Automatic and feature-specific prediction-related neural activity in the human auditory system

Abstract

Prior experience enables the formation of expectations of upcoming sensory events. However, in the auditory modality, it is not known whether prediction-related neural signals carry feature-specific information. Here, using magnetoencephalography (MEG), we examined whether predictions of future auditory stimuli carry tonotopic specific information. Participants passively listened to sound sequences of four carrier frequencies (tones) with a fixed presentation rate, ensuring strong temporal expectations of when the next stimulus would occur. Expectation of which frequency would occur was parametrically modulated across the sequences, and sounds were occasionally omitted. We show that increasing the regularity of the sequence boosts carrier-frequency-specific neural activity patterns during both the anticipatory and omission periods, indicating that prediction-related neural activity is indeed feature-specific. Our results illustrate that even without bottom-up input, auditory predictions can activate tonotopically specific templates.

Fig. 1

Experimental design. a Transition matrices used to generate sound sequences according to the different conditions (random (RD), midminus (MM), midplus (MP), and ordered (OR)). b Schematic examples of different sound sequences generated across time. 10% of sound stimuli were randomly replaced by omission trials (absence of sound) in each context

Fig. 2

Decoding carrier frequencies from random sound sequences. a Robust increase of decoding accuracy is obtained rapidly, peaking ~100 ms after which it slowly wanes. Note however that significant decoding accuracy is observed even after 700 ms, i.e. likely representing a memory trace that is (re-)activated even when new tones (with other carrier frequencies) are processed. b Source projection of classifier weights (relative change baseline (-100-0) ms, 50% threshold) for an early (W1) and late (W2) reveals informative activity to mainly originate from auditory cortices, with a right hemispheric dominance. During the later (W2) period informative activity spreads to also encompass e.g. frontal regions

Fig. 3

Analysis for pre- and post-stimulus decoding using time-generalization of classifier trained on random sound sequences. a, b “Raw” decoding time-generalization maps (grand average across subjects), tested on sound- (a) and omission- (b) locked trials, in increasing entropy from top to bottom. c, d Regression results (sound left, omission right), using entropy level as the independent variable; red colors indicate increased decoding accuracy for more regular sequences. t-values are thresholded at uncorrected p < 0.05. The areas framed in black are clusters significant at p_cluster < 0.05. e, f Decoding accuracy for individual conditions averaged for training times between the dashed lines, testing on sound (left) and omission (right). c Display of effects pre- and post-sound, showing a clear anticipation effect and a late effect commencing after ~400 ms. The latter effect is more clearly visualized in (e). Interestingly, different train times appear to dominate the anticipation and post-stimulus effects. d Display of effects pre- and post-omission, showing a single continuous positive cluster. However, the actual t-values suggest temporally distinct maxima within this cluster underlining the dynamics around this event. Analogous to sounds a clear anticipation effect can be observed, driven by increased pre-omission decoding accuracy for events embedded in regular sequences (see f). A similar increase can be seen immediately following the onset of the omission which cannot be observed following actual sound onset. Interestingly this increase is long lasting with further peaks emerging approximately at 330 ms and 580 ms

Fig. 4

Decoding entropy level of sound sequence and correlation main prediction effects gained from time-generalization analysis (see Fig. 3). a A classifier was trained to decode the entropy level from MEG activity elicited by sounds and tested around relevant events, i.e. sound or omission onset. Robust increases of decoding accuracy can be observed following sound onsets. Right temporal and parietal regions appear to contribute most informative activity (small inset, relative change baseline (-100-0 ms), 50% threshold). While overall decoding accuracy is above chance level throughout most of the entire period, this pattern breaks down briefly following an omission. b Average entropy level decoding following sound onset (0–330 ms) was taken and (Spearman) correlated with the time-generalized decoding accuracy of the low entropy condition. Nonparametric cluster permutation test yields a significant negative correlation especially with early training time-window patterns (W1) in the anticipation period towards a sound that was; however, not observed prior to omissions. Following the onset of omissions nonparametric cluster permutation testing pointed to a late positive correlation with the late activation patterns (W2)