[Paper Review] A foundation model for electrodermal activity data

Leonardo Alchieri, Matteo Garzon+5 more|arXiv (Cornell University)|Feb 20, 2026

Emotion and Mood Recognition0 citations

TL;DR

This paper introduces UME, an open-source foundation model for electrodermal activity (EDA) trained on EDAMAME, a large, open multi-dataset collection of wearable EDA data, and shows it outperforms baselines on downstream tasks while using far fewer resources.

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.

Motivation & Objective

• Motivate the need for open large-scale EDA data to enable foundation models for wearable signals.
• Create EDAMAME, a diverse, open collection of EDA datasets to train and evaluate models.
• Develop UME, an open foundation model for EDA data trained with self-supervised contrastive learning.
• Evaluate UME on diverse downstream tasks and compare to handcrafted features and other time-series models.
• Provide open access to datasets, model weights, and code to foster reproducibility and further research.

Proposed method

• Assemble EDAMAME from 24 public EDA datasets, totaling over 25,000 hours and 634 users, all collected with Empatica E4 devices and 4 Hz sampling.
• Preprocess data with Butterworth low-pass filtering (0.4 Hz) and cvxEDA decomposition into phasic and tonic components, using phasic, tonic, and original EDA signals as inputs.
• Segment data into 60-second windows with 0.25 s overlap for training, yielding ~275 million windows for UME training.
• Train UME, an open foundation model with an EfficientNet backbone adapted for 1D time series (240-length input, 3 channels), approximately 1M parameters and latent size d=64, using contrastive learning with InfoNCE loss.
• Use two augmentations per EDA segment to create positive pairs; negatives come from other segments in the batch; perform ablations including a masked reconstruction attempt which was not successful.
• Evaluate representations via linear probing on a downstream EDAMAME-test suite, comparing against generic handcrafted features, EDA-specific handcrafted features, and generalist time-series foundation models (Chronos, MOMENT, Mantis).

Experimental results

Research questions

• RQ1 Can a foundation model trained on a large, open EDA multi-dataset corpus learn transferable representations for downstream EDA tasks?
• RQ2 How does UME perform compared with handcrafted feature baselines and existing generalist time-series foundation models on EDA classification tasks?
• RQ3 What are the data characteristics and challenges of real-world EDA that affect model learning and generalization?

Key findings

• UME outperforms baselines trained on generic handcrafted features in 8 of 10 downstream tasks.
• UME matches the performance of generalist time-series foundation models while using at least 20× less computational resources.
• EDAMAME contains approximately 25,000 hours of EDA data from 634 users, with EDA sampled at 4 Hz across all datasets.
• The UME model uses an EfficientNet backbone with ~1M parameters and a latent size of 64, operating on 60-second windows and three input components (phasic, tonic, original).
• Balanced accuracy across tasks is generally below 0.7 and exhibits substantial variability, highlighting intrinsic challenges in EDA modeling.
• All datasets, model weights, and code are released to support further research.