A phenomenological cartography of misophonia and other forms of sound intolerance

Abstract

People with misophonia have strong aversive reactions to specific "trigger" sounds. Here we challenge this key idea of specificity. Machine learning was used to identify a misophonic profile from a multivariate sound-response pattern. Misophonia could be classified from most sounds (traditional triggers and non-triggers) and, moreover, cross-classification showed that the profile was largely transferable across sounds (rather than idiosyncratic for each sound). By splitting our participants in other ways, we were able to show-using the same approach-a differential diagnostic profile factoring in potential co-morbidities (autism, hyperacusis, ASMR). The broad autism phenotype was classified via aversions to repetitive sounds rather than the eating sounds most easily classified in misophonia. Within misophonia, the presence of hyperacusis and sound-induced pain had widespread effects across all sounds. Overall, we show that misophonia is characterized by a distinctive reaction to most sounds that ultimately becomes most noticeable for a sub-set of those sounds.

Keywords: Behavioral neuroscience; Cognitive neuroscience; Health sciences.

Graphical abstract

Figure 1

A heatmap depicting misophonic > control effect sizes (Cohen’s d) for 32 sounds and 17 response features The sounds are ranked according to the ability of a machine learning classifier to predict group membership (AUC). ∗p < 0.05 significance relative to chance classification. See also Figure S2.

Figure 2

A heatmap depicting cross-classification accuracy (AUC) when a classifer trained on responses to one sound is used to predict group membership from a pattern of responses to another sound (note that the on-diagonal values are not cross-classifications) The grid lines show divisions between the four sound categories (human oral/nasal, human actions, non-human, and scrambled). The percentages within each grid cell refer to the number of cross-classifications that achieved significance (at p < 0.05 FDR corrected).

Figure 3

Top: Spearman’s correlations between the psychoacoustic region of interest identified by Kumar et al. and phenomenological ratings to the same sounds by misophonics and controls considering scrambled and real sounds (left and right) Note, the critical values for alpha = 0.05 are 0.643 (n = 8, i.e. scrambled sounds) and 0.344 (n = 24, i.e. natural sounds). ∗p < 0.05.

Figure 4

As an example, mean ratings for “annoyance” ranked from most to least annoying (according to the misophonic mean rating), with the mean correlation between ranks being 0.912 Equivalent plots for all 17 descriptors are provided in Figures S8–S12. The right panel shows Spearman’s correlations between ratings to the same sounds across groups. The critical value for alpha = 0.05 is rho ≥ 0.296 (for N = 32 sounds).

Figure 5

The ability to classify misophonics based on their response pattern to 32 different sounds (x axis) correlates with the number of misophonic triggers reported (top figure) and the severity of the misophonia (bottom) Each point is an individual misophonic participant.

Figure 6

Heatmaps (showing Cohen’s d effect sizes) for 32 sounds (y axis) and 17 responses (x axis) for within misophonia comparisons, contrasting presence versus absence of hyperacusis (top) and high versus low scores on the SMS pain factor (bottom)

Figure 7

Heatmaps (showing Cohen’s d effect sizes) for 32 sounds (y axis) and 17 responses (x axis) for comparisons within non-misophonics, contrasting high versus low AQ scorers

Figure 8

Heatmaps (showing Cohen’s d effect sizes) for (top) 32 sounds (y axis) and 17 responses (x axis) and (bottom) the eight ASMR triggers, contrasting ASMR responders and non-responders (excluding misophonics)