A foundation model for electrodermal activity data
The model, trained on 25,000+ hours of skin conductance data, matches larger models with 20x less compute.
A team of researchers has published a paper introducing UME, the first foundation model specifically designed for electrodermal activity (EDA) data. EDA measures skin conductance, a key indicator of sympathetic nervous system activity used to infer cognitive load, stress, and emotional engagement. The project addresses a critical gap: the lack of large-scale, open datasets for this physiological signal, which has hindered AI progress in the domain. The team compiled EDAMAME, a massive new resource aggregating over 25,000 hours of EDA traces from 24 existing public datasets, covering 634 unique users.
Using the EDAMAME dataset, the researchers trained the UME foundation model. In benchmark tests across ten different scenarios, UME outperformed existing baseline methods in eight of them. Notably, it achieved performance comparable to larger, generalist timeseries foundation models while requiring 20 times fewer computational resources for training—a significant efficiency gain. The paper also candidly discusses the intrinsic challenges of modeling the noisy, complex nature of EDA signals, highlighting areas for future research to fully unlock its potential for health and wellness applications. In a major boost for open science, the team has released the complete EDAMAME dataset, the UME model weights, and all associated code publicly.
- Built the EDAMAME dataset: 25,000+ hours of EDA data from 634 users across 24 public sources.
- Trained UME foundation model: Matches larger generalist models' performance using 20x less computational power.
- Full open-source release: All datasets, model weights, and code are publicly available for research.
Why It Matters
Enables scalable, efficient AI for stress and cognitive load monitoring in wearables, moving beyond proprietary data silos.