CHAP2.0 Datasets and Models

This page summarizes the datasets and checkpoints used in the CHAP2.0 workflows. For preprocessing commands and input arguments, see CHAP2.0 Data Preprocessing.

Table of contents

  1. Datasets
    1. About
      1. SOL/PASOS
      2. iWatch
    2. Summary
  2. Data Format
    1. SOL/PASOS input files
    2. iWatch input files
  3. Recommended Checkpoints
  4. Submitted Checkpoints

Datasets

About

SOL/PASOS

SOL PASOS (Peripheral Artery Disease Study of SOL) is an ancillary study to the NHLBI-sponsored Hispanic Community Health Study/Study of Latinos (HCHS/SOL), a population-based prospective cohort study of 4,370 Hispanic/Latino adults and older adults aged 40 to 80 years in four regions of the U.S. SOL PASOS was conducted with HCHS/SOL Visit 3 in 2020-2024 and included up to 8 days of 24-hour measurement using wrist-worn ActiGraph monitors, with a subsample of 621 participants also wearing an activPAL. The activPAL measured sitting/lying and standing, and time in bed was excluded using participant bed logs and activPAL in-bed detection. The ActiGraph recorded raw triaxial acceleration in gravitational units, primarily at 80 Hz. The study was approved by the Institutional Review Board (IRB) at the University of North Carolina, and all participants provided informed consent.

iWatch

The iWatch study was an observational study of sitting and physical activity patterns in 148 adults aged 24 to 85 years, of whom 50% were female, 22% were non-White, and 20% were Hispanic. Daily activities were measured using a person-worn SenseCam (Vicon Revue), which captured first-person images approximately every 20 seconds while participants wore the camera on a lanyard for 7 days, excluding nighttime charging periods. SenseCam annotations were collapsed into two CHAP2.0 classes: Sitting, including Sedentary/Sitting and Vehicle, and Nonsitting, including Standing Still, Standing Moving, and Walking/Running. Nonwear and uncodable intervals were excluded. Data collection also included hip- and wrist-worn ActiGraph GT3X+ accelerometers, worn at the hip/waist or on the non-dominant wrist, with raw triaxial acceleration recorded in gravitational units at 30 Hz. The study was approved by the Institutional Review Board (IRB) at the University of California, San Diego, IRB #111160, date 07/12/2011, and all participants provided informed consent.

Summary

DatasetDevicesWear locationLabel sourceSampling frequency
SOL/PASOSActiGraph and activPALWrist ActiGraph; activPAL subsampleactivPAL sitting/lying and standing labelsActiGraph primarily 80 Hz
iWatchActiGraph GT3X+ and SenseCamHip and wrist ActiGraphSenseCam image annotations collapsed to sitting and nonsittingActiGraph 30 Hz

Data Format

CHAP2.0 preprocessing converts raw ActiGraph CSV or CSV.GZ files and dataset-specific annotation files into daily HDF5 files.

SOL/PASOS input files

For SOL/PASOS, preprocessing expects wrist ActiGraph raw files and can use supporting annotation files for valid wear days, sleep or in-bed time, non-wear intervals, and activPAL posture labels.

InputDescription
--gt3x-dirDirectory containing SOL/PASOS ActiGraph raw CSV files. Use the matching --gt3x-frequency for the input files.
--valid-days-fileOptional CSV identifying concurrent wear or valid analysis days.
--sleep-logs-fileOptional sleep or in-bed interval file. SOL/PASOS sleep logs use subject ID plus sleep start and sleep end timestamps.
--non-wear-times-fileOptional non-wear interval file. SOL/PASOS non-wear files use subject ID plus non-wear start and end timestamps.
--activpal-dirDirectory containing activPAL label files. Leave --event-file unset for 1-second epoch files with TS_LOCAL_COR and PL_ACTIVITY_NEW columns.

The ActiGraph raw files are gzipped CSV files. The first few lines of a sample file are as follows:

------------ Data File Created By ActiGraph GT3X+ ActiLife v6.13.5 Firmware v1.7.2 date format M/d/yyyy at 80 Hz  Filter Normal -----------
Serial Number: TAS1H45190074
Start Time 09:06:00
Start Date 8/18/2023
Epoch Period (hh:mm:ss) 00:00:00
Download Time 15:16:26
Download Date 8/30/2023
Current Memory Address: 0
Current Battery Voltage: 3.75     Mode = 12
--------------------------------------------------
Accelerometer X,Accelerometer Y,Accelerometer Z
-0.227,0.977,-0.059
-0.207,0.961,-0.051

The activPAL files provide 1-second epoch posture labels:

HCHSID,TS_LOCAL_COR,TS_LOCAL,TS_UTC,PL_ACTIVITY,PL_ACTIVITY_NEW,PL_DUR,PL_WAKE_PRD,PL_STEPS,PL_MET_H,PL_WAKESLEEPDAY_ID
M7148383,2022-04-19T11:20:49Z,2022-04-19T11:20:49Z,2022-04-19T15:20:49Z,0,0,10546.3,1,0,0.000347222222222222,1
M7148383,2022-04-19T11:20:50Z,2022-04-19T11:20:50Z,2022-04-19T15:20:50Z,0,0,10546.3,1,0,0.000347222222222222,1

The valid-days file identifies subject days with concurrent wear:

"ID","valid_days"
"C6002331",2023-11-25
"C6002331",2023-11-26
"C6002331",2023-11-27

The sleep-log file contains subject-level sleep or in-bed intervals:

ID,startsleep,endsleep
B5025951,3/28/23 23:00,3/29/23 9:09
B5025951,3/30/23 20:29,3/31/23 9:14

The non-wear file contains subject-level non-wear intervals:

ID,startNW,endNW
C6003361,6/20/22 5:39,6/20/22 7:16
C6003361,6/20/22 23:58,6/21/22 10:12

For datasets with mixed raw sampling frequencies, split raw files by frequency and run preprocessing separately with the matching --gt3x-frequency. For example, if a SOL/PASOS input directory contains both 60 Hz and 80 Hz files, process those subject groups separately rather than relying on one mixed run.

iWatch input files

For iWatch, preprocessing expects hip or wrist ActiGraph raw files and can use SenseCam-derived event labels and location-specific non-wear intervals.

InputDescription
--gt3x-dirDirectory containing iWatch ActiGraph raw CSV or CSV.GZ files. Use --gzipped for .csv.gz files.
--activpal-dirDirectory containing event label CSV files matched to the raw ActiGraph file names. Use --event-file for this format.
--non-wear-times-fileOptional iWatch non-wear interval file. If the file includes both hip and wrist records, pass --loc hip or --loc wrist.
--valid-days-fileOptional valid-day CSV, if available.
--sleep-logs-file / --wear-logs-fileOptional sleep or wear interval files, if available for the dataset version being processed.

The ActiGraph raw files are gzipped CSV files. The first few lines of a sample file are as follows:

------------ Data File Created By ActiGraph GT3X+ ActiLife v6.13.4 Firmware v3.2.1 date format M/d/yyyy at 30 Hz  Filter Normal -----------
Serial Number: MRA1F08120057
Start Time 00:00:00
Start Date 4/16/2014
Epoch Period (hh:mm:ss) 00:00:00
Download Time 13:16:21
Download Date 4/29/2014
Current Memory Address: 0
Current Battery Voltage: 3.85     Mode = 12
--------------------------------------------------
Accelerometer X,Accelerometer Y,Accelerometer Z
-0.393,0.094,0.962
-0.387,0.094,0.956

The event label files are CSV files with start time, duration, and activity code columns:

"Time","DataCount (samples)","Interval (s)","ActivityCode (0=sedentary, 1= standing, 2=stepping)","CumulativeStepCount","Activity Score (MET.h)","Abs(sumDiff)"
"41820.3555555556","",120,1,"","",""
"41820.3569444444","",780,0,"","",""
"41820.3833333333","",360,-1,"","",""
"41820.3881944444","",120,-2,"","",""

The iWatch non-wear file can include both hip and wrist records. Use --loc to select the wear location:

"ID","wearLoc","NW_DT","int.min","wearDate","timeNum","filename","loc"
"i0234A","N",2014-04-16,1440,2014-04-16,0,"i0234A_N2005760sec.agd","Wrist"
"i0234A","N",2014-04-17,874,2014-04-17,0,"i0234A_N2005760sec.agd","Wrist"
"i0234A","N",2014-04-17 15:40:00,500,2014-04-17,940,"i0234A_N2005760sec.agd","Wrist"

The preprocessing script converts these inputs into daily HDF5 files used by CHAP2.0 prediction and fine-tuning. See Data Preprocessing for the output HDF5 structure.

Use these checkpoints as the default recommendations for CHAP2.0 prediction:

CheckpointRecommended for
SUBMIT_RESULT/SOL_W/CHAP_FT/checkpoint-submit.pthActiGraph wrist data.
SUBMIT_RESULT/iWatch_H/CHAP-ZS/checkpoint/checkpoint-submit.pthActiGraph hip data (ACT and AusDiab trained).

The SOL/PASOS fine-tuned wrist checkpoint is the recommended wrist model. The iWatch wrist checkpoint is retained for reproducibility, but is not the default recommendation.

Submitted Checkpoints

CHAP2.0 submitted checkpoints are stored under CHAP2/SUBMIT_RESULT/ and can be used with the CHAP2.0 prediction workflow.

CHAP-ZS means zero-shot prediction: the CHAP/MSSE-2021 pre-trained checkpoint is applied directly to the target dataset without fine-tuning. CHAP-FT checkpoints are initialized from CHAP/MSSE-2021 weights and then fine-tuned on the target dataset.

The following checkpoints are included for reproducibility and comparison; not all are recommended as default models.

CheckpointTypeTarget dataNotes
SOL_W/CHAP_FT/checkpoint-submit.pthCHAP-FTSOL/PASOS wristFine-tuned on SOL/PASOS wrist data; recommended for ActiGraph wrist data.
iWatch_W/CHAP-FT/checkpoint/checkpoint-submit.pthCHAP-FTiWatch wristFine-tuned on iWatch wrist data; retained for reproducibility.
iWatch_W/CHAP-ZS/checkpoint/checkpoint-submit.pthCHAP-ZSiWatch wristZero-shot baseline for iWatch wrist data.
iWatch_H/CHAP-FT/checkpoint/checkpoint-submit.pthCHAP-FTiWatch hipFine-tuned on iWatch hip data; retained for reproducibility.
iWatch_H/CHAP-ZS/checkpoint/checkpoint-submit.pthCHAP-ZSiWatch hipRecommended checkpoint for ActiGraph hip data trained on ACT and AusDiab.