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

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.

• The paper presents BioX-Bridge, a model bridging method that efficiently transfers knowledge across biosignal modalities using a prototype-based low-rank network.
• It achieves state-of-the-art results by reducing trainable parameters by 88–99% while outperforming baselines by 1–2% in key metrics.
• Practical evaluations on datasets like WESAD, FOG, and ISRUC demonstrate robustness, efficiency, and flexibility across different biosignal types.

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

Motivation and Problem Setting

The paper addresses the challenge of unsupervised cross-modal knowledge transfer in biosignal analysis, where models trained on one modality (e.g., EEG, ECG, PPG, EMG) are leveraged to enable predictive modeling on another modality with limited or no labeled data. This is motivated by the practical constraints of biosignal acquisition: different modalities vary in sensor comfort, cost, and signal fidelity, and large labeled datasets are often unavailable for new or underrepresented modalities. Existing approaches—primarily knowledge distillation and data translation—are either computationally expensive or limited in their applicability across diverse biosignal pairs.

Figure 1: Comparison of unsupervised cross-modal knowledge transfer methods for biosignals. The red arrow indicates loss computation.

BioX-Bridge Framework

BioX-Bridge introduces a model bridging paradigm, constructing a lightweight bridge network that projects intermediate representations from a new modality model into the representation space of an old modality model. This enables the use of powerful, pre-trained foundation models for inference on new modalities without retraining large models or requiring extensive paired data.

Figure 2: Overview of BioX-Bridge. (a) Training: the bridge learns to project intermediate representations from the new modality to the old modality, mimicking the output of the old modality model. (b) Inference: the bridge enables predictions on new modality data. (c) The bridge consists of a low-rank approximation module and a prototype set.

Bridge Position Selection

A critical aspect of BioX-Bridge is the selection of optimal input and output positions for the bridge within the respective models. The framework employs a two-stage strategy:

1. Input Position (New Modality): Linear probing is used to identify the layer whose representations are most discriminative with respect to pseudo-labels generated by the old modality model.
2. Output Position (Old Modality): Linear CKA is used to select the layer whose representations are most similar to those of the selected input layer in the new modality model.

This approach avoids brute-force search over all possible layer pairs, significantly reducing computational overhead while maximizing transfer performance.

Figure 3: BioX-Bridge learning procedure.

Bridge Architecture

The bridge network is designed for parameter efficiency and flexibility. Instead of a full-rank linear projection (which would require billions of parameters for high-dimensional representations), BioX-Bridge employs a prototype network comprising:

• Prototype Set: Learnable vectors initialized from the old modality model's representations.
• Low-Rank Approximation Module: Factorized matrices generate aggregation weights for the prototypes, enabling effective high-dimensional projection with orders-of-magnitude fewer parameters.

This architecture supports both CNN and transformer-based foundation models and is compatible with arbitrary biosignal modalities.

Experimental Evaluation

BioX-Bridge is evaluated on three biosignal datasets (WESAD, FOG, ISRUC) spanning four modalities and six transfer directions. The backbone models include LaBraM (EEG), HuBERT-ECG (ECG), PaPaGei (PPG), and NormWear (EMG), all initialized with pre-trained weights.

Main Results

BioX-Bridge achieves comparable or superior performance to state-of-the-art baselines (knowledge distillation, contrastive distillation, CardioGAN) while reducing the number of trainable parameters by 88–99%. For example, in WESAD (PPG→ECG), BioX-Bridge uses only 1.3% of the parameters required by KD and outperforms it by 1–2% across balanced accuracy, F1-macro, and F1-weighted metrics.

Notably, in several cases, BioX-Bridge and KD-Contrast surpass the supervised Oracle in balanced accuracy, indicating improved recall at the expense of precision—a consequence of the transfer learning setup.

Ablation Studies

• Bridge Rank and Prototype Set: Performance peaks at intermediate values of rank and prototype set size, with under- or over-parameterization leading to degradation.
• Paired Dataset Size: Transfer performance degrades gracefully with reduced paired data, demonstrating robustness in low-data regimes.
• Bridge Position Selection: The proposed selection strategy yields higher performance than fixed layer choices.
• Foundation Model Choice: BioX-Bridge maintains efficiency and performance even when the old modality model is replaced with larger or alternative foundation models, with more pronounced gains as model size increases.

  Figure 4: Bridge Training Ablation. Blue: Balanced Accuracy. Red: Number of Parameters. Varying bridge rank, number of prototypes, and paired dataset size demonstrates robustness and efficiency.

Implementation Considerations

BioX-Bridge requires only the bridge network to be trained, with all backbone models frozen. The framework is compatible with PyTorch's forward hooks for extracting and replacing intermediate representations. Hyperparameters (bridge rank, prototype set size) are selected via grid search. Training is feasible on commodity GPUs (V100, 32GB VRAM), with single runs ranging from minutes to a few hours depending on model size and dataset.

The method assumes the availability of pre-trained models for each modality and paired data for bridge training. While this is a limitation for emerging biosignals, the approach is extensible to unpaired data via future work in unsupervised domain adaptation.

Figure 5: Illustration of Dataset Split. The dataset is divided into four subject-independent subsets for training, validation, and testing.

Theoretical and Practical Implications

BioX-Bridge demonstrates that model bridging is a viable alternative to knowledge distillation and data translation for cross-modal transfer in biosignals. The prototype-based, low-rank architecture enables efficient alignment of high-dimensional representations, facilitating interoperability between foundation models. The two-stage bridge position selection strategy is empirically validated to maximize transfer performance.

Practically, BioX-Bridge enables the deployment of biosignal models in resource-constrained environments, supporting real-world health monitoring applications where labeled data and computational resources are limited. The framework is modality-agnostic and compatible with both traditional and foundation models.

Future Directions

Potential extensions include:

• Transfer learning with unpaired data via adversarial or contrastive objectives.
• Application to emerging biosignal modalities lacking pre-trained models.
• Integration with multimodal foundation models and connector architectures for general-purpose biosignal analysis.
• Exploration of bridge architectures for other domains (e.g., vision-language, sensor fusion).

Conclusion

BioX-Bridge provides an efficient, flexible framework for unsupervised cross-modal knowledge transfer in biosignal analysis. By leveraging model bridging and prototype-based low-rank projection, it achieves strong transfer performance with minimal computational overhead. The approach is robust across modalities, tasks, and model architectures, and offers a scalable solution for real-world biosignal applications where data and resources are limited.