Automated Detection of Onset and Offset of Multiunit Inspiratory Burst Activity Using a Threshold and Window Independent Approach

Abstract

In animal models, inspiratory electromyogram (EMG) or neurogram (ENG) multiunit activity recorded with various electrode configurations are commonly used for assessing multiple aspects of central respiratory network control. The cyclic inspiratory electrogram ‘bursts’ are complex, and generally characterized by an augmenting fast‐oscillatory amplitude pattern that begins to decrement after peak activity. Accurate automated capture of the bounds of these bursts in time would allow for in‐depth analysis of intra‐burst features, but is difficult to achieve due to high stochasticity of the underlying network interactions, particularly at the decrementing phase of the burst. Common approaches to estimate these bounds often use threshold and/or window‐based smoothing or binning methods that minimize signal variance around the overall trend. These approaches rely on conservation of burst‐to‐burst characteristics to maintain accuracy and consistency; however, features underlying the onset and/or offset of the burst may be obscured and not be adequately captured. To overcome this limitation, we present an unbiased method for automated estimation of burst bounds that is not dependent on thresholding or windowing methods, and assess the capabilities of this method on rectified diaphragm EMG and phrenic ENG time‐series bursts recorded from urethane anesthetized Sprague‐Dawley rats. For our method, each burst is initially identified and approached using an arbitrary time point between adjacent bursts that contains the burst onset and extends toward the center of the burst. A multidimensional time‐invariant distribution is then achieved using time‐delay embedding of all amplitude values that occur at time‐series slope inversions. The variability and magnitude of the distribution are grouped using spectral clustering to identify potential timings for transitions from non‐burst to burst activity, and the maximum second differential of a running variance is identified as the most likely transition and established as the burst onset. The entire process is then repeated from an arbitrary point containing the burst offset to similarly identify the most likely offset transition, after which the next burst is identified and evaluated. This process is then repeated for each burst in the time‐series. The algorithm was evaluated by comparing automated and manual selection of burst bounds in both phrenic nerve ENG and diaphragm EMG recordings. Preliminary comparisons show a mean difference of inspiratory burst onset: 9.0±6.2 ms, offset: 16.6±8.7 ms, and duration: 24.2±11.8 ms in phrenic ENG, and a mean difference of inspiratory burst onset: 7.2±2.8 ms, offset: 36.4±7.3 ms, and duration 37.4±5.8 ms in diaphragm EMG, which is similar to that observed for manual bounds selection between individuals. This algorithm rapidly processes data (approximately 25 bursts/s) and maintains accuracy despite changes in burst amplitude and/or frequency; thus, it presents an efficient and consistent method for high throughput burst identification that may facilitate burst analysis for long duration time series data.Support or Funding InformationNIH NS101737; Thomas Hartman Center for Parkinson’s Disease Research at Stony Brook University