A Time-Series Augmentation Method Based on Empirical Mode Decomposition and Integrated LSTM Neural Network.

Area of Science:

• Biomedical Engineering
• Data Science
• Machine Learning

Background:

• Collecting extensive patient movement data for disease assessment is challenging due to the repetitive nature of evaluations.
• Limited data quantity hinders accurate analysis of conditions like stroke and Parkinson's disease.
• Existing methods struggle with insufficient data for robust motor function assessment.

Purpose of the Study:

• To develop a novel data augmentation technique to address the scarcity of patient movement data.
• To improve the accuracy of disease assessment through enhanced data quantity and quality.
• To validate the effectiveness of the proposed method in clinical settings.

Main Methods:

• Proposed a data augmentation method combining Empirical Mode Decomposition (EMD) and an integrated Long Short-Term Memory neural network (EMD-ILSTM).
• Utilized EMD to decompose raw patient movement signals into intrinsic mode functions.
• Applied LSTM for data augmentation on the decomposed signal components.

Main Results:

• The EMD-ILSTM method demonstrated a 5.2% improvement in classification accuracy on the Ninaweb dataset.
• Clinical trials confirmed that augmented data, after dimensionality reduction, exhibit smaller intra-class and larger inter-class distances compared to raw data.
• The findings indicate enhanced data separability and improved diagnostic potential.

Conclusions:

• The EMD-ILSTM data augmentation technique effectively addresses the challenge of insufficient patient data in motor function assessments.
• The method significantly improves classification accuracy and data discriminability, offering a valuable tool for disease assessment.
• This approach holds promise for advancing quantitative analysis in neurological and rehabilitative medicine.