Fig 2. Classifying imminent feeding from other behavior with VS oscillations.
A. Overview of analysis flow. Data was binned and power and coherence extracted as in Fig 1. Orange box: Eighty percent of all pre-feeding bins immediately before feeding bins and all other bins outside of feeding (not-feeding and rest) are used for model training and the other 20% are used for testing. Using the same data, the assignment of category to predictor variables is permuted and trained on 80% of the data and tested on the left out 20%. This analysis produces two receiver operator (ROC) curves representing the performance of models built from actual and permuted data. To build rare event detectors (feeding is only ~18% of the data) the adaptive synthetic sampling approach for imbalanced learning was found to be the best option (Fig 2-1A-B). To be able to compare across models it was also necessary to determine the number of samples were required to achieve stable model performance (Fig 2-1A-D). B. Average performance of differentiating pre-feeding immediately before feeding begins and bins outside of feeding (not-feeding and rest); performance distribution statistics reported as AUC mean ± 95% confidence interval and the effect size between Actual and Permuted given as Cohen’s d: Actual = 0.81±0.03; Permuted = 0.55±0.06; d = 2.68. C. Average areas under the ROC curves () obtained for 20 iterations of model building using models built from the pre-feeding bins immediately preceding feeding and tested on all other pre-feeding bins grouped by amount of time before feeding up to bins centered 42.5 seconds before feeding. D. Percent of pre-feeding data which overlaps with approach behavior; pre-feeding bins centered up to 12.5 seconds before feeding have at most 30% overlap with approach behavior. E-F. Normalized representative LFP features plotted through time around feeding. Features were normalized by either total power or by average coherence of the channel(s) of interest and averaged across all bins and animals around the beginning and end of feeding epochs. Shading indicates ±1 standard deviation. Black traces represent averages before and after feeding and blue traces represent averages during feeding. Dashed blue and solid black lines indicate average feature activity for either feeding bins or all other bins including those outside of these plots. E. Core left core right high gamma coherence increases during feeding. F. Shell left high gamma power decreases during feeding. Although both E and F exhibit a mean shift; a decrease in feature variance around feeding was also found (Fig 2–2).