wt_cnn: Wear Time Estimation of Raw Data from a Convolutional Neural...

In muschellij2/weartime: Wear Time Estimation from Accelerometer Data

Wear Time Estimation of Raw Data from a Convolutional Neural Network

Description

Wear Time Estimation of Raw Data from a Convolutional Neural Network

Usage

wt_cnn(
  df,
  sample_rate = NULL,
  std_threshold = 0.004,
  distance_in_min = 5L,
  episode_window_sec = 7L,
  edge = TRUE,
  start_stop_label_decision = c("and", "or"),
  min_segment_length = 1L,
  sliding_window = 1L,
  verbose = TRUE,
  model_path = NULL,
  outdir = NULL
)

Arguments

df	activity data, usually output from read.gt3x::read.gt3x(), and then imputed, or a data.frame of time/X/Y/Z
sample_rate	sample rate (integer) of the sampling frequency in Hertz from the header
std_threshold	standard deviation threshold in g
distance_in_min	merge distance to group two nearby candidate nonwear episodes
episode_window_sec	define window length to create input features for the CNN model
edge	default classification when an episode does not have a starting or stop feature window (happens at t=0 or at the end of the data)
start_stop_label_decision	logical operator to see if both sides need to be classified as non-wear time (AND) or just a single side (OR)
min_segment_length	minimum length of the segment (in minutes) to be candidate for non-wear time, coerced to integer
sliding_window	sliding window (in minutes) that will go over the acceleration data to find candidate non-wear segments, coerced to integer
verbose	print diagnostic messages
model_path	path to the model h5 file, overrides outdir and must match corresponding episode_window_sec.
outdir	output directory to download the CNN model, passed to download_cnn_model

Details

Detect candidate non-wear episodes: Perform a forward pass through the raw acceleration signal and calculate the SD for each 1-minute interval and for each individual axis. If the standard deviation is <= 0.004 g for all axes, record this 1-minute interval as a candidate non-wear interval. After all 1-minute intervals have been processed, merge consecutive 1-minute intervals into candidate non-wear episodes and record their start and stop timestamps.

Merge bordering candidate non-wear episodes: Merge candidate non-wear episodes that are no more than 5 minutes apart and record their new start and stop timestamps. This step is required to capture artificial movement that would typically break up two or more candidate non-wear episodes in close proximity.

Detect the edges of candidate non-wear episodes: Perform a backward pass with a 1-second step size through the acceleration data from the start timestamp of a candidate non-wear episode and calculate the SD for each individual axis. The same is applied for the stop timestamps with a forward pass and a step size of 1 second. If the standard deviation of all axes is <= 0.004 g, include the 1-second interval into the candidate non-wear episode and record the new start or stop timestamp. Repeat until the standard deviation of the 1-second interval does not satisfy <= 0.004 g. As a result, the resolution of the edges is now recorded on a 1-second resolution.

Classifying the start and stop windows: For each candidate non-wear episode, extract the start and stop segment with a window length of 3 seconds to create input features for the CNN classification model. For example, if a candidate non-wear episode has a start timestamp of tstart a feature matrix is created as (tstart – w , tstart) x 3 axes with w = 3 seconds, resulting in an input feature with dimensions (300 x 3) for 100Hz data. If both (i.e., logical ‘AND’) start and stop features are classified (through the CNN model) as non-wear time, the candidate non-wear episode can be considered true non-wear time. If tstart is at t = 0, or tend is at the end of the acceleration data—meaning that those candidate non-wear episodes do not have a preceding or following window to extract features from—classify the start or stop as non-wear time by default.

Value

A list of data

Examples

if (isTRUE(wt_packages_installed())) {
  path = system.file("extdata", "TAS1H30182785_2019-09-17.gt3x",
                     package = "weartime")
  df = read.gt3x::read.gt3x(path, asDataFrame = TRUE,
                            imputeZeroes = TRUE,
                            verbose = TRUE)
  model_path = system.file("extdata", "cnn_v2_7.h5", package = "weartime")
  out = wt_cnn(df, outdir = tempdir(), model_path = model_path)
}
## Not run: 

url = "https://github.com/THLfi/read.gt3x/files/3522749/GT3X%2B.01.day.gt3x.zip"
destfile = tempfile(fileext = ".zip")
dl = download.file(url, destfile = destfile)
gt3x_file = unzip(destfile, exdir = tempdir())
gt3x_file = gt3x_file[!grepl("__MACOSX", gt3x_file)]
path = gt3x_file

df = read.gt3x::read.gt3x(path, asDataFrame = TRUE,
                           imputeZeroes = TRUE,
                           verbose = TRUE)
out = wt_cnn(df, outdir = tempdir())

path = file.path(tempdir(), "AI12_NEO1F09120034_2017-09-25.gt3x.gz")
if (!file.exists(path)) {
  curl::curl_download(
    "https://ndownloader.figshare.com/files/21855675",
    destfile = path,
    quiet = FALSE)
}
df = read.gt3x::read.gt3x(path, asDataFrame = TRUE,
                           imputeZeroes = TRUE,
                           verbose = TRUE)
out = wt_cnn(df, outdir = tempdir())

## End(Not run)

muschellij2/weartime documentation built on April 5, 2025, 4:01 a.m.