josefheidler/har_ee_adults_2024-lendt

Lendt et al. 2024 — thigh-worn accelerometry + energy expenditure, healthy adults

Thigh-worn accelerometry (Axivity AX6, 100 Hz, ±8 g) paired with indirect-calorimetry energy expenditure in 69 healthy adults across a standardized laboratory protocol of sitting, standing, walking (4–5 speeds), running (4–5 speeds), and cycling (3–5 power levels). Harmonized from the original Zenodo release into per-subject parquet — one row per raw accelerometer sample, with 1 Hz activity labels and steady-state energy expenditure asof-joined on. 106 M samples total, 33 % labelled (97.9 h labeled out of 294.9 h total).

Sibling cohort to lendt_adults_2024 (different device, different subjects, no free-living condition, adds EE signal).

Source

• Paper: Lendt, C., Hansen, N., Froböse, I., & Stewart, T. (2024). Composite activity type and stride-specific energy expenditure estimation model for thigh-worn accelerometry. International Journal of Behavioral Nutrition and Physical Activity, 21(1), 99. https://doi.org/10.1186/s12966-024-01646-y
• Raw data: Zenodo record 13477128
• License: CC-BY-4.0

Protocol

• Participants: 69 healthy adults (25.2 ± 5.8 y, 49 % female). Divided by the authors into a training subset (n = 49) and a validation subset (n = 20). The dataset includes the full cohort; the authors' split is preserved in the split column.
• Sensor: Axivity AX6 (Axivity Ltd, Newcastle upon Tyne, UK). Triaxial accelerometer, ±8 g, 100 Hz. Pre-session static calibration via rectangular plastic cube (6 axes × 10 s); this window is labeled calibration in the harmonized data. Worn on the lateral side of the dominant thigh, midway between hip and knee, taped and secured with an elastic bandage (harmonized data transformed to front-thigh frame — see Axis orientation).
• Protocol: standardized laboratory sequence — sitting → standing → walking (4–5 speeds incl. 6 % incline) → running (4–5 speeds incl. 6 % incline) → cycling (3–5 power levels at 60–80 rpm self-selected). Each stage lasts 6 min with 5 min rest between conditions. The training and validation groups use slightly different speed/power lists — see speed_kph and power_w column values.
• Reference EE: Cortex Metalyzer 3B (Cortex Biophysik, Leipzig). Breath-by-breath VO2/VCO2 converted to kcal/min via Weir's equation. Steady-state EE averaged over the last 3 min of each 6-min stage.

Ground-truth provenance

• Stage start times were recorded by researchers during data collection.
• Each stage lasted 6 min; the 5-min inter-stage rest is unlabeled.
• Reference resolution: 1 s.

Schema

All timestamps are tz-aware Europe/Berlin (recording-site local time).

column	dtype	notes
subject	string	pseudonymized alias, e.g. happy-alpaca
split	string	train (n=49) or test (n=20)
timestamp	timestamp[ns, Europe/Berlin]	~10 ms cadence (100 Hz)
acc_x	float64	hub standard frame (x=up), units = g
acc_y	float64	hub standard frame (y=right), units = g
acc_z	float64	hub standard frame (z=forward), units = g
label	string (nullable)	activity label; null outside any labeled window
variant	string (nullable)	intensity tier (slow · moderate · fast); null for sit/stand
speed_kph	float64 (nullable)	treadmill speed in km/h (walk/run only)
power_w	float64 (nullable)	cycling power in W (bicycle only)
incline_pct	float64 (nullable)	treadmill incline in % (6.0 for incline stages, null otherwise)
vo2	float64 (nullable)	L/min, mean over last 3 min of stage
vco2	float64 (nullable)	L/min, mean over last 3 min of stage
ee_kcal_min	float64 (nullable)	Weir (3.9·VO2 + 1.1·VCO2), kcal/min
ee_kcal_min_kg	float64 (nullable)	kcal/kg/min — paper's primary EE unit
met	float64 (nullable)	metabolic equivalents (≈ kcal/kg/hour)

Example rows:

subject       split  timestamp                   label        variant   speed_kph  power_w  incline_pct  ee_kcal_min
happy-alpaca  train  2023-09-07 08:45:00+02:00   calibration  null      null       null     null         null
happy-alpaca  train  2023-09-07 09:14:32+02:00   stand        null      null       null     null         1.71
happy-alpaca  train  2023-09-07 09:22:10+02:00   walk         slow      2.1        null     null         3.37
happy-alpaca  train  2023-09-07 09:55:00+02:00   walk         moderate  3.7        null     6.0          5.84
happy-alpaca  train  2023-09-07 10:15:00+02:00   bicycle      fast      null       150.0    null         10.32
happy-alpaca  train  2023-09-07 08:50:00+02:00   null         null      null       null     null         null

Label vocabulary

label (6 values): calibration, sit, stand, walk, run, bicycle.

variant — intensity tier for walk/run/bicycle; null for sit/stand:

variant	walk (km/h)	run (km/h)	bicycle (W)
slow	2.1 (train) · 2.5 (val)	7.5 (train)	30
moderate	2.9–3.7 (train) · 3.3 (val)	8.5–9.0	60
fast	4.5 (train) · 4.1 (val)	9.5–10.5 (train) · 10.0 (val)	90–150

Multiple protocol stages can share the same variant — the exact value is always in speed_kph or power_w. Incline stages (6 % treadmill gradient) share the base variant of their speed but have incline_pct = 6.0.

All unique (label, variant, speed_kph, power_w, incline_pct) combinations present in the data:

label	variant	speed_kph	power_w	incline_pct	group
sit	—	—	—	—	test
stand	—	—	—	—	both
walk	slow	2.1	—	—	train
walk	slow	2.5	—	—	test
walk	moderate	2.9	—	—	train
walk	moderate	3.3	—	—	test
walk	moderate	3.3	—	6.0	test
walk	moderate	3.7	—	—	train
walk	moderate	3.7	—	6.0	train
walk	fast	4.1	—	—	test
walk	fast	4.5	—	—	train
run	slow	7.5	—	—	train
run	moderate	8.0	—	—	test
run	moderate	8.0	—	6.0	test
run	moderate	8.5	—	—	train
run	moderate	9.0	—	—	test
run	fast	9.5	—	—	train
run	fast	9.5	—	6.0	train
run	fast	10.0	—	—	test
run	fast	10.5	—	—	train
bicycle	slow	—	30	—	both
bicycle	moderate	—	60	—	both
bicycle	fast	—	90	—	both
bicycle	fast	—	120	—	train
bicycle	fast	—	150	—	train

Label glossary

Full meaning of every (label, variant) combination in the dataset. The exact numeric intensity is always in speed_kph, power_w, or incline_pct; the variant is the intensity tier.

label	variant	meaning	labeled min
bicycle	slow	Cycle ergometer 30 W	384.1
bicycle	moderate	Cycle ergometer 60 W	384.1
bicycle	fast	Cycle ergometer 90–150 W (training: 90/120/150; validation: 90)	904.2
calibration	—	Pre-session static calibration: 6 sensor orientations × 10 s on a rectangular plastic cube, immediately before the activity protocol	97.1
run	slow	Treadmill 7.5 km/h (training protocol only)	282.0
run	moderate	Treadmill 8.5 km/h (train) · 8.0–9.0 km/h (val); includes an 8.0 km/h + 6 % incline stage	620.5
run	fast	Treadmill 9.5–10.5 km/h (train) · 10.0 km/h (val); includes a 9.5 km/h + 6 % incline stage	862.1
sit	—	Seated at rest (validation protocol only; training protocol starts from standing)	118.2
stand	—	Upright stationary standing (both protocols)	404.0
walk	slow	Treadmill 2.1 km/h (train) · 2.5 km/h (val)	405.9
walk	moderate	Treadmill 2.9–3.7 km/h (train) · 3.3 km/h (val); includes a 6 % incline stage	1103.5
walk	fast	Treadmill 4.5 km/h (train) · 4.1 km/h (val)	407.8

Total activity labeled: 5876.4 min / 97.9 h across 69 participants (calibration excluded).

Coverage

Every raw accelerometer sample is kept, even outside labeled windows. Rows outside any labeled stage have label = null and all derived columns null — typically the 5-min inter-stage rest periods and the pre/post-protocol margins.

Each sample carries the most recent stage label whose 1 Hz timestamp falls at or before the sample's timestamp, within 1 s. Samples outside any labeled window get label = null.

The calibration label marks the pre-session static calibration block (≈90 s per subject, immediately before the activity protocol). Exclude them from HAR or EE analysis with df[~df['label'].isin(['calibration', None])] or keep them if your pipeline needs a static reference period.

The EE columns (vo2, vco2, ee_kcal_min, ee_kcal_min_kg, met) are stage-level steady-state summaries — constant across all samples within a stage, not breath-by-breath.

Axis orientation

Accelerometer values use the hub standard axis convention, shared across all datasets on this hub:

• acc_x runs along the thigh toward the head — reads +1 g when the person stands upright.
• acc_y points to the person's right — positive when tilting right, negative when tilting left.
• acc_z points forward — positive when leaning or stepping forward, negative when leaning backward.
• At rest standing upright: acc_x ≈ +1 g, acc_y ≈ 0, acc_z ≈ 0.

Harmonized accelerometer data is presented in a front-thigh frame — as if the sensor were mounted on the anterior surface of the dominant thigh. The physical sensor (Axivity AX6) was originally worn on the lateral side of the dominant thigh (midway between hip and knee); corrections transform the native lateral-thigh axes to the front-thigh convention.

Native lateral-thigh axes, verified from the raw sit/cycling ground truth (median gravity per posture), standing upright:

axis	meaning	right thigh	left thigh
x	long axis	down	down
y	anterior–posterior	+y backward	+y forward
z	medio-lateral	lateral	lateral

The anterior–posterior axis is y (not z), and the medio-lateral axis is z — the reverse of an earlier assumption that rotated the harmonized frame 90° (anterior gravity landed on −y). The y-sign is mirrored between sides because the mirrored placement rotates the device ≈180° about the long axis.

Correction applied to reach front-thigh (hub) frame — x=up, y=right, z=forward — a proper rotation:

• Both: acc_x ← -acc_x (long axis up).
• Right-dominant: acc_y ← -acc_z, acc_z ← -acc_y.
• Left-dominant: acc_y ← acc_z, acc_z ← acc_y.

A per-subject safety normalization then enforces sitting acc_z positive (falling back to cycling), guarding against any mislabeled dominant leg. Verified: every subject reads sitting acc_z positive (and cycling acc_z positive), matching the other thigh datasets on this hub.

plucky-python and royal-robin have dominant_leg overridden to right per the upstream code/main.ipynb correction, so they receive the right-dominant correction path.

For raw Axivity-native (lateral-thigh) values, ingest the upstream Zenodo record directly (record 13477128).

Gravitational calibration not applied. No autocalibration (e.g. sphere-fitting) has been performed on the harmonized data. The calibration label marks the 6-position static calibration block recorded before each session — these rows can be used to derive and apply a gravitational calibration if your downstream analysis requires it.

Harmonization notes

All 69 participants have both raw accelerometry and CPET data — no subjects excluded.

Participants plucky-python and royal-robin have dominant_leg overridden to right in the participant manifest (config.yaml), matching the upstream code/main.ipynb correction — they follow the right-dominant (no y-negation) path.

Per-participant demographics (split, gender, age, dominant_leg, height, weight, bmi, leg_length) are available in participants.csv at the repository root, also loadable as the participants dataset config — see Loading.

Label format uses bare base labels only (stand, sit, walk, run, bicycle). Stage intensity is preserved in the variant, speed_kph, power_w, and incline_pct columns rather than base[modifier] encoding — see Label vocabulary.

Use

Intended for energy expenditure estimation and human activity recognition from thigh-worn accelerometry. Predict ee_kcal_min or met from the acc_x/y/z signal, or label for HAR.

With 69 subjects split into a predefined train/validation partition (the split column contains train or test), the standard evaluation protocol respects this split: train on the train split, evaluate on the test split.

Loading

Each subject is stored as a separate parquet file under harmonized/. Filenames are pseudonymized aliases (happy-alpaca.parquet, brave-bison.parquet, …). Every row carries a subject column with the same alias, so identity is preserved when loading all files together.

Load train and test splits:

from datasets import load_dataset

ds = load_dataset("josefheidler/har_ee_adults_2024-lendt")
train_df = ds["train"].to_pandas()
test_df = ds["test"].to_pandas()

Load all subjects into one table (both splits combined):

import pandas as pd
import datasets

ds = load_dataset("josefheidler/har_ee_adults_2024-lendt")
df = pd.concat([ds["train"].to_pandas(), ds["test"].to_pandas()], ignore_index=True)

Load participant demographics:

from datasets import load_dataset

participants = load_dataset(
    "josefheidler/har_ee_adults_2024-lendt",
    name="participants",
)["participants"].to_pandas()

Load a single subject with pandas:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/josefheidler/har_ee_adults_2024-lendt/harmonized/train/happy-alpaca.parquet"
)

Citation

Lendt, C., Hansen, N., Froböse, I., & Stewart, T. (2024). Composite activity type and stride-specific energy expenditure estimation model for thigh-worn accelerometry. International Journal of Behavioral Nutrition and Physical Activity, 21(1), 99. https://doi.org/10.1186/s12966-024-01646-y