Explainable and Robust Deep Forests for EMG-Force Modeling.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Machine and deep learning models are increasingly used for estimating finger forces from high-density surface electromyography (HDsEMG) for neural interfacing.
• Existing models often function as black boxes and struggle with performance degradation due to noisy signals.

Purpose of the Study:

• To propose and evaluate an explainable and noise-robust forest ensemble model for HDsEMG-based finger force estimation.
• To investigate the impact of forest model depth on EMG-force modeling accuracy.
• To assess the model's performance in realistic scenarios with delayed data acquisition and signal noise.

Main Methods:

• Utilized a forest ensemble model, specifically exploring deep forests, for HDsEMG-force modeling.
• Evaluated model performance on a finger force estimation task using data acquired days apart.
• Introduced artificial signal distortion to assess robustness against noise.
• Employed the mean decrease impurity (MDI) metric for model explainability.

Main Results:

• Deep forests significantly outperformed other machine learning models in finger force estimation.
• The deep forest model demonstrated superior robustness against artificial signal distortion, reducing error by 50% compared to baselines.
• Model explanations using MDI revealed a strong correspondence between the model's feature importance and physiological principles.

Conclusions:

• Deep forest models offer a promising, explainable, and noise-robust approach for HDsEMG-based finger force estimation.
• The findings suggest that deep forests can enhance the reliability and accuracy of neural interfaces.
• The explainability of the model aids in understanding the relationship between EMG signals and finger forces.