Brain Responses to Emotional Stimuli During Breath Holding and Hypoxia: An Approach Based on the Independent Component Analysis

Abstract

Voluntary breath holding represents a physiological model of hypoxia. It consists of two phases of oxygen saturation dynamics: an initial slow decrease (normoxic phase) followed by a rapid drop (hypoxic phase) during which transitory neurological symptoms as well as slight impairment of integrated cerebral functions, such as emotional processing, can occur. This study investigated how breath holding affects emotional processing. To this aim we characterized the modulation of event-related potentials (ERPs) evoked by emotional-laden pictures as a function of breath holding time course. We recorded ERPs during free breathing and breath holding performed in air by elite apnea divers. We modeled brain responses during free breathing with four independent components distributed over different brain areas derived by an approach based on the independent component analysis (ICASSO). We described ERP changes during breath holding by estimating amplitude scaling and time shifting of the same components (component adaptation analysis). Component 1 included the main EEG features of emotional processing, had a posterior localization and did not change during breath holding; component 2, localized over temporo-frontal regions, was present only in unpleasant stimuli responses and decreased during breath holding, with no differences between breath holding phases; component 3, localized on the fronto-central midline regions, showed phase-independent breath holding decreases; component 4, quite widespread but with frontal prevalence, decreased in parallel with the hypoxic trend. The spatial localization of these components was compatible with a set of processing modules that affects the automatic and intentional controls of attention. The reduction of unpleasant-related ERP components suggests that the evaluation of aversive and/or possibly dangerous situations might be altered during breath holding.

Appendix: Component Adaptation Analysis

Component adaptation analysis is a parametric approach designed to investigate the changes of independent components related to a target condition compared to a reference, well characterized, condition. For example, in the present application a satisfactorily large number of ERP trials was collected during the free-breathing condition, thus allowing a robust estimate of independent components; the goal was then to determine possible changes in ERP components during breath holding phases. In the present application, the free breathing condition was characterized by means of a satisfactory number of ERP trials, which allowed a robust estimate of r independent components (the number r was determined based on the average quality index provided by ICASSO). On the other side, the number of ERP trials collected during each breath holding (breath holding) phase was low and could not allow stable results. Thus, in order to overcome this issue and compare the two conditions, we introduce CAA in order to estimate, under some assumptions, the components during breath holding as variations from those estimated during free breathing.

Indeed, CAA assumes that the activation maps of the components embedded in the signals do not change from the reference condition (the free breathing) throughout breath holding, instead both amplitude and latency of each component are free to change.

Thus, CAA derives the independent components of the target condition by imposing the same demixing transformation (matrix W) of the reference condition to trial data (x_j with j = 1, …, n channels) related to the target condition and estimates how to rescale (by the multiplicative factor A) and temporally shift (to anticipate or to delay by the additive factor T) each component of the reference condition in order to fit into the corresponding target component. The search of optimal parameters is the core of CAA and consists in minimizing the error in reproducing each target component starting from the reference one. The optimal parameters are those that minimized the least square difference between the back-projected target trials w_iX and the amplitude-scaled, time-shifted reference component. Mathematically the problem can be described as that of solving

$$\arg \begin{array}{*{20}c} {\hbox{min} } \\ {A_{i} ,T_{i} \in D} \\ \end{array} \sum {\left( {w_{i} X - S_{i} \left( {A_{i} ,T_{i} } \right)} \right)^{2} } ,$$

where the sum Σ is over time samples; w_i is the i-th column of the demixing matrix W that projects the target signals X into the component space of the reference condition, and S_i (i = 1, … r) is the corresponding component of the reference condition, amplitude scaled and time shifted:

$$S_{i} \left( {A_{i} ,T_{i} } \right) = A_{i} \cdot S_{i} \left( {t + T_{i} } \right).$$

It can be shown that minimizing the error on reproducing each component is mathematically equivalent to minimize the global error on reproducing the target signals starting from the reference components. In other words, thanks to the mutual independence of the components the optimal parameters search problem is separable and it is equivalent to the minimizations carried out separately on each component. This property makes the minimization problem mathematically tractable since we can search several (i = 1, … r components) minimum errors in the two dimension space of the Ai and Ti pairs instead of searching a global minimum in a multidimensional space (A₁, T₁, …, A_i, T_i, …, A_r, T_r).

In addition to this, the joint usage of CAA and ICASSO allows to find the optimal A and T parameters with their confidence errors. Component adaptation analysis is performed in fact using the sets of components provided by the multiple runs of ICASSO. Indeed, in the present CAA application we scaled and shifted each BAT and we used the other activity templates from the ICA runs within ICASSO related to the same component in order to derive the confidence intervals of the optimal A and T parameters.