A foundation model for electrodermal activity data

Leonardo Alchieri; Matteo Garzon; Lidia Alecci; Francesco Bombassei De Bona; Martin Gjoreski; Giovanni De Felice; Silvia Santini

cs.LGFeb 20, 2026

A foundation model for electrodermal activity data

Authors: Leonardo Alchieri, Matteo Garzon, Lidia Alecci, Francesco Bombassei De Bona, Martin Gjoreski, Giovanni De Felice, Silvia Santini

Organizations: Università della Svizzera Italiana (USI), Via la Santa 1, 6962 Lugano, Switzerland

Abstract

Foundation models have recently extended beyond natural language and vision to timeseries domains, including physiological signals. However, progress in electrodermal activity (EDA) modeling is hindered by the absence of large-scale, curated, and openly accessible datasets. EDA reflects sympathetic nervous system activity and is widely used to infer cognitive load, stress, and engagement. Yet very few wearable devices provide continuous, unobtrusive sensing, and the only large-scale archive to date is proprietary. To address this gap, we compile EDAMAME, a collection of EDA traces from 24 public datasets, comprising more than 25,000 hours from 634 users. Using this resource, we train UME, the first dedicated foundation model for EDA. In eight out of ten scenarios, UME outperforms baselines and matches generalist timeseries foundation models while using 20x fewer computational resources. Our findings, however, also highlight the intrinsic challenges of EDA modeling, motivating further research to unlock its full potential. All datasets, model weights, and code are released to support further research.

Figures & tables

Clinical data
Foundation model				train data
Ref.	Signal(s)	Params.	Avail.	Dataset(s)	Data type	Hours (k)	Users	Avail.
[ 72 ]	PPG	$\sim$ 30 - 150 M	Open source	VitalDB [ 57 ]	Clinical	$\sim$ 17	$\sim$ 6k	✓
				MIMIC-III [ 51 ]	Clinical	$\sim$ 20	$\sim$ 6k	✓
				MESA [ 94 ]	Clinical	$\sim$ 20	$\sim$ 2k	✓
[ 32 ]	PPG	$\sim$ 1-8 M	Code only	[ 32 ]	Clinical	$\sim$ 300	$\sim$ 29k	✓

Table 1: Overview of selected foundation models for physiological data and the datasets used to train them.

Dataset Name	Duration (h)	# Users	Scenario	Environment
APSync [ 39 ]	168	27	Real-world Stress	Wild
BIG IDEAS [ 11 ]	2607	16	Daily Living	Wild
BiHeartS [ 2 ]	2911	11	Sleep Monitoring	Wild
DREAMT [ 92 ]	882	100	Daily Living	Wild
Dynamics in the workplace [ 61 ]	1813	42	Workplace Analysis	Wild
EmpaticaE4Stress [ 20 ]	5	29	Stress/Emotion Induction	Lab

Table 2: Summary of the datasets included in the training split. The collection spans diverse physiological scenarios and environments.

Figure 1: Example of six EDA signals (randomly drawn). Each signal is from a different dataset, each for one of the six scenarios.

Figure 2: Ridgeline plot with distribution of EDA values from each dataset. The colors represent the six scenarios each dataset was collected in. The plot has been truncated, for visualization’s sake, in the range

0-2.5~\mu\mathrm{S}

. The original range was

0-40~\mu\mathrm{S}

(the E4’s max theoretical value is

100~\mu\mathrm{S}

Figure 3: Number of users and data per user across the EDAMAME collection of datasets. Marker size is proportional to the size of the dataset (in terms of hours of EDA data).

Figure 4: Overview of the train and downstream evaluation process we used for the UME foundation model. For reference, we report a sketch of the EfficientNet architecture, which we use as backbone for UME.

Dataset	Binary Tasks
APSync	Low/High Engagement
HeartS	Sleep/Wake
USILaughs	Cog. Load/Relaxation
WESAD	High/Low Arousal, High/Low Valence (self-report)
Nurses’ Stress	Low/High Stress
DREAMT	Sleep/Wake

Table 3: Summary of Datasets and Associated Binary Tasks

Figure 5: Pairplot with comparison between UME and the baseline features for the selected downstream tasks. Colors represent the two validation methods, TA and LOPO.

		Dummy	Handcrafted (HC)		Generalist Foundation			Ours
Dataset	Binary Task	Dummy	Gen.	EDA-spec.	Mantis	MOMENT	Chronos	UME
		Balanced accuracy ${}_{standard~error}$
Time-Aware cross-validation (TA)
APSYNC	Low/High engagement	$.45_{.20}$	$.80_{.10}$	$.80_{.10}$	$.86_{.07}$	$.47_{16}$	$.61_{.20}$	$.89_{.11}$
Dreamt	Deep Sleep/REM	$.48_{.02}$	$.55_{.05}$	$.59_{.04}$	$.64_{.03}$	$.65_{01}$	$.63_{02}$	$.63_{.05}$
Dreamt	Sleep/Wake	$.50_{.01}$	$.65_{.02}$	$.69_{.01}$	$.73_{.01}$	$.69_{01}$	$.73_{01}$	$.70_{.01}$

Table 4: Results of binary classification experiments across datasets and tasks, in terms of balanced accuracy

{}_{standard~error}

. Reported are results for both TA and LOPO cross-validation methods. Acronyms: Gen. HC stands for generic handcrafted features ; EDA HC stands for EDA-specific handcrafted features .

Panel of Figure 6 — Figure 6: Computational complexity analysis. (a) Average (across 20 samples) per-sample computation time for a 60-second window of EDA data for all three components (phasic, tonic, and original signal). Note: hc stands for handcrafted ; generic handcrafted features are not shown as they require negligible computation time. (b) GFLOPs for the different foundation models.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset Name	License
BiHeartS [ 2 ]	✓- data share agreement
BIG IDEAS [ 11 ]	✓- with original license (ODC-BY 1)
Dynamics in the workplace [ 61 ]	✓- data share agreement
EmpaticaE4Stress [ 20 ]	✓- (CCBY 4) with original license
EPM-E4 [ 37 ]	✓- (CCBY 4) with original license
LAUREATE [ 55 ]	✓- data share agreement

Appendix

Table A.1: Summary of datasets with their corresponding data-sharing agreements. This table includes all datasets, including citations and the "Reshareable" column, for inclusion in the appendix.

Handcrafted Methods
Model / Feature Set	Model Size	Feature Dim. ( $d$ )
Generic HC	N/A	12
EDA-specific HC	N/A	45
Generalist Foundation Models
Mantis [ 35 ]	$\sim$ 8 M	768
Chronos [ 9 ]	$\sim$ 200 M	1536

Appendix

Table A.2: Overview of the models and feature sets used in the experimental evaluation. We report the model size (number of trainable parameters) and the dimensionality of the feature space (

d

) used for linear probing.

Feature	Mathematical Notation or Formula	Feature Set(s)
Time-domain
Mean	$\frac{1}{N}\sum_{i=1}^{N}x_{i}$	Generic & EDA-specific
Minimum	$\min(x_{1},x_{2},\ldots,x_{N})$	Generic & EDA-specific
Maximum	$\max(x_{1},x_{2},\ldots,x_{N})$	Generic & EDA-specific
Standard Deviation	$\sqrt{\frac{1}{N-1}\sum_{i=1}^{N}(x_{i}-\bar{x})^{2}}$	Generic & EDA-specific
Dynamic Range	$\max(x_{1},x_{2},\ldots,x_{N})-\min(x_{1},x_{2},\ldots,x_{N})$	EDA-specific only

Appendix

Table A.3: Mathematical formulations for the 15 base hand-crafted features extracted from EDA time series signals (

x

) of length

N

. All features are computed independently for the tonic, phasic, and original EDA components. The "Feature Set(s)" column indicates whether the feature was used in the Generic baseline (12 total features) or the EDA-specific baseline (45 total features).

Hyperparameter	Values
Regularization strength ( $C$ )	0.01, 0.1, 1, 10
Solver	lbfgs, liblinear
Penalty	L2
Max iterations	10000
Class weight	balanced

Appendix

Table A.4: Hyperparameter Grid for Logistic Regression

Augmentation	Type	Description	Parameter Range
Frequency Domain & Component Isolation
Low-Pass Filter	EDA-Specific	Extracts tonic component; removes high-freq noise.	Cutoff $f\in[0.25,1.0]$ Hz
High-Pass Filter	EDA-Specific	Extracts phasic component; removes slow drifts.	Cutoff $f\in[0.05,0.25]$ Hz
Band-Pass Filter	EDA-Specific	Isolates information-rich EDA frequency bands.	Pass $f\in[0.05,0.25]$ Hz
Band-Stop Filter	EDA-Specific	Rejects specific frequency bands.	Reject $f\in[0.75,1.0]$ Hz
High Freq. Noise	EDA-Specific	Adds Gaussian noise only to frequencies $>1$ Hz.	Noise $\sigma\in[0,0.5]$

Appendix

Table A.5: Summary of Electrodermal Activity (EDA) Data Augmentations used to train the UME foundation model. The table details both EDA-specific transforms designed to isolate physiological components or simulate artifacts, and generic time series transforms. The set is the same proposed by Matton et al. [64] .

Figure B.1: Example of EDA signal reconstruction using MAE.

Figure B.2: Validation loss during training for different configurations of masking and masking weight.

Figure B.3:

Extractor	CPU Time (mean) [ms]	CPU Time (SE) [ms]
MOMENT (128M)	4823.53	161.11
Chronos (28M)	721.34	28.91
Mantis (8M)	350.23	39.06
UME (1M)	184.56	3.12
EDA-specific hc	35.39	1.40
generic hc	0.09	0.01

Appendix

Table C.1: CPU Performance Comparison of Extractors

Dataset	Binary Task	Model	Metrics
Dataset	Binary Task	Model	Balanced Acc.	MCC	F1
Dreamt	Deep Sleep/REM	Dummy	$0.73\pm 0.05$	$-0.01\pm 0.01$	$0.16\pm 0.06$
		Gen. HC	$0.64\pm 0.07$	$0.00\pm 0.03$	$0.10\pm 0.06$
		EDA-spec. HC	$0.54\pm 0.05$	$0.02\pm 0.03$	$0.19\pm 0.07$
		Mantis	$0.59\pm 0.05$	$0.02\pm 0.04$	$0.18\pm 0.07$
		MOMENT	$0.64\pm 0.03$	$0.06\pm 0.03$	$0.21\pm 0.07$

Appendix

Table C.2: Results of binary classification experiments, using the Leave One Participant Out (LOPO) cross-validation. Results shown as metric

\pm

standard error .

Dataset	Binary Task	Model	Metrics
Dataset	Binary Task	Model	Balanced Acc.	MCC	F1
Dreamt	Deep Sleep/REM	Dummy	$0.51\pm 0.01$	$0.01\pm 0.01$	$0.20\pm 0.06$
		Gen. HC	$0.55\pm 0.10$	$0.09\pm 0.20$	$0.28\pm 0.13$
		EDA-spec. HC	$0.59\pm 0.08$	$0.16\pm 0.13$	$0.37\pm 0.13$
		Mantis	$0.64\pm 0.07$	$0.22\pm 0.11$	$0.40\pm 0.12$
		MOMENT	$0.65\pm 0.04$	$0.24\pm 0.07$	$0.42\pm 0.10$

Appendix

Table C.3: Results of binary classification experiments for Time-Aware (TA) cross-validation. Results shown as metric

\pm

standard error . Experiments whose result is reported as OoM (out of memory) means that the required memory size for the embeddings computation exceeded the computational resources available.

Explore similar work

EMG-FM-Bench: A Comprehensive Benchmark for Foundation Model Transfer and Adaptation on Electromyography
Oct 5, 2026Tianhao Wu, Jiawei Yu, Xiangdong Hu +8Few-Shot Domain Adaptation Neural Decoding
University of Georgia · University of Oklahoma · University of Massachusetts Amherst
Artifact Annotations Partially Substitute for Per-User Calibration: SAFE-EDA and a Normalization-Controlled Evaluation of Wrist-EDA Affect Recognition
Oct 1, 2026Haochen Chai, Xinbi Luo, Zining Liu +1Self-Supervised Pre-Training Affective Computing
College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China
NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces
May 14, 2026Konstantinos Kontras, Trui Osselaer, Stylianos G. Mouslech +12Electroencephalography Brain-Computer Interfaces
KU Leuven · MIT

Oct 5, 2026J/K move · Enter open · S save

EMG-FM-Bench: A Comprehensive Benchmark for Foundation Model Transfer and Adaptation on Electromyography

Tianhao Wu, Jiawei Yu, Xiangdong Hu, Xu Wu+7

University of Georgia · University of Oklahoma · University of Massachusetts Amherst

Few-Shot Domain Adaptation Neural Decoding Transfer Learning Time-Series Foundation Models

Open paper AskPDF arXiv

Foundation models (FMs) are increasingly being developed for general time series and physiological signals, yet their transferability to downstream physiological tasks remains poorly understood. This question is particularly challenging for electromyography (EMG), where signal distributions vary substantially across users, sensing configurations, acquisition hardware, and downstream tasks. We introduce EMG-FM-Bench, a systematic benchmark for studying foundation-model transfer and adaptation on EMG. EMG-FM-Bench unifies 20 public datasets with over 1 million EMG segments and evaluates nine pretrained foundation models across four questions: how pretrained models perform when frozen or fully fine-tuned, how much pretraining helps compared with training the same model from scratch, how well models generalize to new users with limited labeled data, and how performance changes across different EMG tasks. Across the benchmark, linear probing provides useful information about pretrained representations, but full fine-tuning can substantially change downstream EMG performance. Comparing each pretrained model with the same model trained from scratch shows that the benefit of pretraining varies substantially across models and is not universal. Performance decreases when models are evaluated on new users, while five-shot adaptation improves macro-F1 in 70.2% of evaluated model-dataset combinations but recovers only part of the lost performance. Model performance is highly consistent between upper- and lower-limb classification and remains strongly correlated with continuous EMG-to-text decoding. Together, these results provide a systematic view of when pretrained time-series models transfer effectively to EMG and how their performance depends on fine-tuning, user variation, and downstream task.

Abstract

Explore similar work

EMG-FM-Bench: A Comprehensive Benchmark for Foundation Model Transfer and Adaptation on Electromyography

Artifact Annotations Partially Substitute for Per-User Calibration: SAFE-EDA and a Normalization-Controlled Evaluation of Wrist-EDA Affect Recognition

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces