Clustering Continuous Wavelet Transform Characteristics of Heart Rate Variability through Unsupervised Learning.

Area of Science:

• Biomedical Engineering
• Data Science
• Cardiology

Background:

• Non-invasive physiological signal analysis is crucial for Smart Health and precision medicine.
• Signal complexity and variability hinder accurate analysis, necessitating dimensionality reduction and clustering.
• Machine learning, particularly unsupervised learning, offers powerful tools for physiological time series analysis.

Purpose of the Study:

• To present a novel unsupervised autoencoder architecture, deep convolutional embedded clustering (DCEC), for analyzing heart rate variability (HRV) signals.
• To explore the time-frequency characteristics of HRV using a data-driven approach.
• To investigate the potential of low-dimensional representations for understanding cardiac activity.

Main Methods:

• An unsupervised autoencoder network was trained on continuous wavelet transforms (CWT) of HRV signals.
• The HRV signals were derived from publicly available, annotated electrocardiogram (ECG) records.
• The model learned low-dimensional latent variables representing CWT features for clustering.

Main Results:

• The learned clusters from the autoencoder often corresponded to distinct beat morphologies in the ECG.
• Dimensionality reduction of time-frequency features provided potential new insights into cardiac function.
• The approach demonstrated the utility of deep learning for analyzing complex physiological time series.

Conclusions:

• Deep convolutional embedded clustering is a viable method for analyzing HRV time-frequency characteristics.
• This technique can aid in understanding beat morphologies and provide deeper insights into cardiac activity.
• The findings support the application of machine learning in precision medicine and biomedical research.