.png){kind=icon}
.png){kind=icon}
Return
Multi-scale and multi-modal contrastive learning network for biomedical time series
.png){kind=icon}
OA
AI
DOI：10.1016/j.bspc.2025.107697.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multi-modal biomedical time series (MBTS) data, integrating physiological information from multiple sources, is essential for various biomedical tasks. However, existing deep learning models struggle to capture robust MBTS representations due to intricate dynamics and inter-modal distribution disparities. This paper introduces the Multi-scale and Multi-modal Biomedical Time Series Contrastive Learning (M2CL) network to overcome these limitations. M2CL initially groups MBTS data based on inter-modal distances, minimizing variation within each group. Separate encoders are then utilized to model each group effectively. To facilitate multi-scale feature extraction, we design diverse patch lengths and mask ratios, generating tokens that encapsulate semantic information at multiple scales and varied contextual perspectives. Furthermore, cross-modal contrastive learning is employed to enhance consistency across inter-modal groups, ensuring the retention of valuable information while reducing noise. Experimental results reveal that M2CL significantly outperforms state-of-theart (SOTA) models, achieving a 33.9% reduction in mean average error (MAE) on respiration rate dataset using Photoplethysmography (PPG) signals, a 13.8% decrease in MAE on exercise heart rate dataset using PPG and 3-axis acceleration signals, a 1.41% increase inaccuracy on the human activity recognition dataset using 9-axis acceleration signals, and a 1.14% improvement in F1-score on obstructive sleep apnea-hypopnea syndrome (OSAHS) dataset using PPG and SpO2. Additionally, on a self-collected OSAHS dataset, which includes PPG and SpO2 signals from smart ring, M2CL surpasses SOTA models by 2.72% in F1-score and 1.40% inaccuracy.
Keywords:
Machine learning
Multi-scale
Contrastive learning
Obstructive sleep apnea-hypopnea syndrome
Heart rate
Smart ring
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
4.9
Papers:
9.8K
Citations:
2.4W