How does downsampling affect needle electromyography signals? A generalisable workflow for understanding downsampling effects on high-frequency time series

TL;DR

This work addresses the challenge of high-frequency nEMG signal analysis by proposing a generalisable five-step workflow to evaluate downsampling effects on information content and classifier performance. It combines shape-based distortion metrics, a feature-based classification pipeline using tsfresh and Boruta with a random forest, and a pairwise ranking model (XGBoost) to identify downsampling configurations that preserve diagnostic information while significantly reducing computation. Using the EMGLAB dataset for a three-class neuromuscular disease task, shape-aware downsampling methods (notably LTTB and MinMaxLTTB) maintained classification accuracy up to factors around $k=30$ with ~60× feature-extraction speedups, while SHAP analysis highlighted envelope correlation and amplitude distributions as key predictors of good performance. The workflow provides practical guidance for near real-time nEMG analysis and is extensible to other high-frequency time series, enabling standardized assessment of data reduction versus model performance across domains.

Abstract

Automated analysis of needle electromyography (nEMG) signals is emerging as a tool to support the detection of neuromuscular diseases (NMDs), yet the signals' high and heterogeneous sampling rates pose substantial computational challenges for feature-based machine-learning models, particularly for near real-time analysis. Downsampling offers a potential solution, but its impact on diagnostic signal content and classification performance remains insufficiently understood. This study presents a workflow for systematically evaluating information loss caused by downsampling in high-frequency time series. The workflow combines shape-based distortion metrics with classification outcomes from available feature-based machine learning models and feature space analysis to quantify how different downsampling algorithms and factors affect both waveform integrity and predictive performance. We use a three-class NMD classification task to experimentally evaluate the workflow. We demonstrate how the workflow identifies downsampling configurations that preserve diagnostic information while substantially reducing computational load. Analysis of shape-based distortion metrics showed that shape-aware downsampling algorithms outperform standard decimation, as they better preserve peak structure and overall signal morphology. The results provide practical guidance for selecting downsampling configurations that enable near real-time nEMG analysis and highlight a generalisable workflow that can be used to balance data reduction with model performance in other high-frequency time-series applications as well.

Figures (8)

• Figure 1: Five steps of the proposed workflow to investigate the effects of downsampling on time series.
• Figure 2: Effects of five downsampling methods on a synthetic MUAP-like waveform. Shaded regions indicate the value groups (when applicable) used by each algorithm. Typical effects are visible: B) Decimate preserves MUAP duration but loses peak amplitude and phase detail. C) MinMax preserves amplitude and duration but alters phase count. D) M4 maintains amplitude but distorts duration and phases. E) and F) LTTB and MinMaxLTTB yield the most faithful approximations, preserving amplitude, duration, and phase structure.
• Figure 3: Accuracy per downsampling factor on the EMGLAB dataset after downsampling. Each line represents a downsampling algorithm, and the highlighted stars indicate the critical factor ($p < 0.05$) for each algorithm. The horizontal dashed line corresponds to the ROC AUC achieved on the original (non-downsampled) data, and the grey area represents its standard deviation across the 10 folds of cross-validation. Stars represent for each algorithm the first factor with a critical difference in classification accuracy.
• Figure 5: Reported SHAP values for the distance metrics used to predict winning downsampling configuration. For interpretability, the colour scale corresponds to the degree of similarity with the original signal: bad indicates a metric value representing greater dissimilarity from the original, while good indicates higher similarity. To ensure consistency across metrics, the values of correlation-based measures were inverted. The preservation trade-off between envelope correlation, Euclidean PSD, Kurtosis and ZCR can be seen in Figure \ref{['fig:signal_diff']}.
• Figure 6: Comparison of the LTTB and Decimate downsampling algorithms on a single EMGLAB signal (factor 25). (A) and (B) illustrate the time-series preservation quality of Decimate and LTTB, respectively. (C) presents the Metric Profiles using four normalised distance metrics, where a lower value indicates better preservation. While Decimate demonstrates superior preservation of high-frequency content and zero-crossing dynamics (Euclidean PSD and Zero Crossing Rate), LTTB achieves significantly better preservation of the signal's morphology and amplitude distribution, as characterised by lower Pearson Envelope Correlation and Kurtosis distances, which is considered a key signal characteristic for clinicians.