A foundation model for electrodermal activity data

2026-03-19 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Internet of Things (IoT) & Connected Devices · Depth: Advanced, extended

Summary

Researchers have developed UME, the first dedicated open-source foundation model for electrodermal activity (EDA) data, addressing a critical gap in large-scale, curated, and openly accessible physiological datasets. To train UME, they compiled EDAMAME, a collection of EDA traces from 24 public datasets, totaling over 25,000 hours from 634 users, all collected using Empatica E4 devices. UME, an EfficientNet-based model trained using contrastive learning on approximately 275 million 60-second EDA windows, demonstrates strong performance. In 8 out of 10 downstream classification tasks, UME outperforms traditional handcrafted feature baselines and matches the performance of generalist time-series foundation models like Mantis, while requiring 20x fewer computational resources. The project releases all datasets, model weights, and code to foster further research in wearable EDA modeling.

Key takeaway

For AI Scientists and Research Scientists working with physiological signals, UME and the EDAMAME dataset offer a significant advancement in electrodermal activity (EDA) modeling. You should explore integrating UME's pre-trained weights or EDAMAME into your projects to improve performance on EDA-related tasks, especially given its computational efficiency compared to generalist models. This work provides a robust foundation for developing more accurate and resource-friendly wearable health applications.

Key insights

UME is the first open-source foundation model for EDA, outperforming baselines with 20x less computation.

Principles

• Large, diverse datasets are crucial for foundation model training.
• Specialized foundation models can match generalist performance with fewer resources.
• Contrastive learning effectively trains physiological signal foundation models.

Method

The method involves compiling EDAMAME from 24 public datasets, pre-processing EDA signals (filtering, phasic/tonic decomposition, 60-second windowing), and training an EfficientNet architecture with contrastive learning using InfoNCE loss and EDA-specific data augmentations.

In practice

• Utilize EDAMAME for new EDA foundation model development.
• Apply UME's features for cognitive load, stress, and engagement assessment.
• Consider specialized, computationally efficient models for on-device deployment.

Topics

• Electrodermal Activity
• Foundation Models
• Time Series Analysis
• Wearable Sensors
• Self-Supervised Learning

Best for: AI Scientist, Research Scientist, AI Researcher, Machine Learning Engineer, Data Scientist

Related on AIssential

• Towards a General Intelligence and Interface for Wearable Health Data — Large-scale pretraining on wearable data creates a foundation model for diverse health insights.
• PULSE: Privileged Knowledge Transfer from Rich to Deployable Sensors for Embodied Multi-Sensory Learning — PULSE transfers privileged EDA knowledge to low-cost sensors, improving stress detection without requiring EDA at inference.
• Evaluation of EEG Foundation Models for Event-Based Burst-Suppression Detection in ICU — EEG Foundation Models show promise for robust, patient-agnostic burst-suppression detection in ICU settings, even with limited data.
• CogAdapt: Transferring Clinical ECG Foundation Models to Wearable Cognitive Load Assessment via Lead Adaptation — Adapting clinical ECG foundation models via lead transformation and progressive fine-tuning enables robust wearable cognitive load assessment.
• Personalized and Context-Aware Transformer Models for Predicting Post-Intervention Physiological Responses from Wearable Sensor Data — Transformer models can predict personalized physiological responses to stress interventions using wearable sensor data.
• Zyphra Releases ZUNA: A 380M-Parameter BCI Foundation Model for EEG Data, Advancing Noninvasive Thought-to-Text Development — ZUNA is a 380M-parameter BCI foundation model for EEG, using 4D RoPE and diffusion for robust signal reconstruction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.