Analyze data with PAAT

import h5py
import paat
import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
import matplotlib.dates as mdates
from matplotlib.patches import Rectangle, Patch
%matplotlib inline

plt.rcParams['figure.figsize'] = [15, 3]
plt.rcParams['font.size'] = '13'

Load the example data

Normally, the paat.io.read_gt3x would be used to load the acceleration data. However, to save storage space and runtime, we provide a gunzipped CSV file with an already calibrated one day recording to visualize the analysis options currently available in PAAT.

data, sample_freq = pd.read_csv("data/example_day_calibrated.csv.tar.gz", compression='gzip'), 100
data["Timestamp"] = pd.to_datetime(data["Timestamp"])
data = data.set_index("Timestamp")

Alternatively, you can load and directly recalibrate the data with actipy:

data, info = actipy.read_device(
    "path/to/file.gt3x",
    calibrate_gravity=True,
    detect_nonwear=False,
    lowpass_hz=False,
    resample_hz=False
)
sample_freq = info["SampleRate"]

Visualize the raw data

It is often good, when developing analysis scripts, to have sneak peeks into the data to get a feeling for the quality and what to expect. There is no dedicated function for this in PAAT as what you are looking for might severly vary between applications and research questions. One of the easiest ways to visualize the data is using Seaborn. Seaborn provides a simple, but yet powerful, interface to the way more versatile Matplotlib library. To visualize the data, we recommend to resample it to 1min or 1s resolution as plotting otherwise takes ages. However, certain applications might require higher resolutions.

agg = data.resample("1s").mean()

ax = sns.lineplot(data=agg[['X', 'Y', 'Z']], legend=True)

While this plot already provides some insights, we recoomend to pretify the plot a little bit to facilitate readability:

agg = data.resample("1s").mean()

fig, ax = plt.subplots(figsize=(12,3))
ax = sns.lineplot(data=agg[['X', 'Y', 'Z']], ax=ax, legend=True)

ax.xaxis.set_major_formatter(mdates.DateFormatter('%H:%M'))
ax.set_xlabel("Time")
ax.set_ylabel("Acceleration [g]")

plt.tight_layout()
print('')

Detect non-wear periods

Different methods to infer non-wear time from the raw acceleration signal are implemented in the paat.wear_time module. We suggest to use paat.wear_time.detect_non_wear_time_hees2011 published by Van Hees et al. (2011,2013) as it is the most validated algorithm. This function can also be called from the top-level:

data.loc[:, "Non Wear Time"] = paat.detect_non_wear_time_hees2011(data, sample_freq)

However, there are also other non-wear time algorithm implemented in the paat.wear_time module. For instance, the CNN-based NWT method by Syed et al. (2021) is also implemented and can be used by calling:

data.loc[:, "Non Wear Time Syed"] = paat.detect_non_wear_time_syed2021(data, sample_freq)

Finally, also a very naive NWT algorithm is implemented in paat.wear_time.detect_non_wear_time_naive. For a comparison of the different algorithms see also Syed et al. (2020).

data.loc[:, "Non Wear Time Naive"] = paat.detect_non_wear_time_naive(
    data,
    sample_freq,
    std_threshold=.003,
    min_interval=5
)

Identify Time in Bed periods

Time in Bed identification is implemented in the paat.detect_time_in_bed_weitz2024 which is based on a bidirectional LSTM models as described by Weitz et al. (2025). Time in bed is detected on a 1min resolution with averaged acceleration values per minute, so it is also possible to apply this algorithm after the data has been resampled. However, note that in this case, you have to adjust the sampling frequency accordingly (e.g., if you have resampled to 1s, then the sample frequency should be 1, if you have resampled to 1min then the sample frequency should be 1/60, etc.).

data.loc[:, "Time in bed"] = paat.detect_time_in_bed_weitz2024(data, sample_freq)

Estimate physical activity and sedentary behavior

The most common approach to analyze physical activity data is the use of cutpoints. There are various published cutpoints for different demographical groups. In this example, we use the cutpoints proposed by Sanders et al. (2019).

data.loc[:, ["MVPA", "SB"]] = paat.calculate_pa_levels(
    data,
    sample_freq,
    mvpa_cutpoint=0.069,
    sb_cutpoint=0.015
)

Combine the estimates into one column

To simplify the data, paat has a create_activity_column functions, which creates one activity column from the different columns. The importance of column is the revised order of the columns argument. So in this example, first NWT is marked. Then everything which was not marked as NWT is marked as time in bed, and so on.

data.loc[:, "Activity"] = paat.create_activity_column(
    data,
    columns=["SB", "MVPA", "Time in bed", "Non Wear Time"]
)
data =  data[["X", "Y", "Z", "Activity"]]
data["ENMO"] = paat.calculate_enmo(data)

Visualize and analyze the results

Finally, it can be helpful to visually investigate the results. Below a simple example is provided how the resulting data frame can be visualized. However, note that especially with larger datasets visualizing this data on 1s resolution takes a considerable amount of time.

As physical activity, in this example, is dependent on the ENMO of the acceleration signal, it can be useful to plot the resulting estimates against the ENMO:

agg = data.resample("1s").apply({
    "X": "mean",
    "Y": "mean",
    "Z": "mean",
    "ENMO": "mean",
    "Activity": pd.Series.mode
})

COLOR = {
    'Non Wear Time': "grey",
    'Time in bed': "blue",
    'SB': "green",
    'LPA': "yellow",
    'MVPA': "red"
}

fig, axes = plt.subplots(2, 1, sharex=True, figsize=(12,6.5))

ax = axes[0]
ax = sns.lineplot(data=agg[['X', 'Y', 'Z']], ax=ax, legend=True)
ax.set_ylabel("Acceleration [g]")

ax = axes[1]
ax = sns.lineplot(data=agg['ENMO'], ax=ax, legend=True)

# Add background
ymin, ymax = ax.get_ylim()
height = ymax - ymin
for ii, row in agg.iterrows():
    ax.add_patch(
        Rectangle(
            (ii, ymin),
            pd.Timedelta("1s"),
            height,
            alpha=.2,
            facecolor=COLOR[row["Activity"]]
        )
    )


ax.xaxis.set_major_formatter(mdates.DateFormatter('%H:%M'))
ax.set_xlabel("Time")
ax.set_ylabel("ENMO [g]")

handles = [Patch(color=value, label=key, alpha=.2) for key, value in COLOR.items()]
ax.legend(
    handles=handles,
    loc='upper center',
    bbox_to_anchor=(0.5, -.4 * ymax),
    fancybox=False,
    shadow=False,
    ncol=5)

plt.tight_layout()
plt.subplots_adjust(wspace=0, hspace=0)
print('')

References

Sanders, G. J., Boddy, L. M., Sparks, S. A., Curry, W. B., Roe, B., Kaehne, A., & Fairclough, S. J. (2019). Evaluation of wrist and hip sedentary behaviour and moderate-to-vigorous physical activity raw acceleration cutpoints in older adults. Journal of Sports Sciences, 37(11), 1270–1279. https://doi.org/10.1080/02640414.2018.1555904

Syed, S., Morseth, B., Hopstock, L. A., & Horsch, A. (2020). Evaluating the performance of raw and epoch non-wear algorithms using multiple accelerometers and electrocardiogram recordings. Scientific Reports, 10(1), 1. https://doi.org/10.1038/s41598-020-62821-2

Syed, S., Morseth, B., Hopstock, L. A., & Horsch, A. (2021). A novel algorithm to detect non-wear time from raw accelerometer data using deep convolutional neural networks. Scientific Reports, 11(1), 8832. https://doi.org/10.1038/s41598-021-87757-z

Van Hees, V. T., Gorzelniak, L., León, E. C. D., Eder, M., Pias, M., Taherian, S., Ekelund, U., Renström, F., Franks, P. W., Horsch, A., & Brage, S. (2013). Separating Movement and Gravity Components in an Acceleration Signal and Implications for the Assessment of Human Daily Physical Activity. PLOS ONE, 8(4), e61691. https://doi.org/10.1371/journal.pone.0061691

Van Hees, V. T., Renström, F., Wright, A., Gradmark, A., Catt, M., Chen, K. Y., Löf, M., Bluck, L., Pomeroy, J., Wareham, N. J., Ekelund, U., Brage, S., & Franks, P. W. (2011). Estimation of Daily Energy Expenditure in Pregnant and Non-Pregnant Women Using a Wrist-Worn Tri-Axial Accelerometer. PLOS ONE, 6(7), 7. https://doi.org/10.1371/journal.pone.0022922

Weitz, M., Syed, S., Hopstock, L. A., Morseth, B., Henriksen, A., & Horsch, A. (2025). Automatic time in bed detection from hip-worn accelerometers for large epidemiological studies: The Tromsø Study. PLOS ONE, 20(5), e0321558. https://doi.org/10.1371/journal.pone.0321558