BioCross Repository

This repository contains the source code developed for the study:

BioCross: A Cross-Modal Framework for Unified Representation of Multi-Modal Biosignals with Heterogeneous Metadata Fusion

Publish:Information Fusion 123 (2025) 103302. https://doi.org/10.1016/j.inffus.2025.103302.

Overview

BioCross is a cross-modal framework designed to unify representations of multi-modal biosignals while incorporating heterogeneous metadata. It enables seamless integration of physiological signals such as ECG, PPG, and ABP with additional contextual metadata for enhanced diagnostic and predictive capabilities.

Features

• A mask-merge strategy to Variational Autoencoders architecture to align various modalities in a Gaussian latent space, which facilitates the effective representation of multi-sensor physiological data.

• Utilize the frequency-based attention mechanism and cross-attention to fuse embeddings of biosignals and metadata, includes circadian rhythms, enhancing the interaction of heterogeneous data.

• The product-of-experts technique to handle missing waveform inputs by enabling composable biosignal inputs. The output can either be certain modality generation or disease prediction via distinct decoders.

Repository Structure

├── model/                # Model architectures and configurations
├── utils/                 # Helper functions for data handling and model utilities
├── requirements.txt       # Python dependencies
├── Dt.py                  # Data processing script
├── S1_main_BioCross.py    # Main script for training and evaluation
├── S2_downstream_BioCross.py    # Downstream task training script
├── README.md              # Repository documentation (this file)

Contact

For questions or collaborations, please contact:

• Name: Mengxiao Wang
• Email: [email protected]