PSD-Based Outlier Removal for Enhancing ENG Signal Analysis and Classification

Accurate classification of electroneurographic (ENG) signals is crucial for applications like prosthetic control and neurorehabilitation. However, noise and artifacts in ENG signals can degrade model performance. To address this, we propose a preprocessing method based on power spectral density (PSD) variance for outlier detection. This approach uses a tunable threshold to balance data retention and removal, effectively eliminating windows with abnormal spectral characteristics. Additionally, we optimize training through early stopping based on the F1-score. This combination of filtering and training optimization enhances ENG signal consistency, improving classification performance by up 8%, particularly for animal experiments with higher levels of noise, while reducing performance variability up to 7 times.