Foundation Models for Biosignals: A Survey (www.techrxiv.org)

A recent survey explores the emerging landscape of foundation models tailored for biosignals—physiological data like ECG, EEG, PPG, and inertial measurement units that serve as vital indicators of human health and behavior. Building on successes in natural language processing and computer vision, these pretrained neural networks aim to learn robust, generalizable representations across diverse biomedical tasks. This shift is significant for the AI/ML community as it promises to unlock richer, more scalable analyses of complex biosignals, potentially transforming clinical diagnostics and monitoring. The survey identifies three key approaches shaping this field: training foundation models from scratch using large-scale biosignal datasets, adapting existing general-purpose time series models to biomedical contexts, and leveraging multimodal large language models to interpret biosignal patterns. Each strategy presents unique technical challenges and opportunities, from handling diverse data modalities to optimizing pretraining techniques specific to biomedical signals. By synthesizing these efforts, the work charts a comprehensive roadmap for advancing biosignal modeling, highlighting open research directions crucial to realizing clinically impactful AI tools that integrate seamlessly with wearable and medical sensing technologies.