FEATURE EXTRACTION OF BRAIN ACTIVITY PRECEDING PUPIL DILATION USING MACHINE LEARNING

We investigated the dynamics of cerebral cortical activity that precedes and contributes to stimulus-evoked pupil dilation. Our approach involved modeling these cortical features using machine learning techniques and performing a subsequent feature importance analysis. Specifically, pupil dilation and corresponding large-scale cortical activity in mice were simultaneously visualized and recorded using wide-field calcium imaging across various conditions, including somatosensory stimulation, visual pattern stimuli, visual context stimuli, and a no-stimulus condition. To quantify the predictive power of the neural signals, a Recurrent Neural Network (RNN) model was developed to predict pupil dilation based on the preceding cerebral cortical activity. The features of brain activity essential for this accurate prediction were subsequently identified through a combination of Independent Component Analysis (ICA) for robust feature extraction and SHAP (SHapley Additive exPlanations) analysis to determine feature importance. Furthermore, we employed Principal Component Analysis (PCA) on the SHAP results to classify and examine potential differences in these predictive features across the various stimulus types. The analysis revealed that it is possible to accurately predict the time course of pupil dilation from the cortical activity measured from stimulus onset up to the dilation's peak. A significant finding was that the predictive brain activity features were largely conserved and independent of the specific sensory modality of the stimulus. However, the features diverged and presented a distinct pattern when predicting spontaneous pupil dilation, suggesting a unique underlying neural mechanism for internally generated arousal compared to externally driven responses.