DUDE: deep unsupervised domain adaptation using variable nEighbors for physiological time series analysis.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Medical Informatics

Background:

• Deep learning excels with consistent data distributions for physiological signals like ECG.
• Real-world deployment faces challenges due to distribution shifts between training and deployment data.
• Non-overlapping source and target domain supports present a significant hurdle for model generalization.

Purpose of the Study:

• To introduce a novel framework, Deep Unsupervised Domain adaptation using variable nEighbors (DUDE), for continuous physiological signal analysis.
• To address the challenge of domain adaptation where source and target data distributions differ significantly.
• To improve the generalizability of deep learning models in medical applications.

Main Methods:

• Developed a novel contrastive loss function between source and target domains.
• Implemented a dynamic neighbor selection strategy, adaptively determining neighbors based on latent space density.
• Utilized multiple real-world datasets with diverse target domain characteristics (demographics, ethnicities, geographies, comorbidities).

Main Results:

• DUDE demonstrated superior performance compared to existing baseline methods.
• Achieved up to a 16% improvement over the Nearest-Neighbor Contrastive Learning of Visual Representations strategy.
• Validated effectiveness across real-world datasets with significant domain shifts.

Conclusions:

• DUDE effectively bridges the domain adaptation gap in medical applications.
• The framework shows potential for enhancing the precision and adaptability of diagnostic tools.
• This work paves the way for more robust AI in patient care through improved model generalizability.