EDAPT: Towards Calibration-Free BCIs with Continual Online Adaptation

This repository contains research code for the Journal of Neural Engineering paper: EDAPT: Towards Calibration-Free BCIs with Continual Online Adaptation by Haxel*, Kapoor*, Ziemann and Macke (2026) (* equal contribution; arXiv preprint).

EDAPT pretrains an EEG decoder on data from many users and then keeps adapting it to a new user during use, with supervised continual finetuning (CFT) after every trial and optional unsupervised domain adaptation (UDA). No calibration session is needed.

Installation

We recommend a conda environment. A GPU is recommended but not necessary.

git clone https://github.com/mackelab/EDAPT.git
cd EDAPT
conda create --name edapt python=3.11
conda activate edapt
pip install -r requirements.txt

Data

We use nine public EEG datasets across three paradigms: motor imagery (Yang2025, Lee2019_MI, BNCI2014_001), P300 (BI2015a, Huebner2017, Huebner2018) and SSVEP (Lee2019_SSVEP, MAMEM2, Kalunga2016). All but Yang2025 are downloaded automatically through MOABB. Download Yang2025 from its data paper and set YANG2025_DATA_DIR to its location.

Paths default to the repository and can be changed with environment variables (EDAPT_DATA_ROOT, YANG2025_DATA_DIR, EDAPT_RESULTS_DIR, EDAPT_SLURM_PARTITION, EDAPT_CONDA_ENV).

Running the experiments

The model code is in train_transfer.py (pretraining and online adaptation) and tta_wrapper.py (unsupervised adaptation). Experiments are launched on a SLURM cluster with submitit; add --dry-run to any launcher to print the jobs without submitting them.

Experiment	Command
main results: PRE / UDA / CFT ablation (Fig. 2, Table 3)	python submit_component_ablation.py --include-no-pretrain
scaling with number of pretraining subjects (Fig. 4)	python submit_pretrain_subs_ablation.py
scaling with trials per subject (Fig. 4)	python submit_trial_ablation.py
fixed data budget: subjects vs trials (Fig. 4)	python submit_isoscaling_analysis.py
hyperparameter sensitivity	python submit_hyperparam_sweep.py
UDA components, incl. the BN-1 baseline of Wimpff et al. (2024)	python submit_tta_comparison.py
latency	python analyze_latency.py

Figures and tables

All figures and tables of the paper can be regenerated from the aggregated results shipped in figures/figure_data and figures/prepared_data, without re-running any experiment (a few minutes on a CPU):

python figures/make_all.py --out figures/out

After re-running experiments, rebuild these files from your own results with the figures/generate_csv_*.py scripts.

Citation

@article{haxel2026edapt,
    title   = {EDAPT: towards calibration-free BCIs with continual online adaptation},
    author  = {Haxel, Lisa and Kapoor, Jaivardhan and Ziemann, Ulf and Macke, Jakob H.},
    journal = {Journal of Neural Engineering},
    year    = {2026},
    volume  = {23},
    number  = {2},
    pages   = {026025},
    doi     = {10.1088/1741-2552/ae5689}
}

Contact

Please open a GitHub issue for questions, or send an email to lisa.haxel@uni-tuebingen.de or jaivardhan.kapoor@uni-tuebingen.de.

License

MIT, see LICENSE.