Feature extraction across individual time series observations with spikes using wavelet principal component analysis

Area of Science:

• Biomedical Signal Processing
• Time Series Analysis
• Clinical Data Mining

Background:

• Clinical studies often involve longitudinal data with repeated measurements over time.
• Identifying common temporal patterns in individual time series is crucial for understanding disease progression and treatment effects.
• Extreme observations, such as spikes, can represent significant temporal phenomena requiring specialized analysis.

Purpose of the Study:

• To present a novel methodology for extracting common temporal features from multiple individual time series.
• To explore the utility of wavelet analysis for localized feature extraction in time-varying signals.
• To apply this methodology to fetal movement (FM) count data as a potential screening tool for fetal compromise.

Main Methods:

• Wavelet basis function decomposition was applied to individual time series data.
• Wavelet shrinkage was employed for noise reduction.
• Linear principal component analysis (PCA) was used on wavelet coefficients to extract common features.
• Inverse transformation returned features to the time domain for clinical interpretation.

Main Results:

• The methodology successfully extracted common temporal features from simulated and real-world time series data.
• Analysis of fetal movement (FM) count data revealed distinct temporal patterns.
• The approach demonstrated potential for identifying clinically relevant temporal phenomena in longitudinal data.

Conclusions:

• The proposed methodology offers an effective way to identify common temporal features in complex time series data.
• Wavelet analysis combined with PCA provides a robust tool for analyzing longitudinal clinical data.
• This approach holds promise for developing new screening tools, such as formal FM counting, for improved perinatal outcomes.