Autoencoded deep features for semi-automatic, weakly supervised physiological signal labelling.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Physiological Signal Processing

Background:

• Machine learning (ML) models require extensive labeled data, which is often scarce in medical research.
• Missing labels in large datasets, such as microneurography recordings, hinder ML model training.
• Developing efficient data labeling methodologies is crucial for advancing ML in healthcare.

Purpose of the Study:

• To develop a novel semi-supervised, iterative group-labeling methodology using autoencoders.
• To overcome the challenge of missing labels in a large dataset of microneurography recordings.
• To enable robust ML model training for physiological signal analysis.

Main Methods:

• Systematic optimization of autoencoders for feature extraction from 478,621 microneurography signal excerpts.
• K-means clustering of extracted features, followed by iterative expert labeling of signal bursts (valid/non-valid muscle sympathetic nerve activity - MSNA).
• Training a deep neural network on the fully labeled dataset.

Main Results:

• Three autoencoders (fully connected and convolutional neural network-based) were selected for feature learning.
• An iterative labeling process successfully labeled all 478,621 signal excerpts within 13 iterations.
• Deep neural networks trained on the labeled data achieved 93.13% accuracy and 91% AUC ROC in cross-validation.

Conclusions:

• The autoencoder-based iterative labeling procedure efficiently labeled a large physiological signal dataset using expert input.
• This methodology is broadly applicable to various unlabeled datasets requiring expert annotation for ML applications.
• The approach facilitates the development of ML models even with limited initial labeled data.