Adapting LLMs for Ballistocardiographic Signals: A Multi-Task Learning Framework for BCG to ECG Reconstruction

Ballistocardiography (BCG) offers a non-invasive method for continuous cardiac monitoring, yet its weak and noisy signals pose challenges for accurate heart rate (HR) estimation. We propose a novel multi-task learning framework that leverages large language models (LLMs) to reconstruct electrocardiography (ECG) signals from BCG inputs, with auxiliary tasks of heart rate prediction and heart rate anomaly detection aimed at enhancing reconstruction accuracy. Utilizing a Transformer-based architecture, our model captures both local and global features of BCG signals, facilitating precise ECG waveform reconstruction. The multi-task approach enhances performance through shared representations across tasks. Experimental results demonstrate that our framework achieves a Pearson correlation coefficient of 0.927 for ECG reconstruction and a relative Root Mean Square Error (rRMSE) of 0.113 for heart rate prediction, outperforming single-task models. Our primary contribution lies in adapting pre-trained LLMs to BCG signals, achieving significant advancements in modality alignment and multimodal fusion. This work substantially improves the accuracy of ECG signal reconstruction from BCG inputs, providing a robust solution for continuous and non-invasive cardiovascular health monitoring.