Detecting Social Breathing : Quantitative signs of interaction in conversational data

Abstract: Human interaction has been a focus of study in psychology for a long time, but until recentlyonly very specific aspects of interaction have been considered. This approach leads to studiesthat gave insights into certain contexts, like parent-child or patient-clinician dynamics [1], [2].There are many theories on narrow subtopics, but no overarching unifying theory describingthe essence of human interaction. To fill this gap, Niclas Kaiser and Emily Butler recentlydeveloped the theory of Social Breathing, considering the participants as parts of a complexsystem [3] which currently lacks a quantitative description. According to their theory, the presence of Social Breathing in an interaction leads to com-plex patterns in multi-variate physiological time series of interacting people. How exactlythis effect could manifest in experimental data remains unclear. However, to distinguishinteraction data from non-interaction data from three independent sources, we tested ex-ploratory analysis methods, including principal component analysis (PCA), cross-wavelettransform and a convolutional neural network. We also developed a Maximum SpectralSimilarity Estimation method based on the cross-wavelet transform. All three data analyzed sets shared the general setup of two participants being in differentvariations of a conversation while one or more (neuro-)physiological variables were tracked. APCA of correlation coefficients we applied to the first data set by Guan et al. from 2015 [4]showed differences in participants’ dynamics, which a support vector machine could capitalizeon with a maximum classification accuracy of 72%. Because physiological dynamics during an interaction are not stable over time, we usedcross-wavelet transforms for time-resolved frequency information. To check for any transientspectral patterns that could be attributed to Social Breathing, we developed Maximum SpectralSimilarity Estimation. It showed that some variables contained spectra that were more similarwithin interaction data compared with less interactional or fake data. This pattern was trueboth for the previously named data source and for the second data source gathered by NiclasKaiser and described in ref. [5]. In this setup, two participants engaged in different stages ofvarying interactional intensity. Contrary to our expectations, interaction distractions resultedin increased similarity. The final experimental setup called NUNA was specifically designed for investigating SocialBreathing in neuro-physiological time series. We used early pilot data from NUNA in this workfor a proof of concept. Training a convolutional neural network on cross-wavelet transformsof functional near-infrared spectroscopy (fNIRS) brain data to recognize reoccurring frequencypatterns of Social Breathing was unsuccessful. Maximum Spectral Similarity Estimation didnot show convincing differences in spectral similarity between different modes of conversationand fake data. We propose adaptations to the experimental setup and the preprocessing of thedata to better identify Social Breathing.