BioX-Bridge: Model Bridging for Unsupervised Cross-Modal Knowledge Transfer across Biosignals

TL;DR

BioX-Bridge addresses the challenge of unsupervised cross-modal knowledge transfer between biosignals under limited labeled data and large foundation models. It introduces a lightweight bridge network that aligns intermediate representations across modalities, using a two-stage bridge position selection and a prototype-based, low-rank projection to enable information flow from a new modality to an old modality. The training objective focuses on aligning the bridged representations with the old-modality representations, while only training the bridge parameters to minimize the alignment loss. Across ISRUC, FOG, and WESAD, BioX-Bridge achieves transfer performance comparable to or better than baselines while reducing trainable parameters by about 88–99%, demonstrating practical efficiency for modality-agnostic, resource-constrained biosignal applications.

Abstract

Biosignals offer valuable insights into the physiological states of the human body. Although biosignal modalities differ in functionality, signal fidelity, sensor comfort, and cost, they are often intercorrelated, reflecting the holistic and interconnected nature of human physiology. This opens up the possibility of performing the same tasks using alternative biosignal modalities, thereby improving the accessibility, usability, and adaptability of health monitoring systems. However, the limited availability of large labeled datasets presents challenges for training models tailored to specific tasks and modalities of interest. Unsupervised cross-modal knowledge transfer offers a promising solution by leveraging knowledge from an existing modality to support model training for a new modality. Existing methods are typically based on knowledge distillation, which requires running a teacher model alongside student model training, resulting in high computational and memory overhead. This challenge is further exacerbated by the recent development of foundation models that demonstrate superior performance and generalization across tasks at the cost of large model sizes. To this end, we explore a new framework for unsupervised cross-modal knowledge transfer of biosignals by training a lightweight bridge network to align the intermediate representations and enable information flow between foundation models and across modalities. Specifically, we introduce an efficient strategy for selecting alignment positions where the bridge should be constructed, along with a flexible prototype network as the bridge architecture. Extensive experiments across multiple biosignal modalities, tasks, and datasets show that BioX-Bridge reduces the number of trainable parameters by 88--99\% while maintaining or even improving transfer performance compared to state-of-the-art methods.

Figures (5)

• Figure 1: Comparison of unsupervised cross-modal knowledge transfer methods for biosignals. The red arrow indicates loss computation.
• Figure 2: Overview of BioX-Bridge. (a) At the training stage, the bridge learns to project intermediate representations from the new modality to the old modality, such that it mimics the output of the old modality model. (b) At the inference stage, the bridge has been constructed and enables the flow of information between the two models in order to make predictions on data from the new modality. (c) The bridge consists of a low-rank approximation module and a prototype set. The low-rank approximation module generates aggregation weights for the prototype vectors.
• Figure 3: BioX-Bridge learning procedure
• Figure 4: Bridge Training Ablation. Blue: Balanced Accuracy. Red: Number of Parameters. We vary (a) bridge rank, (b) number of prototypes, and (c) pair dataset size to understand the robustness of BioX-Bridge and its performance under a low-data regime.
• Figure A1: Illustration of Dataset Split. The dataset is divided into four subject-independent subsets. We first use the old modality data from $\mathcal{D}^{\text{(old)}}$ to train the linear prober $g_\omega^{\text{(old)}}$ for experiment setup, followed by unsupervised training on $\mathcal{D}^{\text{(pair)}}$. The subsets $\mathcal{D}^{\text{(val)}}$ and $\mathcal{D}^{\text{(new)}}$ are used for hyperparameter selection and testing, respectively.