IMU Activity Analysis (skdh.activity)#
Pipeline Activity Processing#
|
Classify accelerometer data into different activity levels as a proxy for assessing physical activity energy expenditure (PAEE). |
Activity Endpoints#
|
Base class for activity endpoints. |
|
Compute the gradient of the acceleration movement intensity. |
|
Compute the maximum acceleration over windows of the specified length. |
|
Compute the total time spent in an intensity level. |
|
Compute the time spent in bouts of intensity levels. |
|
Compute fragmentation metrics for the desired intensity level. |
Compute the threshold such that the bouts are of equal duration on either side of the threshold. |
|
|
Compute various signal features on the raw acceleration metric. |
Accelerometer Metrics#
|
Compute the euclidean norm. |
|
Compute the euclidean norm minus 1. |
|
Compute the band-pass filtered euclidean norm. |
|
Compute the high-pass filtered euclidean norm. |
|
Compute the high-pass filtered euclidean norm plus the low-pass filtered euclidean norm minus 1g. |
|
Compute the Mean Amplitude Deviation metric for acceleration. |
Background Information#
Activity level classification is the process of using accelerometer metrics to derive estimates of the time spend in different energy expenditure states, typically classified by Metabolic Equivalent of Task (MET). The MET is the rate of energy expenditure for an individual of a certain mass performing physical activities, relative to a baseline. The different classifications are typically
sedentary
light (< 3 MET)
moderate (3-6 MET)
vigorous (> 6 MET)
Different research has yielded different methods of estimating these thresholds, with different acceleration metrics and cutpoints for those metrics. The ones available as part of Scikit-Digital-Health are the following:
"esliger_lwrist_adult"[1] Note that these use light and moderate thresholds of4 and 7 METs
"esliger_rwirst_adult"[1] Note that these use light and moderate thresholds of4 and 7 METs
"esliger_lumbar_adult"[1] Note that these use light and moderate thresholds of4 and 7 METs
"schaefer_ndomwrist_child6-11"[2]"phillips_rwrist_child8-14"[3]"phillips_lwrist_child8-14"[3]"phillips_hip_child8-14"[3]"vaha-ypya_hip_adult"[4] Note that this uses the MAD metric, and originallyused 6 second long windows
"migueles_wrist_adult"[7] these are the default cutpoints used
The thresholds have been automatically scaled to the average values, and can be used with any length windows (though most originally use 1s windows), and use the appropriate acceleration metric.
Adding Custom Endpoints#
Custom endpoints are simple to add - each custom endpoint should be a subclass
of the ActivityEndpoint. Endpoint classes can generate multiple endpoints
at once, but the names for all of the endpoints need to be specified in the custom
class, as this is how the results dictionary is populated.
from skdh.activity import ActivityEndpoint
class CustomEndpointSingle(ActivityEndpoint):
def __init__(self, arg1, arg2=None, state='wake'):
super().__init__("custom endpoint name", state)
self.arg1 = arg1
self.arg2 = arg2
def predict(self, results, i, accel_metric, epoch_s, epochs_per_min, **kwargs):
super().predict()
# desired processing
# save the results
results[self.name][i] = custom_endpoint_res
class CustomEndpointMultiple(ActivityEndpoint):
def __init__(self, arg1, state='wake'):
super().__init__(
[
"custom ept 1",
"custom ept 2",
],
state
)
The only required parameter for the custom endpoint class __init__ is state, which should be set when initialized to the state in which the endpoint should be calculated. The .predict method is run for every block of wear data during state, meaning it could potentially get run multiple times during the same day. For some custom endpoints, this may not be a problem, however if one value needs to be calculated on all the data for the day, the state of the class is kept/left alone between runs. This leads to the below example, which saves several values between runs until the full day is done (at which point .reset_cache() is called:
class IntensityGradient(ActivityEndpoint):
def __init__(self, state="wake"):
super(IntensityGradient, self).__init__(
["intensity gradient", "ig intercept", "ig r-squared"], state
)
# default from rowlands
self.ig_levels = (
array([i for i in range(0, 4001, 25)] + [8000], dtype="float") / 1000
)
self.ig_vals = (self.ig_levels[1:] + self.ig_levels[:-1]) / 2
# values that need to be cached and stored between runs
self.hist = zeros(self.ig_vals.size)
self.ig = None
self.ig_int = None
self.ig_r = None
self.i = None
def predict(self, results, i, accel_metric, epoch_s, epochs_per_min, **kwargs):
super(IntensityGradient, self).predict()
# get the counts in number of minutes in each intensity bin
self.hist += (
histogram(accel_metric, bins=self.ig_levels, density=False)[0]
/ epochs_per_min
)
# get pointers to the intensity gradient results
self.ig = results[self.name[0]]
self.ig_int = results[self.name[1]]
self.ig_r = results[self.name[2]]
self.i = i
def reset_cached(self):
super(IntensityGradient, self).reset_cached()
# make sure we have results locations to set
if all([i is not None for i in [self.ig, self.ig_int, self.ig_r, self.i]]):
# compute the results
# convert back to mg to match existing work
lx = log(self.ig_vals[self.hist > 0] * 1000)
ly = log(self.hist[self.hist > 0])
if ly.size <= 1:
slope = intercept = rval = nan
else:
slope, intercept, rval, *_ = linregress(lx, ly)
# set the results values
self.ig[self.i] = slope
self.ig_int[self.i] = intercept
self.ig_r[self.i] = rval ** 2
# reset the histogram counts to 0, and results to None
self.hist = zeros(self.ig_vals.size)
self.ig = None
self.ig_int = None
self.ig_r = None
self.i = None
Using Custom Cutpoints/Metrics#
If you would like to use your own custom cutpoints/metric, they can be supplied in a dictionary as follows:
from skdh.activity import ActivityLevelClassification
from skdh.utility import rolling_mean
def metric_fn(accel, wlen, \*args, \*\*kwargs):
# compute acceleration metric for non-overlapping windows of length wlen
metric = compute_metric()
return rolling_mean(metric, wlen, wlen)
custom_cutpoints = {
"metric": metric_fn, # function handle
"kwargs": {"metric_fn_kwarg1": value1},
"sedentary": sedentary_max, # maximum value for sedentary (min value for light)
"light": light_max, # maximum value for light acvitity (min value for moderate)
"moderate": moderate_max # max value for moderate (min value for vigorous)
}
mvpa = ActivityLevelClassification(cutpoints=custom_cutpoints)
References#
D. W. Esliger, A. V. Rowlands, T. L. Hurst, M. Catt, P. Murray, and R. G. Eston, “Validation of the GENEA Accelerometer,” Medicine & Science in Sports & Exercise, vol. 43, no. 6, pp. 1085–1093, Jun. 2011, doi: 10.1249/MSS.0b013e31820513be.
C. A. Schaefer, C. R. Nigg, J. O. Hill, L. A. Brink, and R. C. Browning, “Establishing and Evaluating Wrist Cutpoints for the GENEActiv Accelerometer in Youth,” Med Sci Sports Exerc, vol. 46, no. 4, pp. 826–833, Apr. 2014, doi: 10.1249/MSS.0000000000000150.
L. R. S. Phillips, G. Parfitt, and A. V. Rowlands, “Calibration of the GENEA accelerometer for assessment of physical activity intensity in children,” Journal of Science and Medicine in Sport, vol. 16, no. 2, pp. 124–128, Mar. 2013, doi: 10.1016/j.jsams.2012.05.013.
H. Vähä-Ypyä et al., “Validation of Cut-Points for Evaluating the Intensity of Physical Activity with Accelerometry-Based Mean Amplitude Deviation (MAD),” PLOS ONE, vol. 10, no. 8, p. e0134813, Aug. 2015, doi: 10.1371/journal.pone.0134813.
M. Hildebrand, V. T. Van Hees, B. H. Hansen, and U. Ekelund, “Age Group Comparability of Raw Accelerometer Output from Wrist- and Hip-Worn Monitors,” Medicine & Science in Sports & Exercise, vol. 46, no. 9, pp. 1816–1824, Sep. 2014, doi: 10.1249/MSS.0000000000000289.
M. Hildebrand, B. H. Hansen, V. T. van Hees, and U. Ekelund, “Evaluation of raw acceleration sedentary thresholds in children and adults,” Scandinavian Journal of Medicine & Science in Sports, vol. 27, no. 12, pp. 1814–1823, 2017, doi: https://doi.org/10.1111/sms.12795.
J. H. Migueles et al., “Comparability of accelerometer signal aggregation metrics across placements and dominant wrist cut points for the assessment of physical activity in adults,” Scientific Reports, vol. 9, no. 1, Art. no. 1, Dec. 2019, doi: 10.1038/s41598-019-54267-y.