A Deep Machine Learning Method for Classifying Cyclic Time Series of Biological Signals Using Time-Growing Neural...

Area of Science:

• Machine Learning
• Time Series Analysis
• Biomedical Signal Processing

Background:

• Cyclic time series (CTS) analysis is crucial in various fields, including medicine.
• Existing methods often struggle to efficiently capture dynamic patterns in CTS.
• The need for robust classification methods with validated performance is evident.

Purpose of the Study:

• To introduce a novel deep time-growing neural network (DTGNN) for learning cyclic contents of stochastic time series.
• To enhance the classification of cyclic time series using a multiscale learning structure.
• To propose a systematic procedure for parameter selection and a novel validation method for structural risk assessment.

Main Methods:

• Developed the deep time-growing neural network (DTGNN), combining supervised and unsupervised learning.
• Implemented a multiscale learning structure to preserve dynamic time series content.
• Utilized a systematic procedure for design parameter optimization in one-versus-multiple class problems.
• Introduced a novel validation method for quantitative and qualitative structural risk evaluation.

Main Results:

• The DTGNN significantly improves classification performance on cyclic time series.
• The method demonstrates optimal structural risk across multiple medical datasets.
• Statistical validation using repeated random subsampling confirmed the DTGNN's efficacy.

Conclusions:

• The DTGNN offers a powerful and efficient approach for cyclic time series classification.
• The proposed validation method provides robust assessment of classification models.
• DTGNN shows significant promise for applications in biomedical signal analysis.