Skip to main content Link Search Menu Expand Document (external link)

Datasets and Models

We developed two deep-learned posture classifiers, one for adults aged 35 years and older, a second for children aged 8 - 11 years. These models were named CHAP-ADULT and CHAP-CHILD, respectively.

Table of contents

  1. Datasets
    1. About
      1. Adult Dataset
      2. Child Dataset
    2. Summary
    3. Data format
  2. Pre-trained Models
    1. Additional Models



Adult Dataset

Data for the CHAP-ADULT model came from two distinct studies: the Australian Diabetes, Obesity, and Lifestyle study (AusDiab) and the Adult Changes in Thought (ACT) study. AusDiab is a population-based epidemiologic cohort, initiated in 1999, that enrolled adults aged 25 years and older throughout Australia; the data for our study comprised a sub-sample of 1,014 ambulatory community-living participants who in 2011 were invited to join an ancillary activity monitor study. Participants were asked to wear the activPAL3TM (thigh) and the Actigraph GT3X+ (hip) for 7 consecutive days, with the ActiGraph removed for sleeping and the activPAL worn continuously. ACT is an ongoing longitudinal cohort study, initiated in 1994, investigating risk factors for development of dementia and a wide range of cognitive and noncognitive factors of healthy aging. ACT recruited dementia-free adults age 65 years and older from random samples of the membership of Group Health Cooperative (now Kaiser Permanente Washington) in King County, Washington, USA . Starting in 2016, ACT participants were invited to participate in an activity monitoring sub-study (ACT-AM). Those who were eligible (ambulatory, no recent critical illness diagnosis) and provided consent to participate were invited to wear an ActiGraph GT3X+ accelerometer on the hip, an activPAL micro3 on the thigh, or, if willing, both simultaneously, continuously for 7 consecutive days. The sample of 1001 ACT participants who agreed to wear both devices comprise the dataset for our study. Protocols for both studies were approved by their respective institutional ethics and review boards. To develop and test the posture classification models, we leveraged data from 1397 participants (n=688 AusDiab; n=709 ACT) of these two cohort studies who concurrently wore and had valid data on ActiGraph and activPAL accelerometers for up to 7 days while going about their usual (i.e., free-living) behavior pattern. Models were developed to predict sitting versus not-sitting for the Actigraph time series outputs, using the activPAL posture labels (sitting versus not-sitting) as the criterion measure. We randomly split this sample into a training set (n=978) for model development and selection, and a test set (n=419) for independent model evaluation.

Child Dataset

Data for the CHAP-CHILD model came from the Patterns of Habitual Activity across SEasons (PHASE) Study, which recruited students in Years 4 and 5 (ages 8 - 11 years) in the Melbourne (Australia) vicinity to evaluate seasonal variation in physical activity behavior across socio-economic and male/female categories. Each participant was asked to complete a physical activity assessment (simultaneous wear of the ActiGraph and activPAL) in the winter, spring, summer, and fall. Parental consent was obtained and the study protocol was approved by the participating institutional review boards. The sample used for this project comprised 718 participant-seasons from 278 participants (mean of 2.6 seasons per participant due to data loss and noncompliance). Participants were randomly divided into the following: a training set (n=194), used to train and select CHAP-CHILD model; and a testing set (n=84), used to evaluate the performance of the final CHAP-child model selected.


Training DatasetDescription
ACTACT is a cohort of community dwelling older adults age 65+. At time of accelerometer wear, the sample had a mean age of 76.7 years and was approximately 59% female and 90% non-Hispanic White.
AusDiabAusDiab is a population-based epidemiologic cohort of adults 35+ years in Australia. At time of accelerometer wear, the sample had a mean age of 58.3 years and was approximately 56% female.
PHASEPHASE is a cohort of children ages 8 - 11 years in Australia. At time of accelerometer wear, the sample had a mean age of 10.5 years and was approximately 51% female.

Data format

  • Accelerometer Data: We assume the input data is obtained from ActiGraph GT3X device and converted into single .csv files. The files should be named as **.csv** and files for all subjects should be put in the same directory. First few lines of a sample csv file are as follows:
      ------------ Data File Created By ActiGraph GT3X+ ActiLife v6.13.3 Firmware v3.2.1 date format M/d/yyyy at 30 Hz  Filter Normal -----------
      Serial Number: NEO1F18120387
      Start Time 00:00:00
      Start Date 5/7/2014
      Epoch Period (hh:mm:ss) 00:00:00
      Download Time 10:31:05
      Download Date 5/20/2014
      Current Memory Address: 0
      Current Battery Voltage: 4.07     Mode = 12
      Accelerometer X,Accelerometer Y,Accelerometer Z

    Note: The pre-processing function expects, by default, an Actigraph RAW file at 30hz with normal filter. If you have Actigraph RAW data at a different Hz level, please specify in the pre_process_data function using the gt3x-frequency parameter.

    Note: Our pre-trained models work and make predictions on time-series windows. If your data length is not exactly an integer multiply of the window size, the last a few minutes not enough to make up a whole window will be dropped. This is usually not an issue in most scenarios, but if it matters for your task, you can pad the file with 0s at the end to fill the last window. Note the predictions on this last window would be somewhat unreliable due to the missing data.

  • (Optional) Valid Days File: A .csv file indicating which dates are valid (concurrent wear days) for all subjects. Each row is subject id, date pair. The header should be of the from ID,Date.Valid.Day. Date values should be formatted in %m/%d/%Y format. A sample valid days file is shown below.

  • (Optional) Sleep Logs File: A .csv file indicating sleep records for all subjects. Each row is tuple of subject id, date went to bed, time went to bed, data came out of bed, and time went out of bed. The header should be of the form ID,Date.In.Bed,Time.In.Bed,Date.Out.Bed,Time.Out.Bed. Date values should be formatted in %m/%d/%Y format and time values should be formatted in %H:%M format. A sample sleep logs file is shown below.

  • (Optional) Non-wear Times File: A .csv file indicating non-wear bouts for all subjects. Non-wear bouts can be obtained using CHOI method or something else. Each row is a tuple of subject id, non-wear bout start date, start time, end date, and end time. The header should be of the form ID,Date.Nw.Start,Time.Nw.Start,Date.Nw.End,Time.Nw.End. Date values should be formatted in %m/%d/%Y format and time values should be formatted in %H:%M format. A sample sleep logs file is shown below.

  • (Optional) Events Data: Optionally, you can also provide ActivPal events data-especially if you wish to train your own models-for each subject as a single .csv file. These files should also be named in the **.csv** format and files for all subjects should be put in the same directory. The first few lines of a sample csv file are as follows:
      "Time","DataCount (samples)","Interval (s)","ActivityCode (0=sedentary, 1= standing, 2=stepping)","CumulativeStepCount","Activity Score (MET.h)","Abs(sumDiff)"

Pre-trained Models

We currently support several pre-trained models that can be used to generate predictions. The detailed description of these model architectures can be found in our paper. They have been trained on different training datasets, which have different demographics. The recommended and default model is the CHAP_ALL_ADULTS model. However, users can change the pre-trained model to better match their needs using the --model option. Below we provide a summary of the available pre-trained models and the characteristics of the datasets that they were trained on.

ModelTraining Dataset
CHAP_ALL_ADULTS (default and recommended)ACT + AUSDIAB

Additional Models

Below are some additional pre-trained models described in the previous paper.

ModelTraining Dataset
CHAP (ensemble of A, B, and C)ACT