Deep Learning on 1-D Biosignals: a Taxonomy-based Survey.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Medical Informatics

Background:

• Deep learning, particularly convolutional neural networks (CNNs), has advanced medical imaging.
• Biomedical signal analysis has not yet fully adopted deep learning approaches for computer-aided diagnosis.
• A comprehensive review is needed to organize the growing applications of deep learning in biosignal analysis.

Purpose of the Study:

• To review deep learning techniques applied to biosignal analysis for computer-aided diagnosis.
• To establish a taxonomy for classifying existing deep learning applications in this field.
• To identify trends and gaps in the current research landscape.

Main Methods:

• A systematic literature search was conducted across PubMed, Scopus, and ACM databases.
• Deep learning models were categorized based on biosignal characteristics, application goals, data properties, and network architecture.
• Key classification parameters included signal origin, dimension, type, application goal, ground truth data, learning schedule, and model topology.

Main Results:

• 71 papers published between 2010-2017 focused on deep learning for biosignal analysis.
• Electrocardiography (ECG) signals were the most studied (36 papers), primarily for pattern detection (25 papers).
• Convolutional neural networks (CNNs) were the most frequently used models (34 papers), while restricted Boltzmann machines were least used (15 papers).

Conclusions:

• Significant variability exists in deep learning approaches for biosignal analysis regarding data, applications, and network topology.
• Future research should prioritize standardizing deep learning architectures and optimizing network parameters for enhanced performance and robustness.
• Application-driven research and the incorporation of updated training data from mobile recordings are crucial for advancing the field.