ADABase: Cognitive Driving Public Set

Multimodal physiological and behavioral recordings for cognitive load estimation during an n-back test and a driving-simulator study. This record holds the Public Set: 30 of the 51 ADABase recordings.

How to cite

Please cite this dataset record and both papers: ADABase introduced the acquisition setup, and REVELIO extends it with the gaming cohort.

@article{oppelt2023adabase,
  author  = {Oppelt, Maximilian P. and Foltyn, Andreas and Deuschel, Jessica and Lang, Nadine R.
             and Holzer, Nina and Eskofier, Bjoern M. and Yang, Seung Hee},
  title   = {{ADABase}: A Multimodal Dataset for Cognitive Load Estimation},
  journal = {Sensors},
  year    = {2023},
  volume  = {23},
  number  = {1},
  pages   = {340},
  doi     = {10.3390/s23010340}
}

@article{oppelt2026revelio,
  author  = {Oppelt, Maximilian P. and Foltyn, Andreas and Lang-Richter, Nadine R. and Eskofier, Bjoern M.},
  title   = {{REVELIO}: Multimodal Task Load Estimation for Cross-Domain Generalization},
  journal = {International Journal of Human--Computer Interaction},
  publisher = {Taylor \& Francis},
  year    = {2026},
  doi     = {10.48550/arXiv.2509.01642},
  note    = {In press. DOI refers to the arXiv preprint arXiv:2509.01642}
}

Dataset overview

Participants. ADABase and REVELIO together comprise 96 sessions from 88 subjects (51 in the driving study, 37 in the gaming study). Five people took part in both studies, so the subjects are 83 different individuals. The sessions are split into four Zenodo records:

The full ADABase cohort comprises 51 participants.

Tasks. Each recording contains an n-back test and a driving-simulator study with single- and dual-task workloads of increasing intensity. Each recording lasts more than two hours. See the ADABase paper for the protocol.

Modalities.

• Physiological signals (BIOPAC MP160, recorded at 2000 Hz, stored at 1 kHz): ECG, PPG, EDA, EMG (trapezius), respiration, skin temperature
• Eye tracking (Tobii Pro Fusion, 250 Hz): gaze and pupil diameter
• Behavioral features extracted from video (camera at 25 Hz): facial action units and body landmarks. No raw video or images are released.
• Performance measures (e.g. reaction times) and subjective ratings (NASA-TLX questionnaires)

Public/private split. The 51 recordings are split into two Zenodo records:

• Public Set: 30 recordings (IDs 0000–0029), access on request.
• Private Set: 21 recordings (IDs 0030–0050), retained as a hidden test set.

Data collection was approved by the ethics committee of Friedrich-Alexander-Universität Erlangen-Nürnberg (no. 129_21 B, 21 April 2021). All participants gave written informed consent, and only participants who consented to data sharing are included.

Files

This record contains 30 HDF5 files (27.32 GB in total, IDs 0000–0029), plus checksums.md5.

File naming scheme: <NNNN>-<CODE>.h5

• NNNN: four-digit participant ID. IDs are unique across ADABase and REVELIO, so the same NNNN always denotes the same person. ADABase uses IDs 0000–0050.
• CODE: pseudonym of the session.

Example: 0009-ILAB.h5 is the driving recording of participant 0009. The same participant also appears in REVELIO (0009-WMS15.h5, 0009-VRE30.h5).

0000-ACAA.h5, 0001-AGDJ.h5, 0002-AHEG.h5, 0003-AZFR.h5, 0004-BDCD.h5, 0005-BQAB.h5, 0006-BRJG.h5, 0007-FKDN.h5, 0008-GCAF.h5, 0009-ILAB.h5, 0010-ISJP.h5, 0011-KISN.h5, 0012-KSLI.h5, 0013-KYXT.h5, 0014-LEDG.h5, 0015-LINI.h5, 0016-LTRQ.h5, 0017-MDMS.h5, 0018-NMIL.h5, 0019-PFAC.h5, 0020-RLCE.h5, 0021-RNDM.h5, 0022-RSBC.h5, 0023-SFKF.h5, 0024-SZLE.h5, 0025-TBMO.h5, 0026-XAPL.h5, 0027-YMAD.h5, 0028-ZFML.h5, 0029-ZZOV.h5

HDF5 structure

Each file is written with pandas (PyTables) and holds four tables, which pandas.read_hdf reads directly. Driving (ADABase) and gaming (REVELIO) files share this layout.

Channels in SIGNALS.

The full channel list is printed by the example in "Loading the data in Python".

Sampling rates. The biosignals are stored at 1 kHz. Streams with a lower native rate (eye tracker 250 Hz, video features 25 Hz) are NaN between their samples. Calling .dropna() on a column recovers the native samples. The row number equals the time in ms, so the start and stop arguments of pandas.read_hdf select a time range without reading the whole file. A file has about 9 million rows, so read only the columns and time range you need.

Annotation. Every sample in SIGNALS carries three labels:

• STUDY: the task. Time outside all tasks is n/a.
• PHASE: baseline, train (practice), test, or n/a between phases.
• LEVEL: numbers the conditions within a study and phase.

Loading the data in Python

import pandas as pd

path = "0000-ACAA.h5"

# Tables and channels
with pd.HDFStore(path, "r") as store:
    print(store.keys())  # SIGNALS, SUBJECTIVE, PERFORMANCE, META
    channels = list(store.get_storer("SIGNALS").non_index_axes[0][1])
print(len(channels), channels[:10])

# Small tables
subjective = pd.read_hdf(path, "SUBJECTIVE")
performance = pd.read_hdf(path, "PERFORMANCE").dropna(axis=1, how="all")
meta = pd.read_hdf(path, "META")

# NASA-TLX: unweighted mean (RTLX) and weighted score (TLX)
dims = ["MENTAL", "PHYSICAL", "TEMPORAL", "PERFORMANCE", "EFFORT", "FRUSTRATION"]
weights = subjective[[f"WEIGHT {d}" for d in dims]].to_numpy()
subjective["RTLX"] = subjective[dims].mean(axis=1)
subjective["TLX"] = (subjective[dims].to_numpy() * weights).sum(axis=1) / weights.sum(axis=1)

# Annotation of the whole recording (reads all rows, about 30 s)
annotation = pd.read_hdf(path, "SIGNALS", columns=["STUDY", "PHASE", "LEVEL"])
test_3 = annotation.index[
    (annotation["STUDY"] == "n-back") & (annotation["PHASE"] == "test") & (annotation["LEVEL"] == 3)
]

# Selected channels during n-back test 3 (index = time in ms = row number)
signals = pd.read_hdf(
    path,
    "SIGNALS",
    columns=["RAW_ECG_I", "ECG_I_PEAKS", "EYE_LEFT_PUPIL_DIAMETER"],
    start=test_3[0],
    stop=test_3[-1] + 1,
)
pupil = signals["EYE_LEFT_PUPIL_DIAMETER"].dropna()  # native 250 Hz samples

Checksums

Each file's MD5 checksum is shown on the Zenodo record page. The file checksums.md5 in this record lists the same checksums. After downloading, verify all files with:

md5sum -c checksums.md5          # Linux
# macOS: sudo port install coreutils, then gmd5sum -c checksums.md5

Related records and publications

Supplement to the ADABase paper and the REVELIO paper. Counterpart of the ADABase Private Set. Continued by the REVELIO Public and Private Sets.

Publications:

• Oppelt, M. P., Foltyn, A., Deuschel, J., Lang, N. R., Holzer, N., Eskofier, B. M., & Yang, S. H. (2023). ADABase: A Multimodal Dataset for Cognitive Load Estimation. Sensors, 23(1), 340. https://doi.org/10.3390/s23010340
• Oppelt, M. P., Foltyn, A., Lang-Richter, N. R., & Eskofier, B. M. (2026). REVELIO: Multimodal Task Load Estimation for Cross-Domain Generalization. International Journal of Human–Computer Interaction, in press. Preprint: https://doi.org/10.48550/arXiv.2509.01642

License, access and contact

The files of this record are restricted. Researchers with a permanent position at an academic institution can request access with the Request access button on this Zenodo record. Access is free of charge and limited to non-commercial academic research. In the request, state your name, position, affiliation, postal address, email, up to five additional researchers (with affiliation and email), and a short description of the intended research use.

Use of the data is governed by the End User License Agreement (EULA): https://raw.githubusercontent.com/adabase-dataset/adabase-dataset.github.io/main/LICENSE.md. By requesting access, you accept it.

Copyright (C) 2026 The Authors.

Contact: adabase@iis.fraunhofer.de