Hierarchical Clustering Analysis
Grand-averaged raw and normalized GFP across both groups are depicted in Figure 1, visual inspection of which shows that intensity systematically modulated neural responses, as well as supplementary Figures S3-S6, while the results of the hierarchical clustering analysis are displayed in Figure 2. Groups are referred to as “C1” (referring to “Cluster 1”) through “C4”. Generally, analyses of cluster replicability (albeit within the present dataset) based on drawing and clustering repeated subsamples suggest the present clusters are fairly replicable, with participants being substantially more likely to be clustered alongside other participants from their original clusters than alongside those outside their original clusters (Supplementary Tables S2-S3), although it should be noted that one subgroup of participants within C1 showed some propensity to move back and forth between C1 and C2 on resampling (Supplementary Figures S1-S2).
Diagnostic Group Membership.
The distributions of autistic and typically-developing participants across the clusters differed significantly, X2 (3, N = 213) = 8.42, p = .04 (Table 3). C1 had significantly more autistic participants than expected based on proportion of ASD participants in the study, ASR = 2.70, corrected p = .03. Furthermore, at a trend level, C3 had more typically-developing participants than expected, ASR = 2.18, corrected p = .06.
Table 3. Counts and percentages of autistic and typically-developing participants, separately, in cluster groups.
|
|
C1
|
C2
|
C3
|
C4
|
ASD
|
53 (74.65%)
|
24 (58.54%)
|
32 (50.79%)
|
23 (60.53%)
|
TD
|
18 (25.35%)
|
17 (41.46%)
|
31 (49.21%)
|
15 (39.47%)
|
Cluster Electrophysiological Patterns.
There was a significant between-subjects main effect of cluster on normalized GFP, F(3, 209) = 3.86, p = .01, = .003, there was a significant within-subjects effect of loudness, F(3, 627) = 78.23, p < .0001, = .26, and there was a significant cluster by loudness interaction, F(9, 627) = 55.09, p < .0001, = .43 (Figures 3, 4). This robust interaction confirmed that the hierarchical clustering analysis succeeded in defining clusters that differed in the loudness-dependency of their electrophysiological responses to auditory stimuli.
In the 50 dB condition, normalized GFP significantly differed across the clusters, F(3, 209) = 49.09, corrected p < .0001, = .41. The brain’s electrophysiological response to 50 dB sounds was significantly stronger in C4 than in any other cluster group (compared to C1, Welch’s t(96.90) = 13.91, corrected p < .0001, d = 2.55; compared to C2, Welch’s t(63.42) = -11.36, corrected p < .0001, d = 2.50; and compared to C3, Welch’s t(98.99) = 10.58, corrected p < .0001, d = 1.93). There was also a trend for a stronger 50 dB response in C3 compared to C2, but this was not significant after applying the Holm-Bonferroni correction, Welch’s t(81.40) = 2.45, corrected p = .11, d = 0.50.
In the 60 dB condition, normalized GFP significantly differed across the clusters, F(3, 209) = 85.54, corrected p < .0001, = .55. The brain’s electrophysiological response to 60 dB sounds was significantly stronger in C3 than any other group (compared to C1, Welch’s t(131.85) = 10.04, corrected p < .0001, d = 1.72; compared to C2, Welch’s t(76.02) = 14.30, corrected p < .0001, d = 2.97; and compared to C4, Welch’s t(83.34) = 12.91, corrected p < .0001, d = 2.60). Furthermore, the brain’s response to 60 dB sounds was stronger in C1 than C2, Welch’s t(83.44) = 5.64, corrected p < .0001, d = 1.11, and it was stronger in C1 than C4, Welch’s t(92.12) = 3.15, corrected p = .02, d = 0.59. A trend for the 60 dB response to be stronger in C4 than C2 did not survive post-hoc correction, Welch’s t(75.05) = 2.77, corrected p = .06, d = 0.62.
In the 70 dB condition, normalized GFP significantly differed across the clusters, F(3, 209) = 45.57, corrected p < .0001, = .40. The brain’s electrophysiological response to 70 dB sounds was significantly stronger in C1 than any other group (compared to C2, Welch’s t(88.25) = 5.48, corrected p < .0001, d = 1.06; compared to C3, Welch’s t(124.08) = 11.90, corrected p < .0001, d = 2.01; and compared to C4, Welch’s t(82.60) = 6.62, corrected p < .0001, d = 1.29). Furthermore, the brain’s response to 70 dB sounds was significantly weaker in C3 than in C2, Welch’s t(67.23) = –4.20, corrected p = .0009, d = –0.90, and the 70 dB response was significantly weaker in C3 than C4, Welch’s t(62.34) = –2.91, corrected p = .05, d = –0.64.
In the 80 dB condition, normalized GFP significantly differed across the clusters, F(3, 209) = 42.00, corrected p < .0001, = .38. The brain’s electrophysiological response to 80 dB sounds was significantly stronger in C2 than any other group (compared to C1, Welch’s t(76.23) = 9.74, corrected p < .0001, d = 1.97; compared to C3, Welch’s t(87.61) = 9.13, corrected p < .0001, d = 1.82; and compared to C4, Welch’s t(74.44) = 9.13, corrected p < .0001, d = 2.06). These effects were robust to the removal of the three outlying participants in C2 (based on criterion of 3x the median absolute deviation) visible in Figure 3.
Measures and Demographics.
Caregiver-Reported Sensory Symptoms.
Total scores on the SSP showed a non-normal distribution, Shapiro-Wilk W = .96, p = .0003, as did scores on the three SSP auditory factors defined by Williams et al. [77], Shapiro-Wilk p ≤ .0001. Furthermore, combining across clusters, autistic participants had significantly lower (i.e., more atypical) SSP total scores than typically-developing participants, Wilcoxon-Mann-Whitney W = 746.5, nASD = 99, nTD = 65, p < .0001, δ = –.88, and the same pattern was observed with all three auditory subscores, p ≤ .008. Therefore, one-way non-parametric Kruskal-Wallis tests were used to compare clusters on SSP scores and subscores separately in each diagnostic group. Among autistic participants, total SSP scores differed between clusters, H(3) = 8.67, p = .03 (Table 4, Figure 5A). Wilcoxon-Mann-Whitney tests indicated that sensory processing trended towards being more atypical in C2 than C3, W = 108.5, nC2 = 18, nC3 = 23, corrected p = .06, δ = –.48. Furthermore, among autistic participants, scores on the SSP Auditory Distractibility factor significantly differed between clusters, H(3) = 10.31, p = .02 (Figure 5B). Wilcoxon-Mann-Whitney tests indicated that more auditory distractibility was reported in C2 than C4, W = 90.5, nC2 = 19, nC4 = 19, Bonferroni-Holm corrected p = .05, δ = –.50; there was also a trend for more auditory distractibility in C2 than C3, W = 139.5, nC2 = 19, nC3 = 26, corrected p = .07, δ = –.44. Trends in the ASD sample for effects of cluster on Taste/Smell Sensitivity and Noise Distress did not attain significance. In the typically-developing sample, one-way Kruskal-Wallis tests revealed no significant differences between clusters on the SSP total score or any of the auditory subscores (Supplementary Table S1).
INSERT TABLE 4 ABOUT HERE
Chronological Age.
There was no main effect of cluster group, F(3, 204) = 1.02, p = .38, = .01, or diagnostic group, F(1, 204) = 2.19, p = .14, = .01 on chronological age, nor was there an interaction of cluster and diagnostic groups, F(3, 204) = 0.90, p = .44, = .01.
Cognitive Ability.
MSEL DQ had a non-normal distribution in ASD, Shapiro-Wilk W = .96, p = .0008. Furthermore, autistic participants had significantly lower DQ than typically-developing participants (Table 1). Therefore, one-way non-parametric Kruskal-Wallis tests were used to compare clusters on DQ, VDQ, and NVDQ scores separately in each diagnostic group. Among autistic participants, DQ scores differed between clusters, H(3) = 8.28, p = .04 (Table 5, Figure 5C). Similar effects were found in NVDQ scores, and a similar trend was seen in VDQ scores, although in each case post-hoc comparisons failed to locate significant differences after strict corrections for multiple comparisons were applied. Among typically-developing participants, VDQ scores significantly differed between clusters (Table 6, Figure 5D), H(3) = 9.08, p = .03; after correction, scores in C2 were lower than C4, W = 56.0, nC2 = 17, nC3 = 15, corrected p = .04, δ = –.56. No effects of full-scale DQ and NVDQ were observed in TD.
INSERT TABLES 5 AND 6 ABOUT HERE
Adaptive Behaviour.
Although the overall distribution of VABS composite scores was non-normal, distributions did not violate Shapiro-Wilk tests in each group separately (p ≥ .07). Therefore, one-way ANOVAs were used to compare clusters on VABS composite scores separately in each diagnostic group. VABS scores did not differ between clusters in the ASD sample, F(3, 113) = 2.04, p = .11. Among typically-developing participants, the VABS composite did not differ between clusters, F(3, 65) = 1.48, p = .23.
Anxiety.
CBCL DSM-oriented anxiety T-scores had a non-normal distribution, Shapiro-Wilk W = .63, p < .0001. Furthermore, autistic participants had significantly higher anxiety levels than typically-developing participants, Wilcoxon-Mann-Whitney W = 6043.5, nASD = 126, nTD = 75, p = .0003, δ = .28. Therefore, one-way non-parametric Kruskal-Wallis tests were used to compare clusters on anxiety scores separately in each diagnostic group. Anxiety levels did not differ between clusters among autistic participants, H(3) = 0.25, p = .97, or typically-developing participants, H(3) = 0.97, p = .81.
Cluster-Based Permutation Correlation Analyses
Caregiver-Reported Sensory Symptoms.
In ASD, there was a significant negative Spearman correlation between SSP total scores and normalized GFP to 70 dB sounds in a contiguous series of time points spanning between 97 and 131 ms, p = .009: that is, autistic participants with strong responses to 70 dB sounds in this approximate time period had, overall, more atypical caregiver-reported sensory processing features (Figure 6A). There were no significant associations between SSP total scores and normalized GFP in any other loudness condition in ASD.
In ASD, there was a significant positive correlation between SSP auditory distractibility scores and normalized GFP to 50 dB sounds in a contiguous series of points spanning between 87 and 128 ms, p = .01 (Figure 6B). There was no significant association between SSP auditory distractibility scores and normalized GFP to 60 dB sounds in ASD. However, there was a significant negative association between SSP auditory distractibility scores and normalized GFP to 70 dB sounds between 101 and 132 ms, p = .02. Furthermore, in ASD, there was a significant negative association between SSP auditory distractibility scores and normalized GFP to 80 dB sounds between 79 and 114 ms, p = .02. In other words, autistic participants with relatively weak responses to soft 50 dB sounds and relatively strong responses to louder 70 dB and 80 dB sounds were reported by caregivers to have more auditory distractibility problems.
Cognitive Ability.
In ASD, there were no significant associations between normalized GFP and MSEL DQ or MSEL NVDQ in any loudness condition or diagnostic group. Furthermore, in TD, there were no significant associations between normalized GFP and MSEL VDQ in any loudness condition or diagnostic group.