Device Passport: Enabling Spatio-Temporal Pretrained Models to Generalize Across Input Layouts

Jul 16, 2026

Upload by Logic

New device layouts create a difficult modeling problem because large datasets are rarely available for each layout. Biosignal foundation models can help if they generalize reliably when sensor montages change. To strengthen cross-layout transfer, the work analyzes how channel embedding choices behave when pretraining layouts differ from downstream decoding layouts. It introduces Device Passport, an expert- and mixture-based channel embedding approach using channel functional activity and metadata, outperforming strong learned baselines in motivating layout-transfer regimes.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

Abstract

Introduction

Approach and Methods

Device Passport: Enabling Spatio-Temporal Pretrained Models to Generalize Across Input Layouts Geeling Chau 1 2 Ran Liu 2 Juri Minxha 2 Wenhui Cui 2 Erdrin Azemi 2 Ellen L. Zippi 2 Behrooz Mahasseni 2 Christopher Michael Sandino 2 arXiv :2607 .00249v1 [ cs .LG] 30 Jun 2026 Abstract New device layouts pose a challenging modeling problem due to the lack of large datasets for each specific layout. Biosignal foundation models offer a plausible solution if they are able to generalize to new layouts effectively. To improve crosslayout transfer, we study how different channel embedding techniques behave when pretraining layouts differ substantially from the downstream decoding layout. We propose Device Passport, anew channel embedding technique that learns experts and mixture models that take each channel’s functional activity and metadata as input. This contrasts with prior embedding methods, which typically use only functional information or only metadata to look up learned or fixed positional embeddings. Across controlled subset-transfer experiments and realistic transfer to ear-EEG, Device Passport is competitive overall and improves over the strongest learned baseline in the layouttransfer regimes that motivate this work. These results suggest that channel embedding design is a key consideration when reusing large-scale pretrained biosignal models on new devices. 1. Introduction Across biosignal domains such as EMG, EKG, and EEG, large-scale biosignal foundation models have been shown to improve downstream decoding, especially when pretraining and evaluation share similar sensor layouts (Kaifosh & Reardon, 2025 ; Abbaspourazad et al., 2023 ; Wang et al., 2024) . However, it is common for channels to be placed in new locations when experimenting with new devices, constituting unseen sensor layouts. In these settings, ex- 1 California Institute of Technology, Pasadena, CA, USA 2Apple, Cupertino, CA, USA. Correspondence to: Christopher Michael Sandino <[csandino@apple.com](csandino@apple.com) >. Proceedings of the Workshop on Structured Data for Health atthe 43rd International Conference on Machine Learning, Seoul, South Korea. Copyright 2026 by the author(s) . tensive pretraining data may not be available or only be available in mismatched layouts, so it would be valuable for such pretrained models to generalize to new device layouts. This problem is central to wearable health sensing, where new form factors often collect high-frequency physiological time series before large device-specific cohorts exist. For many prototype or emerging health devices, collecting large labeled cohorts for every new montage is impractical, so the useful question is not only whether a pretrained biosignal model works, but whether its spatial knowledge can be reused when the sensor layout changes. A core challenge in this setting is channel embedding. Transformers offer channel count flexibility, but rely on positional or channel embeddings to contextualize each input (Vaswaniet al., 2017 ; Chau et al., 2025) . Most positional or lookupbased embedding schemes assume repeated identities or enough examples per channel to relearn useful embeddings after transfer. Under strong layout shift and low-data experimental settings, these assumptions break, and the model must recover spatial relationships for sensor channels whose locations or identities were not available during pretraining. We therefore argue that channel embeddings are a core weakness in unseen-layout transfer. Prior channel embedding methods span identity-based, coordinate-based, and activity-conditioned schemes, including strong recent approaches such as asymmetric channel positional embedding (ACPE) (Wang et al., 2024) . However, it remains unclear how well these methods and ACPE transfer when the downstream montage differs or data are scarce. This motivates a focused study of how to best learn and initialize channel embeddings when layout transfer is large. In par

FAQs

Why is generalizing biosignal foundation models to new device layouts difficult?

Because new devices often use sensor channels placed in new locations, creating unseen sensor layouts. Large pretraining datasets may be unavailable for each layout, so the model must reuse learned spatial knowledge under layout shift.

What is Device Passport and how does it differ from prior channel embedding methods?

Device Passport is a channel embedding technique that learns expert and mixture models using each channel’s functional activity and metadata. Prior methods typically rely only on functional information or only metadata to select or build positional embeddings.

How was Device Passport evaluated and what did the results show?

The approach was tested in controlled subset-transfer experiments and in a realistic transfer to ear-EEG sleep staging. It is competitive overall and improves over the strongest learned baseline, especially in layout-transfer settings that motivate this work.

Jul 16, 2026

Upload by Logic