Using Computer Models of Language to Evaluate Movie Viewing in Autistic and Non-Autistic Individuals

In the 1940s, Heider and Simmel (1944) found that most people viewing a short movie of moving geometric shapes attributed human characteristics to the shapes and movements (e.g., “…the two men have a fight… the girl starts to go into the room to get out of the way…She apparently does not want to be with the first man”; Heider & Simmel, 1944, page 247). Decades later, using the same videos, Ami (2000) found that autistic individuals identified many fewer social elements in the story, made attributions that were irrelevant to the social plot, and did not afford the shapes Theory of Mind skills to the same degree as non-autistic individuals. In our study, we use computer models of language to evaluate brain imaging data gathered while autistic and non-autistic adolescents watched a Heider-and-Simmel-like movie. In a 1.5T scanner, 60 adolescents (n=30 non-autistic) passively viewed a 7 minute movie of moving geometric shapes. The movie was accompanied by music, but had no dialogue. It depicted a story about a child (i.e., a small triangle) that included interactions such as chatting with their parent (i.e., a big triangle), having nightmares of a monster, and finding a friend. We created a design matrix representing anthropomorphized nouns and actions seen in each second of the movie, the hypernyms of these words, pre-processing motion parameters of the neuroimaging data, and time delays to account for the shape of the hemodynamic response. For each participant, we then ran an encoding analysis to see which voxels responded most consistently to the information in the design matrix (i.e., Fisher p-value<0.05 and correlation coefficient r >0). Following that, on a group-by-group basis, we concatenated the beta weights of the “responsive” voxels from all participants in that group and ran a principal components analysis, thus carrying out one PCA for each group. Finally, for each of the first four principal components in each group, we determined how much of the variance in the given component was related to each of 13 different social and non-social dimensions. For Group1, the top four dimensions were: ‘Change’, ‘Internal verbs/Strong emotions’, ‘Mobile’, and ‘Size’; lowest was ‘Human’. Top four for Group2 were: ‘Human’, ‘Animacy’, ‘Size’, and ‘Mobile’; lowest was ‘Verbs-2-people’. For Group1, the most salience involved whether things were changing state and when theory of mind verbs were being observed (e.g., “trick”, “plan”) or strong feelings were being portrayed (e.g., “anxious”, “sad”). Whether or not something was identified as “human” was minimally salient for this group. In contrast, for Group 2, the most salience involved whether or not something was identified as human, and whether or not it was animate; whereas there was minimal salience reflected for verbs that require 2 people (e.g., “argue”, “hug”). We have not yet finished analysis, and thus are keeping ourselves blind to group identity. However, we feel that the results are strongly suggestive of group identity and look forward to discussing it at the conference.

Friday, May 26, 2023

10:25 PDT
11:25 MDT
13:25 EDT
14:25 ADT
18:25 BST

Brea Chouinard
Postdoctoral Researcher
University of Alberta