A Neurodevelopmental Shift in Reward Circuitry from Mother's to Nonfamilial Voices in Adolescence

Stimuli

Stimuli consisted of the three nonsense words, “teebudieshawlt,” “keebudieshawlt,” and “peebudieshawlt,” produced by the participant's mother as well as two NF females who were also mothers (Fig. 1a; for audio examples, see Extended Data Figs. 1-1, 1-2, 1-3, 1-4, 1-5, and 1-6). Nonsense words were used to avoid activating semantic neural systems, thereby enabling a focus on the neural responses to each speaker's vocal characteristics (Binder et al., 2000; Raettig and Kotz, 2008). These particular nonsense words were selected for a number of reasons: first, stimuli are exemplars from a standardized behavioral test of phonological abilities (Wagner et al., 1999); second, the use of stimulus contrasts that differ by only one phoneme (i.e., minimal pair) enables a fine-grained assessment of phonological decoding in the auditory system; finally, the use of four syllable nonsense words, differentiated by word-initial, place-of-articulation contrasts, provides challenging but realistic speech-like stimuli. These vocal stimuli were used in both the fMRI task and the Voice Identification behavioral task. A second class of stimuli included in the fMRI task was nonspeech environmental sounds. These sounds, which included brief recordings of laundry machines, dishwashers, and other household sounds, were taken from a professional sound effects library.

Stimulus recording

Recordings of each mother were made individually for use in the voice identification and fMRI tasks. Mother's voice stimuli and NF voices were recorded in a quiet conference room using a Shure PG27-USB condenser microphone connected to a MacBook Air laptop. The audio signal was digitized at a sampling rate of 44.1 kHz and A/D converted with 16-bit resolution. Mothers were positioned in the conference room to avoid early sound wave reflections from contaminating the recordings. To provide a natural speech context for the recording of each nonsense word, mothers were instructed to repeat three sentences, each of which contained one of the nonsense words, during the recording. The first word of each of these sentences was their child's name, which was followed by the words “that is a,” followed by one of the three nonsense words. A hypothetical example of a sentence spoken by a mother for the recording was “Johnny, that is a keebudieshawlt.” Before beginning the recording, mothers were instructed on how to produce these nonsense words by repeating them to the experimenter until the mothers had reached proficiency. Importantly, mothers were instructed to say these sentences using the tone of voice they would use when speaking with their child during an engaging and enjoyable shared learning experience (e.g., if their child asked them to identify an item at a museum). The vocal recording session resulted in digitized recordings of the mothers repeating each of the three sentences ∼30 times to ensure multiple high-quality samples of each nonsense word for each mother.

Stimulus postprocessing

The goal of stimulus postprocessing was to isolate the three nonsense words from the sentences that each mother and NF spoke during the recording session and normalize them for duration and RMS amplitude for inclusion in the fMRI stimulus presentation protocol and the voice identification task. First, a digital sound editor (Audacity: http://audacity.sourceforge.net) was used to isolate each utterance of the three nonsense words from the sentences spoken by each mother. The three best versions of each nonsense word were then selected based on the audio and vocal quality of the utterances (i.e., eliminating versions that were mispronounced, included vocal creak, or were otherwise not ideal exemplars of the nonsense words). These nine nonsense words were then normalized for duration to 956 ms, the mean duration of the nonsense words produced by the NF voices, using Praat software similar to previous studies (Abrams et al., 2008). On average, speech samples were adjusted 8.7% during normalization, and this process did not affect the naturalness of the vocal stimuli. A 10 ms linear fade (ramp and damp) was then performed on each stimulus to prevent click-like sounds at the beginning and end of the stimulus, and then stimuli were equated for RMS amplitude. These final stimuli were then evaluated for audibility and clarity to ensure that postprocessing manipulations had not introduced any artifacts into the samples. The same process was performed on the NF voices and environmental sounds to ensure that all stimuli presented in the fMRI experiment were the same duration and RMS amplitude.

Voice identification behavioral task

All participants who participated in the fMRI experiment completed an auditory behavioral test immediately following the voice processing fMRI scan. The goal of the voice identification behavioral task was to ensure that participants were able to reliably discriminate their mother's voice from NF female voices. Participants were seated in a quiet area of the brain imaging suite in front of a laptop computer and facing a wall with noise-cancellation headphones placed over their ears to prevent distractions. Vocal stimuli were delivered via Eprime on the laptop computer. In each trial, participants were presented with a recording of a multisyllabic nonsense word spoken by either the participant's mother or an unfamiliar, NF voice, and the task was to indicate whether their mother spoke the word. The multisyllabic nonsense words used in the behavioral task were the exact same samples used in the fMRI task. Specifically, the stimuli that were presented in the voice identification task consisted of the three nonsense words produced 3 times each by each child's mother as well as the two NF voices (i.e., all the vocal stimuli identified in Fig. 1a). This yielded a total of 27 stimuli (3 nonsense words × 3 repetitions × 3 different speakers). Each stimulus was presented twice in a random order during the voice identification task for a total of 54 trials, including 18 trials in which mother's voice nonsense words were presented and 18 trials in which each NF voice nonsense word was presented. Participants were instructed to press a button on the laptop keyboard as soon as they knew the answer.

fMRI data acquisition parameters

All fMRI data were acquired in a single session at the Richard M. Lucas Center for Imaging at Stanford University. Functional images were acquired on a 3-T Signa scanner (General Electric) using a custom-built head coil. Participants were instructed to stay as still as possible during scanning, and head movement was further minimized by placing memory-foam pillows around the participant's head. A total of 31 axial slices (4.0 mm thickness, 0.5 mm skip) parallel to the anterior/posterior commissure line and covering the whole brain were imaged by using a T2*-weighted gradient-echo spiral in-out pulse sequence (Glover and Law, 2001) with the following parameters: TR, 3576 ms; TE, 30 ms; flip angle, 80°; one interleave. This TR can be calculated as the sum of the stimulus duration (956 ms), a 300 ms silent interval buffering the beginning and end of each stimulus presentation to avoid backward and forward masking effects, the 2000 ms volume acquisition time, and an additional 22 ms silent interval which helped the stimulus computer maintain precise and accurate timing during stimulus presentation. The FOV was 22 cm, and the matrix size was 64 × 64, providing an in-plane spatial resolution of 3.4375 mm. Reduction of blurring and signal loss arising from field inhomogeneities was accomplished using an automated high-order shimming method before data acquisition.

fMRI task

Auditory stimuli were presented in 10 separate runs, each lasting 4 min. One run consisted of 56 randomized trials of mother's voice and NF voices producing the three nonsense words, environmental sounds, and catch trials. Randomizing the order of presentation for all stimuli, including both mother's and NFs vocal samples, during fMRI data collection creates the same level of (in)dependence between iterations of the nonsense words for both vocal sources included in the data analysis. Each stimulus lasted 956 ms in duration. Before each run, participants were instructed to play the “kitty cat” game during the fMRI scan. While laying down in the scanner, participants were first shown a brief video of a cat and were told that the goal of the cat game was to listen to a variety of sounds, including “voices that may be familiar,” and to push a button on a button box only when they heard cat meows (catch trials). During each run, four or five exemplars of each stimulus type, including three speakers producing three nonsense words, multiple environmental sounds, and three catch trials, were presented. Silent trials were not included in the fMRI task. At the end of each run, participants were shown another engaging video of a cat. Across the 10 runs, a total of 48 exemplars of each stimulus condition were presented to each subject. Auditory stimuli were presented to participants in the scanner using Eprime version 1.0 (Psychological Software Tools, 2002). Participants wore custom-built headphones designed to reduce the background scanner noise to ∼70 dBA (Abrams et al., 2011, 2013a). Headphone sound levels were calibrated before each data collection session, and all stimuli were presented at a sound level of 75 dBA. Participants were scanned using an event-related design. Auditory stimuli were presented during silent intervals between volume acquisitions to eliminate the effects of scanner noise on auditory discrimination. One stimulus was presented every 3.576 s.

fMRI preprocessing

fMRI data collected in each of the 10 functional runs were subject to the following preprocessing procedures. The first five volumes were not analyzed to allow for signal equilibration. A linear shim correction was applied separately for each slice during reconstruction by using a magnetic field map acquired automatically by the pulse sequence at the beginning of the scan. Translational movement in millimeters (x, y, z) was calculated based on the SPM12 parameters for motion correction of the functional images in each subject. To correct for deviant volumes resulting from spikes in movement, we used a de-spiking procedure (Iuculano et al., 2014) similar to those implemented in the Analysis of Functional Neuroimages toolkit maintained by the National Institute of Mental Health (Cox, 1996). Volumes with movement exceeding 0.5 voxels (1.562 mm) or spikes in global signal exceeding 5% were interpolated using adjacent scans. The majority of volumes repaired occurred in isolation. After the interpolation procedure, images were spatially normalized to standard MNI space, resampled to 2 mm isotropic voxels, and smoothed with a 6 mm FWHM Gaussian kernel.

Movement criteria for inclusion in fMRI analysis

For inclusion in the fMRI analysis, we required that each functional run had a maximum scan-to-scan movement of <6 mm and no more than 15% of volumes were corrected in the de-spiking procedure. Moreover, we required that all individual subject data included in the analysis consisted of at least seven functional runs that met our criteria for scan-to-scan movement and percentage of volumes corrected; subjects who had fewer than seven functional runs that met our movement criteria were not included in the data analysis. All 46 participants included in the analysis had at least 7 functional runs that met our movement criteria; 32 of the participants had 10 runs of data that met these movement criteria; 6 subjects had 9 runs of data that met movement criteria, 6 subjects had 8 runs of data, and 2 subjects had 7 runs that met criteria.

Voxel-wise analysis of fMRI activation

The goal of this analysis was to identify brain regions that showed differential activity levels in response to mother's voice, NF voices, and environmental sounds. Brain activation related to each task condition was first modeled at the individual subject level using boxcar functions with a canonical HRF and a temporal derivative to account for voxel-wise latency differences in hemodynamic response. Low-frequency drifts at each voxel were removed using a high-pass filter (0.5 cycles/min), and serial correlations were accounted for by modeling the fMRI time series as a first-degree autoregressive process (Friston et al., 1997). Voxel-wise t-statistics maps for each condition were generated for each participant using the GLM, along with the respective contrast images. Group-level activation was determined using individual subject contrast images and a second-level ANOVA. The contrasts of interest were as follows: (1) [NF voices – environmental sounds]; (2) [mother's voice – environmental sounds]; and (3) [NF voices – mother's voice]. To ensure that the inclusion of two NF voices compared with the (one) mother's voice stimulus did not bias fMRI contrast betas and T-maps, a value of 1 was entered into the contrast matrix for mother's voice, while a value of 0.5 was entered into the contrast matrix for each of the two NF voices. A fourth contrast of interest [NF voice #1 – NF voice #2] served as a control analysis to examine whether brain activation differences in response to the two NF voices used in the study vary as a function of age.

For all voxel-wise analyses, including the age covariate analysis (Figs. 2-4) and main effects of all stimulus contrasts (Fig. 5), significant clusters of activation were determined using a voxel-wise statistical height threshold of p < 0.005, with family-wise error corrections for multiple spatial comparisons (p < 0.05; 67 voxels) determined using Monte Carlo simulations using a custom MATLAB script. Significant correlations are inherent to all scatterplots included in the age covariate analysis (Figs. 2-4) because they are based on results from the whole-brain GLM analysis (Vul et al., 2009); however, the results provide important information regarding the distributions and covariation of activity strength in response to voices and age. For age covariate analyses, effect sizes were computed as Cohen's f according to Equation 1, where t is the mean t-score within a cluster and N is the sample size, as follows: Cohen's d=tsqrt(N2)(1)

To examine GLM results in the NAc, a small subcortical brain structure, we used a small volume correction at p < 0.005 using the Harvard-Oxford probabilistic maps of the NAc thresholded at 25%.

Acoustical analysis

The goal of the acoustical analysis was to examine whether physical attributes of the mother's voice stimuli may have contributed to age-related increases in brain activity for the NF voice versus mother's voice contrast (Fig. 4). Acoustical features of each mother's voice stimulus were extracted, including mean pitch, pitch SD, pitch slope, spectral center of gravity, and spectral SD, using Praat software similar to previous studies (Abrams et al., 2016, 2019). These acoustical values were then averaged across pseudoword stimuli for each mother. We examined the relation between mean acoustical values for each mother and signal levels measured in ROIs identified in Figure 4, including left-hemisphere NAc and right-hemisphere vmPFC, for the [NF – mother's voice] contrast. GLM betas were extracted from left-hemisphere NAc and right-hemisphere vmPFC. Coordinates for these ROIs were based on the voxel with the peak T-score in that region for each of these whole-brain age covariate maps. The cortical ROI (i.e., vmPFC) was defined as a 4 mm sphere, and the subcortical ROI (i.e., NAc) was a 2 mm sphere, and signal level was calculated by extracting the β value from individual subjects' contrast maps within each ROI and computing the mean β value for each ROI. Results were FDR-corrected for multiple comparisons.

Sex difference analysis

A regression analysis was performed to examine the effect of sex on age-related changes in ROI signal level measured in response to the [NF voices > mother's voice] GLM contrast. GLM betas were extracted from left-hemisphere NAc and right-hemisphere vmPFC as described previously (see Acoustical analysis). We then built a regression model with age as the dependent variable and sex, mean β value for each ROI, and their interaction as predictors in the regression equation. Significant results for the [sex × mean β value] interaction are reported (p < 0.05).

Support vector regression (SVR) analysis: brain activity levels and age prediction

The robustness and replicability of fMRI data remain a crucial concern for the field, and an established approach for addressing this issue is to perform a confirmatory cross-validation (CV) analysis (Cohen et al., 2010). We therefore used SVR to perform a confirmatory CV analysis that uses a machine-learning approach with balanced fourfold CV combined with linear regression (Cohen et al., 2010). In this analysis, we extracted individual subject activation β values taken from the [NF voices > environmental sounds], [mother's voice > environmental sounds], and [NF voices > mother's voice] GLM contrasts as described above. Mean β values for each ROI were entered as independent variables in a linear regression analysis with age as the dependent variable. r_{(predicted, observed)}, a measure of how well the independent variable predicts the dependent variable, was first estimated using a balanced fourfold CV procedure. Data were divided into four folds so that the distributions of dependent and independent variables were balanced across folds. Data were randomly assigned to four folds and the independent and dependent variables tested in one-way ANOVAs, repeating as necessary until both ANOVAs were insignificant to guarantee balance across the folds. A linear regression model was built using three folds leaving out the fourth, and this model was then used to predict the data in the left-out fold. This procedure was repeated 4 times to compute a final r_{(predicted, observed)} representing the correlation between the data predicted by the regression model and the observed data. Finally, the statistical significance of the model was assessed using a nonparametric testing approach. The empirical null distribution of r_{(predicted, observed)} was estimated by generating 1000 surrogate datasets under the null hypothesis that there was no association between changes in age and brain activity levels.