Quantifying the effects of vagus nerve stimulation on gastric myoelectric activity in ferrets using an interpretable machine learning approach

Figures

• 

  Fig 1. Correlation heatmap of the engineered features (See Table 1) a) baseline vs. VNS at 10 Hz b) baseline vs. VNS at 30 Hz. Each cell’s color shows to what extent features are correlated.

• 

  Fig 2. Selected features via our feature selection algorithm for baseline vs. VNS at 10 Hz (first scenario) and organized in a descending sequence of significance, as established by the Random Forest classifier. Error bars represent one standard deviation. ’*’ is used to show the time domain features. SmEn: Sample Entropy, RMS: Root Mean Square, MVN: (Entropy of) Multi-variate Normal, SVDEn: Singular Value Decomposition Entropy, LL: Line Length, DP: Dominant Power.

• 

  Fig 3. This illustration compares the AUC for a Random Forest classifier trained using various feature sets. From left to right: Features from Fig 2, features selected for VNS at 30 Hz, all features, a random selection of features equivalent in number to those in Fig 2, and randomly shuffled labels. The comparison is conducted in the context of the baseline versus VNS at 10 Hz scenario.

• 

  Fig 4. Selected features via our feature selection algorithm for baseline vs. VNS at 30 Hz (second scenario) and organized in a descending sequence of significance, as established by the Random Forest classifier. Error bars represent one standard deviation. ’*’ is used to show the time domain features. PFD: Petrosian Fractal Dimension, PrEn: Permutation Entropy, DF: Dominant Frequency, ApEn: Approximate Entropy, DP: Dominant Power, BP2: Band Power in 8–11 cpm, MF: Median Frequency, LZ: Lempel-Ziv Complexity, SmEn: Sample Entropy, SpEn: Spectral Entropy.

• 

  Fig 5. This illustration compares the AUC for a Random Forest classifier trained using various feature sets. From left to right: Features from Fig 4, features selected for VNS at 10 Hz, all features, a random selection of features equivalent in number to those in Fig 4, and randomly shuffled labels. The comparison is conducted in the context of the baseline versus VNS at 30 Hz scenario.

• 

  Fig 6. Electrode placements. a) show the structure of nerve cuff, b) position of surface electrodes (only ‘gastric 2’ was used in this study) and nerve cuff. c) A ferret’s head connection to the vagus nerve stimulator and EGG recording device.

• 

  Fig 7. The analytic pipeline for EGG analysis and biomarker identification.

• 

  Fig 8. Feature selection algorithm.