To build a brain-computer interface (BCI), we must translate the continuous electrical crackle of the brain into discrete digital commands. Historically, neuroscientists faced a stark trade-off: they could record individual neurons with highly invasive penetrating microelectrodes, or record broad, blurry brain waves from the scalp using non-invasive EEG.
Intracranial methods like electrocorticography (ECoG) and intracranial EEG (iEEG) offer a powerful middle ground, sitting directly on or just beneath the surface of the brain. But within these complex intracranial signals, which specific features actually carry the user's precise intent?
The answer lies in high-frequency neural oscillations, primarily the high-gamma band (typically 70–150 Hz). Across motor control, speech production, and sensory processing, high-gamma power has emerged as the premier feature for neural decoding because it acts as a highly localized, high-resolution proxy for the synchronized firing of local neuronal populations.
Focal Activation vs. Global Readiness
Why does the high-gamma band consistently outperform traditional lower-frequency rhythms like alpha (8–12 Hz) or beta (13–30 Hz) when training a decoder? The key lies in how these different frequencies respond to behavior.
As reviewed by Wing-kin Tam et al. in Human motor decoding from neural signals: a review, lower frequencies operate via event-related desynchronization. When a brain region prepares for action, its alpha and beta rhythms suppress over a wide, diffuse area. While this suppression indicates that the motor system is "ready," it lacks the spatial and temporal sharpness required to distinguish between fine movements.
In contrast, high-gamma power increases sharply and locally during active execution. As Hiroaki Hashimoto et al. demonstrated in Swallowing-related neural oscillation: An intracranial EEG study, beta suppression acts as a broad indicator of system-wide motor readiness, whereas high-gamma bursts exhibit extreme spatial and temporal specificity to distinct sub-tasks within a sequence. This focal nature allows decoders to pinpoint exactly what the user intends to do, and exactly when.
Low Frequencies (Alpha/Beta): [ Broad Desynchronization ] --> "The system is ready"
High Frequencies (High-Gamma): [ Focal Burst ] --> "Execute specific sub-task"
A Stable Proxy for Population Dynamics
While recording individual single-unit spikes provides the highest theoretical resolution, individual neurons are notoriously difficult to track over long periods due to micro-movements of the tissue or immune responses to penetrating needles. High-frequency field potentials bypass this issue by capturing aggregate population dynamics.
The authors of Representation and decoding of bilateral arm motor imagery using unilateral cerebral LFP signals, Jiafan Lin et al., showed that high-frequency local field potentials (135–300 Hz) provide superior information for decoding complex motor imagery compared to lower bands, serving as a robust and stable alternative to individual neuronal spikes.
This phenomenon is deeply rooted in how neural populations encode information. Even in subcortical structures, the temporal structure of high-frequency population activity carries immense detail. For example, Aniket S. Kaloti et al. found in Representation of Stimulus Speed and Direction in Vibrissal-Sensitive Regions of the Trigeminal Nuclei: A Comparison of Single Unit and Population Responses that high-frequency components of multi-unit activity (1000–3000 Hz) are more sensitive to fine sensory changes than raw single-unit spike rates. By extracting features from these high frequencies, decoding models leverage the redundant, highly reliable nature of population-level dynamics.
Transforming Raw Waves into Decodable Features
To use high-gamma activity in a machine learning model, the raw, oscillating voltage time-series must be transformed into a stable feature representation. This is typically achieved by extracting the spectral power or analytic amplitude within the target frequency band.
As Jesse A. Livezey et al. explained in Deep learning as a tool for neural data analysis: Speech classification and cross-frequency coupling in human sensorimotor cortex, computing the analytic amplitude of specific bands transforms raw data into features that directly track underlying biological processes, making them highly effective inputs for deep neural networks.
Once extracted, these power features can be fed into diverse model architectures. For instance, Hiroaki Hashimoto et al. in The Analysis and Decoding of Swallowing-related Neural Activities Using Intracranial Electrodes mapped high-gamma power (75–150 Hz) into spatial-temporal representations, effectively converting neural signals into images that could be classified with computer vision models to detect swallowing intentions.
Generalization Across Motor and Cognitive Domains
Because high-gamma power tracks localized population activity, it serves as a universal currency for decoding diverse neural systems, ranging from gross motor control to abstract language.
Gross Motor Control and Prosthetics
In real-time BCI applications, Jennifer L Collinger et al. demonstrated in An Electrocorticographic Brain Interface in an Individual with Tetraplegia that the 70–110 Hz range provides the most accurate representation of motor intent for real-time 3D prosthetic arm control. Crucially, Jennifer L Collinger et al. also showed in Motor-related brain activity during action observation: a neural substrate for electrocorticographic brain-computer interfaces after spinal cord injury that these high-gamma signatures are modulated even when a patient merely observes a movement. This allows BCIs to be calibrated for paralyzed individuals who cannot move to generate training data. This clinical utility is echoed by Masayuki Hirata et al. in Brain-Machine Interfaces : Principles and Clinical Application, who utilized the 80–150 Hz gamma band to reliably classify discrete movements in patients with severe paralysis.
Speech and Semantics
The same principle applies to language decoding. Hemmings Wu et al. in Speech decoding using cortical and subcortical electrophysiological signals extracted power features from the 70–150 Hz high-gamma band across both cortical and subcortical structures to decode the complex phonetic nuances of speech.
Pushing this further into abstract cognition, Richard Antonello et al. developed Neuro2Semantic: A Transfer Learning Framework for Semantic Reconstruction of Continuous Language from Human Intracranial EEG. They used high-gamma features (70–150 Hz) as the foundational input to map continuous brain activity directly to high-level semantic text embeddings, proving that high-frequency neural features capture not just the mechanics of execution, but the deeper meaning behind neural processing.
Key Foundational Work
An Electrocorticographic Brain Interface in an Individual with Tetraplegia (2013) — contributed 2 concept elements to this direction.
Brain-Machine Interfaces : Principles and Clinical Application (2013) — contributed 1 concept elements to this direction.
Motor-related brain activity during action observation: a neural substrate for electrocorticographic brain-computer interfaces after spinal cord injury (2014) — contributed 1 concept elements to this direction.