Distinct beta frequencies reflect categorical decisions

The study examines how neural oscillations, specifically beta-band activity (15–35 Hz), support flexible, context-dependent categorization. Categorization can vary with context, and identical stimuli may fall into different categories depending on a shifting boundary. The authors hypothesized that beta oscillations mediate the activation and reactivation of cortical representations by forming and coordinating neural ensembles, providing distinct frequency channels for different categorical decisions. Using a task in which Rhesus monkeys categorized durations (temporal) or distances (spatial) as “short” or “long” with block-wise shifting category boundaries, they probed whether distinct beta frequencies reflect relative, context-defined categories rather than absolute stimulus properties.

Neural oscillations are implicated in cognition broadly and in categorization specifically. Beta rhythms have been linked to top-down processing, maintenance of current brain states, and content-specific synchronization during endogenous processing. Prior work shows beta as transient, short-lived bursts supporting flexible network dynamics and carrying task-relevant information. Prefrontal cortex is central to abstract categorization, and preSMA participates in time processing and category boundary encoding. Theoretical accounts propose that oscillations route information between regions (communication-through-coherence) and that distinct frequencies can multiplex separate information streams (frequency-division multiplexing). Biophysical models suggest cortical beta bursts arise from synchronous excitatory inputs to pyramidal neuron dendrites, with burst duration inversely related to frequency, potentially influenced by thalamic inputs.

Participants: Two male Rhesus monkeys (Macaca mulatta), monkey 1 (5.5 kg) and monkey 2 (7.2 kg). Ethics approvals were obtained. Task: Monkeys performed temporal (interval) and spatial (distance) categorization. In each trial, two parallel bars appeared briefly twice, separated by a delay. In temporal tasks (T1–T3), the delay varied while bar distance was constant; in spatial tasks (S1–S3), the bar distance varied while the delay was constant (670 ms). After stimulus presentation, a fixed decision delay followed (500 ms for monkey 1; 1000 ms for monkey 2; analyses used the last 500 ms). Response mapping (which color/target corresponded to “short” vs “long”) and target locations were randomized per trial, preventing motor preparation during the decision delay. Block design: Each recording session included multiple blocks (temporal and spatial). Each block contained eight stimuli (four short, four long), with the boundary differing per block; some stimuli overlapped between blocks but switched category (context dependence). Each block began with 24 training trials (only shortest and longest stimuli, color-coded to define prototypes) followed by 96 test trials (all eight stimuli, green bars). Average behavioral accuracy: monkey 1, 69%; monkey 2, 72%. Recordings: Simultaneous LFPs and single-unit activity were recorded in dorsolateral prefrontal cortex (dIPFC) and pre-supplementary motor area (preSMA). LFPs were recorded at 1 kHz with 250 Hz low-pass filtering. Data comprised 217 preSMA experimental blocks (199 monkey 1; 18 monkey 2; one electrode per block) and 199 dIPFC blocks (monkey 1). Acute recordings required daily dural penetrations; noisy channels/trials were visually rejected (~10%). Data were re-referenced to site-wise averages. Spectral analysis: Power spectra (2–36 Hz) were computed with Hanning taper, 1-s padding, 1-Hz resolution. For each block/region, the channel with maximal beta power (15–35 Hz) was selected. Beta peak frequency was extracted as the frequency of maximal power. Prestimulus (500 ms pre-first stimulus) and decision delay (500 ms) intervals were compared using cluster-based permutation tests in 15–35 Hz. Logistic regression and AUROC quantified decision predictability from spectra. Burst analysis: Time-frequency representations (15–35 Hz; 250-ms sliding window, 20-ms step) identified burst events as time-frequency points >2 SD above mean power, lasting at least one cycle, clustered spatio-spectrally. Burst features extracted per trial included count, peak amplitude, time span, frequency span, timing, peak frequency, and burst volume (double integral). Paired t-tests compared conditions. Instantaneous frequency: Raw data were band-pass filtered around the beta peak (e.g., 25–35 Hz for monkey 1 decision delay), Hilbert phase was differentiated; robust median filtering produced instantaneous frequency time courses, averaged per condition. Cluster-based permutation tests probed 500-ms decision delay differences. Connectivity: Inter-areal coherence between dIPFC and preSMA and nonparametric bivariate Granger causality (GC) were computed from Fourier coefficients during the decision delay, separately for short vs long trials. Peak frequencies were compared (paired t-tests for coherence; two-way ANOVA with factors trial type and direction for GC). Spike-field coherence: Category-selective neurons were identified via sliding-window regression (factors: categorical choice, stimulus magnitude, outcome), choice-probability index (>0.6), and chi-squared test significance. Spike-triggered spectra (2–36 Hz; ±250 ms) were used to compute pairwise phase consistency (PPC). Peak PPC frequencies were compared across category-selective cell groups.

• Prestimulus vs decision delay beta: In both regions, lower beta power (15–27/27.5 Hz) decreased and higher beta power (≈28–35 Hz) increased during decision delay, indicating an upward shift in beta peak frequency.
  • Monkey 1 dIPFC: peak shifted 25→29 Hz; t(198)=23.3, p≤1e−12.
  • Monkey 1 preSMA: 24→27 Hz; t(198)=12.7, p≤1e−12.
  • Monkey 2 preSMA: 21→24 Hz; t(17)=6.8, p=3e−6.
  • Burst profiles also showed higher peak frequencies during decision delay; bursts differed in number, amplitude, and time/frequency spans across periods.
• Category-specific beta frequencies during decision delay: Across tasks and sites, beta on “short” trials was ~2 Hz faster than on “long” trials.
  • Monkey 1 dIPFC: t(198)=11.0, p≤1e−15, AUROC=0.751.
  • Monkey 1 preSMA: t(198)=8.5, p=5e−15, AUROC=0.724.
  • Monkey 2 preSMA: t(17)=3.5, p=0.003, AUROC=0.660.
  • Burst peak frequencies mirrored this effect.
• Behavioral relevance: On incorrect trials, the frequency difference reversed (reflecting the monkey’s chosen category rather than true stimulus category).
  • dIPFC: t(198)=6.8, p=1e−10; bursts: t(198)=−6.6, p=5e−10.
  • preSMA: t(216)=5.7, p=4e−8; bursts: t(216)=−5.5, p=2e−7.
• Context dependence: Identical stimuli belonging to different categories across blocks still showed category-consistent frequency differences (e.g., T3 “very short” vs T2 “very long”: dIPFC t(77)=6.1, p=5e−8; preSMA t(86)=3.4, p=9e−4). No differences within subcategory levels (2-way ANOVA showed main effect of long/short only).
• Time course: Instantaneous frequency separation persisted across the entire 500-ms decision window and peaked ≈350–360 ms before response screen onset; cluster-corrected p≤1e−12 in both regions. On incorrect trials, the temporal profiles were similar but reversed in sign.
• Connectivity: dIPFC–preSMA coherence peaks were at higher frequencies for “short” vs “long” trials; t(198)=5.4, p≤1e−7; reversed on incorrect trials t(163)=−3.0, p=0.003.
  • Granger causality showed a significant interaction of direction and category (F(1,792)=9.3, p=0.002); the frequency shift occurred from dIPFC→preSMA (t(198)=3.7, p=2.6e−4), not from preSMA→dIPFC (t(198)=0.54, p=0.59). On incorrect trials, dIPFC→preSMA shift reversed (t(169)=−3.8, p=2e−4).
• Spike-field coherence: In dIPFC, category-selective neurons synchronized with the beta frequency corresponding to their preferred category, supporting ensemble-specific frequency channels. Overall, two distinct beta frequencies, with distinct burst and connectivity profiles, encoded relative categorical decisions independent of absolute stimulus properties or correctness.

The findings support the hypothesis that beta oscillations implement context-dependent categorization by organizing and reactivating neural ensembles that represent abstract category decisions. The category-specific beta frequencies were consistent across temporal and spatial tasks, indicating that the frequencies encode relative, abstract concepts (“short” vs “long”) rather than physical magnitudes. The reversal of the effect on error trials shows the signal tracks the monkey’s internal decision. The separation was evident in sustained instantaneous frequency differences, burst parameters, and inter-areal interactions, with directional effects suggesting top-down influence from dIPFC to preSMA. These results align with theories of communication-through-coherence and frequency-division multiplexing, whereby distinct oscillatory frequencies route separate information streams to downstream targets. A biophysical account links different beta frequencies to the duration of distal inputs driving pyramidal dendrites, potentially via thalamic modulation. The task design eliminated preparatory motor confounds, strengthening the interpretation that beta frequency channels reflect cognitive decisions rather than motor preparation.

This work demonstrates that two distinct beta-band frequency channels reliably encode relative categorical decisions across contexts, tasks (time and distance), and cortical sites (dIPFC and preSMA). Beta bursts and inter-areal coupling reflect these category-specific channels, with evidence for top-down influence from dIPFC. The results support a framework in which beta oscillations (re-)activate relevant neural ensembles and selectively route their information via frequency-division multiplexing. Future research should test causality by perturbing specific beta sub-bands, examine generalization to humans and additional brain regions, elucidate thalamocortical mechanisms generating frequency-specific beta bursts, and explore translational applications for neurostimulation, neurorehabilitation, and BMIs leveraging distinct beta frequencies.

• Connectivity and spike-field coherence results rely on simultaneous recordings from a single animal, limiting generalizability and precluding strong causal claims.
• LFP signal-to-noise in monkey 2 deteriorated after three sessions, leaving only 18 clean preSMA blocks.
• Acute recordings and single-electrode-per-block selection may limit sampling of spatial heterogeneity.
• Behavioral performance was moderate (≈69–72%), which, while sufficient for analyses, may influence effect magnitudes.
• Different decision delay durations across animals (500 vs 1000 ms; analysis restricted to last 500 ms) may introduce inter-subject variability.