Objective assessment of activity via accelerometry can provide valuable insights into dog health and welfare. Common activity metrics involve using acceleration cut-points to group data into intensity categories and reporting the time spent in each category. Lack of consistency and transparency in cut-point derivation makes it difficult to compare findings between studies. We present an alternative metric for use in dogs: the acceleration threshold (as a fraction of standard gravity,1g = 9.81m/s²) above which the animal’s X most active minutes are accumulated (MX_ACC) over a 24-hour period. We report M2_ACC, M30_ACC and M60_ACC data from a colony of healthy beagles (n=6) aged 3-13 months. To ensure that reference values are applicable across a wider dog population, we incorporated labelled data from beagles and volunteer pet dogs (n=16) of a variety of ages and breeds. The dogs’ normal activity patterns were recorded at 200 Hz for 24-hours using collar-based Axivity-AX3 accelerometers. We calculated acceleration vector magnitude and MX_ACC metrics. Using labelled data from both beagles and pet dogs, we characterise the range of acceleration outputs exhibited for a variety of behaviours, enabling meaningful interpretation of MX_ACC. These metrics will help standardise measurement of canine activity, inform development of exercise guidelines and adherence monitoring, and serve as outcome measures for veterinary and translational research. This repository contains two files: 1) `LMM_input_data.csv`, a spreadsheet file containing the raw data used to explore the effect of modifying epoch length and sample frequency on aggregate activity metrics using linear mixed effects models as described in the manuscript. There is one row per dog per epoch length, sample frequency and age combination, and 2) `getMostActiveMinsThresh.m`, a MATLAB function used to compute the MX ACC outcome metrics that are explored in the manuscript.

See detailed methods in accompanying paper, Karimjee et al. 2024, Royal Society Open Science.  The Methods description below are exerpts directly from the methods paper.  See the paper for more details including the equations, tables and figures. 

Data were collected via Axivity AX3 activity monitors (Axivity, U.K.). These have been previously validated for use in dogs and their parameters are easily modifiable by the user through open-source software, unlike many other commercial devices. Devices were mounted onto each dog’s neck collar using duct tape, and positioned ventrally to minimise likelihood of rotation of the logger over time. Additionally, the metrics selected were direction invariant to avoid any influence of device rotation. A sample frequency of 200Hz was used for recording. This was a higher sample frequency in comparison to most industry standard devices (30-100Hz). Considering that one of the main objectives of this work was to quantify the effect of sample frequency on output metrics, we aimed to use a sample frequency that far exceeded that necessary to allow for meaningful comparisons to be carried out by downsampling to lower sample frequencies that are commonly used in industry devices. Using a higher sample rate also ensured that the loggers captured all high frequency activity and behaviour data.

Dog Colony

The loggers were mounted onto the dogs’ collars before 12pm and removed after 12pm, 2 days later. Dogs were habituated to wearing collars from 7 weeks of age. Dogs were sampled at monthly intervals between the ages of 3 and 18 months old and data were recorded for 48 hours continuously. The dogs remained in their kennels during the period of monitoring and followed their normal daily routine. Paddock exercise was not provided during the monitoring period, however all kennels included an outdoor courtyard area which was always accessible. To develop aggregate metrics to represent activity over a 24-hour period, we split the 48-hour recording into 2x 24-hour segments. Each 24-hour segment was carried forward for analysis and calculation of aggregate metrics. The resulting metrics were then averaged across the 2x 24-hour periods per time point for each animal. Using an average of the output metrics from 2x 24-hour segments, we were able to establish a more representative measurement of that dog at a particular age point.

In addition to the data described above, we also collected a smaller sample of video and accompanying accelerometer data (approx. 43 hours across multiple recording sessions). The videos were collected at the end of the 48-hour segments within the kennel environment. Video data were recorded using GoPro Hero 5 black cameras and a gooseneck monopod with clamp attachment. Video was captured in 1080p and at 30 fps using the “Superview” mode. This ensured that the whole kennel was visible with a single camera.

Upon starting recording, the collar was shown to the camera and visibly tapped 3x using a solid object (e.g. pen) in order to create spikes in the accelerometer data whilst within view of the camera. These spikes were used to synchronise the video and accelerometer data during labelling. The collars were then placed on the dogs and the start and end times were logged. The dogs were left undisturbed for a period of between 60-120 minutes.

Volunteer Pet Dogs

We carried out 10-minute sessions of behavioural observation whilst the dog wore the AX3 logger. This observation period followed a protocol to encourage a range of behaviours in the dogs without specific commands to maintain an environment comparable to free-living. Dogs were put into a secure, enclosed pen (3m x 2.5m) and were under video surveillance throughout the experiment. The video footage was live streamed to a screen outside the room to allow constant observation. Dog owners were able to watch the live footage and could decide to end the experiment early if they deemed their dog to be under significant stress due to being left alone in an unfamiliar environment. One experiment was terminated early due to anxiety and that dog’s data has been excluded from the analysis.

Active threshold derivation for epoch length analysis

We have explored the derivation of two key processing parameters: (1) derivation of an appropriate active/inactive threshold and (2) establishment of an appropriate epoch length for our dog data. However, each analysis required the use of both parameters. We chose to begin with the derivation of active/inactive threshold to quantify bouts of activity, which meant that an initial epoch length was required to process our data. The derived threshold was then taken forward and used to analyse potential epoch lengths. Lastly, we repeated the threshold derivation analysis with data processed using the newly established epoch length to confirm that the threshold remained appropriate. Labelled data were imported into MATLAB R2021a for processing. A 6^th order Butterworth band-pass filter was used in both directions (‘filtfilt’ function) to obtain zero phase lag, with cut-off frequencies of 0.28 Hz and 32.76Hz (stop band: 0.1Hz and 50Hz, pass band: 0.5Hz and 20Hz). The filter settings were chosen to both remove the acceleration due to gravity and any background noise. We also anticipated that behaviours of interest would all occur at frequencies between the cut off frequencies listed. The labelled, filtered tri-axial data were used to compute the vector magnitude of acceleration. This allowed for the magnitude of the acceleration captured in each axis to be quantified, independent of direction. The labelled vector magnitude data were then split into 1s epochs and each segment was assigned an “active” or “inactive” label. The mean was calculated for that segment. The 1s epoch duration is commonly used in industry and served as an initial starting point for derivation of activity threshold. In the cases where more than one label was present across a single 1s segment, the label corresponding to > 50% of the points was assigned to the whole segment. If there was no label that was assigned to >50% of points, this segment was discarded and not taken forward for analysis.

The labelled data were then split into active and stationary behaviours and exported to Python 3 for further analysis. Data were split into a training and test fractions (80:20) and a Receiver-operating characteristic (ROC) analysis was carried out. ROC curves visualise the trade-off between sensitivity and specificity for all possible cut-off thresholds for a test, to enable selection of an optimum threshold that maximises both sensitivity and specificity. The area under the curve (AUC) also provides a numerical measure of usefulness of a particular test. We iterated through thresholds between 0.001g and 0.5g in increments of 0.001g. We computed the differences between sensitivity and specificity values per threshold and selected the threshold with the minimum difference between the two quantities. The resulting threshold for the beagle data was 0.154g, and for the pet dogs was 0.159g.

Evaluating the effect of epoch length on activity measures

Once data are collected, smoothing by aggregating the data over a time window of fixed length, or “epoch”, is a tool that can be used to further reduce computational load whilst providing a more reliable estimate of the acceleration frequency. Epoch length refers to a time window of fixed length across which data points are aggregated during a recording. This is particularly helpful for devices with lower memory capacity, which is often the case in wearable devices due to their extremely small size. Epochs are used for several reasons including reducing computational load whilst increasing reliability by averaging data points over a time period, versus simply reducing the sample frequency. The choice of epoch length should be a decision made in the context of the data collected. Given that we are investigating activity patterns, including counts and durations of bouts of activity, the choice epoch length was an important consideration. Here, we adopted a data-driven approach to explore how epoch length influenced the measured outcome variables.

To determine the appropriate epoch length for our dog data, we processed 24-hour activity data using different epoch lengths between 0.005s and 1s, in 0.005s increments. The increment size was chosen as data were collected at 200Hz, meaning that the changes were as granular as possible, incrementing by a single sample period each time. A 6^th order Butterworth band-pass filter was then used in both directions (‘filtfilt’ function) to obtain zero phase lag, with cut-off frequencies of 0.28 Hz and 32.76Hz (stop band: 0.1Hz and 50Hz, pass band: 0.5Hz and 20Hz).

Derivation of MX_ACC metrics

We report an activity metric first presented in humans by Rowlands and colleagues (2019), for activity monitoring in dogs: the acceleration threshold (g) above which the animal’s X most active minutes are accumulated (MX_ACC) over a 24-hour time period. Examining the threshold above which the dog’s X most active minutes are accumulated over a 24-hour period provides an understanding of the varying intensity levels of movement that are exhibited over a longer time frame, and how individuals compare to each other. The values reported are directly comparable between individuals and, importantly, are translatable between populations. The values of X used can be adjusted to suit the data that is being examined, with smaller values reflecting higher intensity activities and larger values reflecting lower intensity activities. In this case study, we present MX_ACC data computed using X-minute values of 2, 30 and 60. To calculate these data points, the epoched vector magnitude was sorted in descending order and the appropriate X-minute increment was indexed. The vector magnitude value at the X time index in the sorted data is the MX_ACC statistic.

Rowlands AV, Sherar LB, Fairclough SJ, Yates T, Edwardson CL, Harrington DM, et al. (2019) A data-driven, meaningful, easy to interpret, standardised accelerometer outcome variable for global surveillance. Journal of Science and Medicine in Sport. 22(10):1132-8.