Physiological sensor data cleaning with autoencoders.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Physiological sensor data is crucial for remote patient monitoring but often suffers from noise.
• Existing feature-based signal cleaning models may not fully capture signal characteristics.

Purpose of the Study:

• To develop a deep learning framework for effective physiological sensor signal cleaning.
• To address the challenge of limited annotated data using a semi-supervised approach.

Main Methods:

• A deep learning framework utilizing dilated convolutions for signal cleaning.
• An autoencoder-based semi-supervised model to learn signal representations and improve interpretability.
• Classification of signals as noisy or clean based on learned features.

Main Results:

• The proposed deep learning models achieved over 8% higher accuracy than feature-based methods.
• False positive and negative rates were halved compared to existing approaches.
• The semi-supervised model, with tuning, outperformed supervised methods, achieving over 90% accuracy.

Conclusions:

• Deep learning, particularly semi-supervised methods, offers a powerful approach for cleaning noisy physiological sensor data.
• This technique can enhance the reliability of remote patient monitoring and support clinical trial decision-making.
• The developed framework can lead to more dependable features for assessing drug efficacy.