouts do occur as well, typically to change lying side (Fig. S3; Videos S5 and S6). Because of physical constraints of how the cow can place the hind legs while lying, false standing bouts in actually lying cows are not expected and have never been observed by us when using the default parameter of triact and acceleration data with sampling frequency of ≥1 Hz. Therefore, it is no surprise that the IceTag sensors with their proprietary algorithms are also reported very reliable when it comes to short standing bouts (Tolkamp et al., 2010).

For users of the triact R package analyzing data sampled at recommended frequencies of ≥1 Hz, we expect good results when using the default parameters of the $add_lying() method. However, we encourage users to experiment with the many parameters in order to increase understanding, particularly in potential validation studies or for “off-label” use with animals other than cows. Besides the moving median filter (default), a bidirectional (zero-lag) Butterworth low-pass filter can be selected to filter for the gravity component in the first step of the algorithm (parameter: filter_method; see Winter (2009) for an explanation of zero-lag Butterworth filter). In the second step, the threshold for binary classification can be specified (crit_lie). In the last step, the minimum duration of lying bouts can be adjusted (minimum_duration_lying). A minimum duration of standing bouts can be specified as well. However, this is not recommended (see above) and not used by default. The filtering methods available for the first step can additionally be fine-tuned if desired. Using median filter (the default), the window size in seconds (window_size) can be adjusted. For the Butterworth filter, both the order (default is first order) and the cutoff frequency (default is 0.1 Hz) can be defined by the user. The default cutoff of 0.1 Hz we found suitable for low-pass filtering to obtain the gravity component is in a similar range as the 0.3 Hz used for high-pass filtering to obtain the body acceleration in studies on dairy cow behaviors by Riaboff et al. (2019) and Smith et al. (2016). Cutoff frequencies between 0.1 and 0.5 Hz are also commonly used in the literature to separate static gravitational from dynamic body acceleration in studies on human posture and movements (Bayat, Pomplun & Tran, 2014; Foerster & Fahrenberg, 2000; Veltink et al., 1996).

Detecting lying laterality

The $add_side() method performs the determination of lying laterality (left or right) based on the information of the right axis (Figs. 2, 4C). For each lying bout detected with $add_lying(), one and the same lying side is assigned for the entire duration of the lying bout. This is justified because change of lying side occurs via a short standing bout, which consequently splits the lying bout in two lying bouts. Furthermore, short standing bouts are reliably detected, even when switching side by standing with only the hind legs fully extended but the front legs resting on the carpal joints (Fig. S3; Video S5). The simple one-step algorithm determines whether the majority of the acceleration values measured during that bout are above (left lying side) or below (right lying side) a threshold (crit_left; Fig. 4C). This is equivalent to comparing the median of the acceleration values with the threshold. The threshold depends on which hind leg the accelerometer was mounted on. It is by default 0.5 g in case of the left hind leg, and −0.5 g in case of the right hind leg. When working with this default value, the user must thus specify which hind leg was used (parameter: left_leg). Alternatively, the threshold can be freely adjusted (crit_left).

Assuming the cow would lay down the leg with the accelerometer in the coronal plane and perfectly horizontal, one would expect that the right axis will detect the gravity component of around ±1 g, with a positive mathematical sign (along its direction) when lying on the left side and a negative mathematical sign (opposing its direction) when lying on the right side. Thus, under this (miss) assumption a threshold of 0 would be a logical choice. However, the above assumptions are poor in situations where the cow is not lying on the side where the accelerometer is attached. Compared to the leg on which the cow is lying, the other hind leg has more freedom of movement and can be placed bent to different degree in the direction of the cow’s head (Fig. 1). This can result in the gravity component partially to nearly completely loading on the forward axis (and not the right axis) resulting in values on the right axis approaching 0. It is therefore necessary to shift the threshold more towards the values expected for the situation with the accelerometer on the lying side, i.e., if the accelerometer is on the left leg towards positive (therefore 0.5), if on the right leg towards negative (therefore −0.5). Figure 4, S2 and S3 exemplarily show that when the cow lies on the left side where the accelerometer was mounted, values tend to be around 1 quite consistently, while when it lies on the right side, the central tendency of values can be quite different from closer to −1 (Fig. 4) to close to around 0 (Fig. S3).

Calculating proxies for physical activity

The $add_activity() method allows calculating proxies for the physical activity level based on the information of all three accelerometer axes. By default, the L2 norm of the dynamic body acceleration (DBA) vector is calculated. The corresponding L1 norm is optionally available. Also, the L1 and L2 norms of the jerk vector can be calculated (Table 1). During lying bouts, all activity values are “adjusted” to 0, i.e., periods when cows are lying are considered as “inactive” by definition (Fig. 6).

DBA is obtained by subtracting the static gravity component from the measured acceleration. The static component is obtained by filtering as described above for the lying behavior (same defaults, same options). Jerk, the change in acceleration and therefore its first derivative, is approximated through the backward finite difference method, i.e., difference in acceleration to one time point back divided by time interval (sampling interval respectively). For a vectorial property $\overrightarrow{v}$ (DBA or jerk) at time point $t_i$, the L1 norm is the sum of absolute values in each axis and the L2 norm is the square root of the sum of the squared value in each axis:

Summarizing lying behavior and activity

The results previously added to the Triact object with the $add_… methods can be summarized using the $summarize_bouts() and the $summarize_intervals() methods. The returned tables are ready to be subjected to downstream statistical analyses (not part of the triact package). With $summarize_bouts(), a summary is created for the individual lying and standing bouts, with duration, mean activity, lying side (for a lying bout), and more. With $summarize_intervals(), the summary is obtained per regular intervals (e.g., a day or an hour). Examples of summary measures are the total duration of lying, the duration of lying on left or right side, the number of lying bouts, the mean duration of lying bouts, and the mean activity. For measures such as the number of lying bouts or mean lying bout duration, a weighted mean is calculated with the weights being the proportion of the individual bout overlapping with the respective interval. For example, a bout completely in the interval contributes with weight 1 vs. a bout covered by the interval only by 30% contributes with weight 0.3 (and with 0.7 to the adjacent interval). This weighting ensures that the totals are not falsely inflated. For both $summarize_… methods, we strictly suppress output (return NA) that depends on ill-defined information of the first and the last bouts (e.g., duration), which are only partly covered by the data. However, the user can consciously override this behavior and force a summary in any case (parameter calc_for_incomplete)—first and last bouts in the data are thereby simply assumed complete.

Extracting posture transitions

Using $extract_liedown() and $extract_standup(), the raw acceleration data (and added analyses) of the posture transitions, i.e., lying-to-standing and standing-to-lying, can be extracted. The data between user-specified times before and after the exact moment of posture transition as detected by triact are returned (parameters sec_before, sec_after). Among other things, this can be useful in developing algorithms able to characterize these posture transitions. For example, Brouwers et al. (2023) used the $extract_… methods to extract regions of interest as a first step in a machine learning–based workflow to detect atypical behaviors performed during lying down and standing up movement patterns.