Device Passport Improves Biosignal Model Generalization Across Device Layouts.

The challenge of adapting biosignal foundation models to new device layouts, particularly when large datasets for each specific layout are scarce, is a significant hurdle in medical and wearable technology. A new channel embedding technique, named Device Passport, has been introduced to address this by improving the cross-layout transfer capabilities of spatio-temporal pretrained models. Unlike previous embedding methods that typically rely solely on functional information or metadata for fixed positional embeddings, Device Passport learns a combination of experts and mixture models. These models take into account both the functional activity and the metadata of each channel as input. This comprehensive approach allows for a more nuanced and adaptable representation of biosignal data across diverse device configurations. Through controlled subset-transfer experiments and realistic transfer scenarios, such as adapting to ear-EEG data, Device Passport demonstrated competitive overall performance. Crucially, it showed significant improvement over the strongest learned baselines specifically in the challenging layout-transfer regimes that motivated its development. These findings underscore that the design of channel embeddings is a critical factor for successfully reusing large-scale pretrained biosignal models on novel devices.