Evaluating Classifier Confidence for Surface EMG Pattern Recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Surface electromyogram (EMG) is a key signal for human-computer interfaces.
• Accurate classification and reliable confidence estimation are crucial for EMG pattern recognition.
• Existing classifiers often struggle to provide well-calibrated confidence scores.

Purpose of the Study:

• To identify classifiers that achieve high accuracy and provide well-calibrated confidence in EMG pattern recognition.
• To compare the performance of discriminative and generative classifiers for EMG data.
• To evaluate the suitability of classifiers for applications requiring reliable confidence estimation.

Main Methods:

• Evaluation of diverse discriminative and generative classifiers on four EMG datasets.
• Quantitative and visual performance analysis of classifier accuracy and confidence calibration.
• Focus on deep neural networks and scale mixture models.

Main Results:

• Deep neural networks demonstrated high accuracy but poor confidence calibration.
• Scale mixture models, a type of generative classifier, showed superior accuracy.
• Scale mixture models provided better confidence estimation, reflecting true probabilities of correctness.

Conclusions:

• Generative classifiers, specifically scale mixture models, are better suited for EMG pattern recognition requiring accurate confidence.
• The ability to account for uncertainty in EMG variance is critical for reliable confidence output.
• This finding has implications for improving the robustness of EMG-based interfaces.