Figure 2. Artificial neural network reliably predicts behavioral state from EEG and EMG features.

(A) Example data from an individual rat depicts how electrophysiological features alone reliably segregate behavioral state for most 4 s epochs, but fail to resolve ambiguity of many boundary cases. (B) Artificial neural network architecture employed in the current study to enhance the predictive utility of EEG/EMG features for behavioral state classification (see “Methods” for detailed characteristics of input features used). (C) Reductions in categorical cross-entropy loss and increased prediction accuracy of training data occur across model training. Data depict training of a single model. (D) Activation maps across the four layers of our artificial neural network in response to EEG/EMG features. Within each layer, artificial neurons are depicted as colored boxes with color representing the amount of activation of that neuron in response to input features from a 4 s epoch. Each of the six rows depict activation maps from a unique 4 s epoch from the same rat; two epochs of each behavioral state were chosen to show similarities and differences in activation within and across data from each behavioral state. Note, activation of specific neurons in the output layer corresponds directly with the predicted output of the model which matches the manual scoring for each of the epochs presented above.