How musicality enhances top-down and bottom-up selective attention: Insights from precise separation of simultaneous neural responses

Abstract

Natural environments typically contain a blend of simultaneous sounds. A substantial challenge in neuroscience is identifying specific neural signals corresponding to each sound and analyzing them separately. Combining frequency tagging and machine learning, we achieved high-precision separation of neural responses to mixed melodies, classifying them by selective attention toward specific melodies. Across two magnetoencephalography datasets, individual musicality and task performance heavily influenced the attentional recruitment of cortical regions, correlating positively with top-down attention in the left parietal cortex but negatively with bottom-up attention in the right. In prefrontal areas, neural responses indicating higher sustained selective attention reflected better performance and musicality. These results suggest that musical training enhances neural mechanisms in the frontoparietal regions, boosting performance via improving top-down attention, reducing bottom-up distractions, and maintaining selective attention over time. This work establishes the effectiveness of combining frequency tagging with machine learning to capture cognitive and behavioral effects with stimulus precision, applicable to other studies involving simultaneous stimuli.

Fig. 1.. Experimental tasks and behavioral results.

(A) Melody attention task. Participants attended selectively to one of two simultaneous melodies with different pitches. The low-pitched and high-pitched melodies were frequency-tagged at 39 and 43 Hz, respectively. When the melody stopped, participants reported the final direction of pitch change for the attended melody, which was either falling, rising, or constant. Identical sounds were presented to both ears, ensuring that the melodies could only be distinguished by pitch or timing. (B) Two variations of the experimental task. Experiment I (left): Alternately overlapping melodies. Tone onsets alternate between the low-pitched and high-pitched melodies. Each tone onset induces bottom-up attention toward it, allowing the dissociation of top-down and bottom-up attention effects toward each melody. Experiment II (right): Completely overlapping melodies. Melody tones overlap completely, engaging bottom-up attention simultaneously toward both melodies. All other parameters including modulation frequencies (f_m) and carrier frequencies (f_c) were identical to experiment I. (C) Correlations between participant task performance and musicality. Task performance correlated positively (Pearson correlation, P < 0.001) with musicality across participants in present (Expt I and II) and past experiments (Expt 0). Musicality was measured by the Goldsmiths MSI (25, 26). Notably, the strength of correlation (Pearson r) between musicality and performance increased with task complexity. The least complex task (Expt 0, r = 0.64, N = 28) involved selectively attending to melodies that were completely separated in time and pitch. The moderately complex task (Expt I, r = 0.71, N = 28) required attending to melodies that were partly separated in time and pitch. The most complex task (Expt II, r = 0.79, N = 20) involved selectively attending to melodies that completely overlapped in time and could only be separated by pitch. Expt, experiment.

Fig. 2.. Repeated splitting classification of selective attention at frequency tags.

We trained a classifier on frequency-tagged neural activities to discriminate between conditions where attention was directed toward the 39-Hz low-pitched or 43-Hz high-pitched melody. (A) Single-subject repeated splitting support vector machine pipeline for ASSR classification. For each condition, epochs were randomly divided into five groups and averaged to produce five evoked ASSRs. Next, the evoked ASSRs were Fourier transformed to acquire five power spectra. For any contrast between two conditions, the corresponding power spectra (five per condition) were classified via cross-validation (chance level = 0.5), obtaining the AUC which was expected to increase with higher selective attention. Subsequently, we repeated the entire process from the initial group division step 1000 times and computed the mean AUC across repetitions. The AUC was computed independently for each frequency from 4 to 45 Hz. For source analysis, the evoked ASSRs were localized to the cortical surface before Fourier transformation (step iii). Below the power spectra, MEG gradiometer topographies of a single fold input to the classifier demonstrated dominant auditory cortical power precisely tagged to 39 and 43 Hz but not at the adjacent 41 Hz. For visualization, the power at each frequency was normalized to the mean across all frequencies, and the subject grand average for one condition in experiment I is shown. Units are arbitrary (a.u.). (B and C) Peak AUC values were specifically observed at the stimulus frequency tags of 39 and 43 Hz (red vertical lines) but not at other frequencies. For each plot, the mean across participants is shown in black, with the AUC peak values at 39 and 43 Hz displayed in gray boxes above. Shading indicates SEM. (B) Experiment I (N = 25), top-down (left) and bottom-up attention (right). (C) Experiment II (N = 20).

Fig. 3.. Top-down and bottom-up selective attention across cortical regions.

Across participants (individual markers; N = 27), bottom-up attention was significantly discriminated above chance (dotted line) only at the right STG, while top-down attention was discriminated above chance at all regions except the right IPL. Significance levels above chance are marked by black asterisks beside the corresponding attention condition on the x axis. The y axis denotes the AUC and is identical for all subplots. Box plots outline the 25th to 75th percentiles of the data, with the center dot indicating the mean. At the left and right orbital gyri (OrG), top-down attention was discriminated significantly better than bottom-up attention (blue brackets and asterisks). These results support the notion that bottom-up attention is triggered by lower-level automatic sensory mechanisms situated predominantly in the primary cortices such as the STG, while top-down attention recruits higher-level executive mechanisms located in the prefrontal cortex. All p-values are computed with permutation tests and FDR-corrected. ***P < 0.001 and **P < 0.01. Neural activity illustrating cortical power within each ROI is displayed over the standard brain at the bottom of each subplot. For visualization, the subject grand average activity at 39 Hz was normalized to the mean power across all frequencies for a single condition. Units are arbitrary (a.u.).

Fig. 4.. Correlations of selective attention with musicality and performance.

Scatterplots showing correlations between classifier AUC values against individual musicality or task performance. The AUC is plotted in the y axis for all subfigures and reflects the degree of selective attention. Neural activity illustrating cortical power within each ROI is displayed over the standard brain at the bottom of each subplot. For visualization, the subject grand average activity at 39 Hz was normalized to the mean power across all frequencies for a single condition. Units are arbitrary (a.u.). (A) In experiment I, top-down attention at the left IPL shows positive correlations with both musicality (left; Pearson correlation, r = 0.49, P < 0.01) and performance (right; Pearson correlation, r = 0.38, P < 0.05) across 27 participants. (B) Across the same participants, bottom-up attention correlated negatively with performance at the right IPL (Pearson correlation, r = −0.47, P < 0.05). (C) Similarly, in experiment II, bottom-up attention correlated negatively with performance at the right IPL and right OrG across 19 participants (Pearson correlation, r = −0.47, P < 0.05 for both regions).

Fig. 5.. Musicality and performance across early and late attendees.

(A) Histogram of the time of peak selective attention across participants (N = 20). We computed the classifier AUC over the 2-s tone duration using a sliding time window and extracted the time of peak selective attention, corresponding to maximum AUC, for each participant in experiment II. The distribution appeared to be bimodal, dividing participants equally into “early” (blue) or “late” (orange) attendees depending on whether their attention peaked before or after 0.5 s, respectively. Colored lines represent normal distributions fitted to each group of attendees. (B) Musicality (left) and performance (right) comparison between early versus late attendees. Late attendees performed significantly better (∆ = 0.19, P < 0.05, n = 10,000, two-tailed permutation test) at the task than early attendees. Moreover, late attendees tended to be more musical than early attendees, although the difference nearly missed significance (∆ = 34.3, P = 0.057, n = 10,000, two-tailed permutation test). Box plots represent the 25th to 75th percentiles of the data, with the center dot indicating the mean. (C) Correlational analysis of lateness index with musicality (left) and performance (right). For each participant, we computed a lateness index that reflects the relative strength of selective attention in the late half of the tone compared to the early half. The lateness index correlated positively with both musicality and task performance at the right OrG (colored green over the adjacent standard brain) (Pearson correlation, P < 0.01 for musicality and performance). Together, these results suggest that musical training sharpens the neural mechanisms for sustaining or improving auditory selective attention over time, particularly in the right prefrontal cortex, thereby boosting task performance.