Integrating heterogeneous classifier ensembles for EMG signal decomposition based on classifier agreement

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyographic (EMG) signal decomposition is crucial for understanding motor control.
• Accurate classification of motor unit potentials (MUPs) is essential for reliable EMG analysis.
• Existing methods face challenges in handling signal variability and complexity.

Purpose of the Study:

• To develop and evaluate a design methodology for integrating heterogeneous classifier ensembles.
• To improve the classification performance of MUPs in EMG signal decomposition.
• To reduce classification errors through a diversity-based hybrid classifier fusion approach.

Main Methods:

• Employed an overproduce and choose strategy for classifier ensemble construction.
• Utilized a diversity-based hybrid classifier fusion approach with two combiner modules.
• Applied the kappa statistic diversity measure to select optimal classifier subsets.
• Integrated adaptive certainty-based, fuzzy k-NN, and matched template filter classifiers.

Main Results:

• The developed system demonstrated superior overall classification performance compared to individual base classifiers.
• Achieved a correct classification rate (CCr) of 93.8% and an error rate (Er) of 2.2% for simulated signals of varying intensity.
• For simulated signals with variability, the system achieved a CCr of 89.1% and an Er of 4.7%.
• On real EMG signals, the system achieved a CCr of 89.4% and an Er of 3.9%.

Conclusions:

• The proposed diversity-based hybrid classifier fusion approach effectively improves MUP classification accuracy.
• The methodology successfully reduces classification errors in EMG signal decomposition.
• This system offers a robust solution for analyzing complex EMG signals.