A Simple Unsupervised, Real-time Clustering Method for Arterial Blood Pressure Signal Classification

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Accurate biomedical signal analysis requires distinguishing normal signals from noise or artifacts.
• Existing methods may be unsuitable for resource-constrained embedded systems.

Purpose of the Study:

• To develop a novel unsupervised clustering method for classifying arterial blood pressure (ABP) beat cycles as normal or abnormal.
• To create a method suitable for resource-constrained embedded computing contexts.

Main Methods:

• A cycle detection algorithm delineates ABP beat cycles.
• Continuous time Fourier series decomposition models each cycle.
• Fourier series parameters and cycle period form a discrete vector representation.
• A clustering algorithm uses a weighted distance function of normalized cycle parameters.

Main Results:

• The method effectively classifies normal and abnormal ABP beat cycles.
• The weighted distance function accounts for physiological modulations while identifying artifacts.
• Demonstrated effectiveness on challenging cardiac surgery patient signals.

Conclusions:

• The proposed unsupervised clustering method is effective for ABP signal analysis in embedded systems.
• It successfully differentiates normal from abnormal cycles, even with physiological variations and artifacts.
• This approach enhances the reliability of biomedical signal analysis in critical care settings.