High Frequency Oscillations (HFOs)

• HFOs are defined as transient bursts of high-frequency signals (above 80 Hz) observed in EEG, astrophysics, and plasma experiments.
• They serve as biomarkers to localize epileptogenic tissue, probe accretion physics near black holes, and characterize wave dynamics in solar and plasma studies.
• Recent detection methods combine time–frequency analysis and machine learning, achieving high sensitivity and precision in controlled settings.

High Frequency Oscillations (HFOs) are oscillatory phenomena identified at the high-frequency end of the spectrum within a given discipline rather than by a single universal band. In EEG and iEEG, HFOs are transient bursts of neuronal activity above about 80 Hz, with canonical ripple and fast-ripple subtypes. In black-hole timing, the related term High-Frequency Quasi-Periodic Oscillations (HFQPOs) denotes narrow peaks in X-ray power-density spectra at tens to a few hundred Hz. In solar physics, HFOs have been reported as coherent, approximately one-minute oscillations above a sunspot umbra, while in partially magnetized plasma discharges they appear as coherent modes in the 60–100 MHz range. Across these contexts, HFOs are used to probe localized structure, transport, stability, and wave propagation, but their interpretation remains domain-specific (Yaser et al., 26 Aug 2025, Varniere et al., 2019, Wang et al., 2018, Przybocki et al., 2023).

1. Terminology and disciplinary scope

In neurophysiology, the standard definition places HFOs above about 80 Hz, with ripples at 80–250 Hz and fast ripples at 250–500 Hz. The literature explicitly distinguishes HFOs from gamma-band activity, since gamma is generally 30–80 Hz whereas HFOs begin at at least 80 Hz. That distinction is central in epilepsy and in emerging work on neurodegeneration, because physiological hippocampal ripples associated with memory consolidation are not equivalent to pathological HFOs indexing abnormal synchrony (Yaser et al., 26 Aug 2025).

In accretion physics, HFQPOs are a specific manifestation of HFOs in strong-gravity flows. They appear as narrow peaks in the X-ray power-density spectrum of accreting stellar-mass black holes at tens to a few hundred Hz, often singly and sometimes in near-integer pairs such as 2:3 and 3:4. Their frequencies lie near dynamical frequencies of the inner disk and are therefore used as probes of relativistic orbital motion, epicyclic structure, and mode excitation close to the innermost stable circular orbit (Varniere et al., 2019).

In solar and laboratory plasma physics, the same label refers to very different frequency ranges and mechanisms. Above a sunspot umbra, a distinct band between 10 and 14 mHz with a peak near 12 mHz has been detected simultaneously from the photosphere to the corona. In a neon direct-current magnetron discharge, coherent oscillations at about 60–65 MHz occur at low current, while higher current produces secondary 5–10 MHz modes together with broadened 60–100 MHz fluctuations (Wang et al., 2018, Przybocki et al., 2023).

This disciplinary spread makes a single physical definition impossible. A consistent reading of the literature is that “high frequency” is always relative to the native dynamical spectrum of the system under study. The shared feature is not a common absolute frequency, but the use of localized high-frequency structure as a diagnostic of hidden dynamics.

2. Neurophysiology, epilepsy, and neurodegeneration

In epilepsy research, HFOs are brief, localized bursts of high-frequency activity that have become established biomarkers of epileptogenic tissue. Clinical usage partitions them into ripples and fast ripples, with pathologic fast ripples associated with hyper-synchronized neuronal clusters and seizure-generating structures. Resection of regions generating high HFO rates is associated with better outcomes, and the co-occurrence of interictal spikes with HFOs strongly predicts post-surgical seizure freedom (Yaser et al., 26 Aug 2025).

The neurophysiological literature also emphasizes that HFOs are heterogeneous. Physiological ripples occur during awake immobility and slow-wave sleep, whereas pathological fast ripples are linked to seizure-generating networks. Big-data rat hippocampal EEG analysis showed that HFOs exhibit on–off intermittency with algebraic scaling laws, with on-interval durations following $P(T)\propto T^{-\alpha}$ and heavier tails after the first seizure. This suggests that HFO organization is not merely event-count based, but reflects changes in the temporal structure of network synchrony across disease phases (Huang et al., 2016).

Noninvasive detection is more difficult than intracranial detection. In scalp EEG, HFOs are brief events between 15 and 100 ms with regular small-amplitude oscillations within 80–500 Hz, but they occur at lower rate, lower amplitude, and are more frequently contaminated by artifacts such as eye and muscle activity. High-density EEG has therefore been used to improve spatial resolution, and expert labeling in one study required a minimum of three regular oscillations clearly distinguishable from background and above 80 Hz (Milon-Harnois et al., 2023).

The same biomarker framework is now being extended beyond epilepsy. Emerging evidence supports HFOs as markers of network hyperexcitability in Alzheimer’s disease, including preclinical stages, and animal models of Alzheimer’s disease exhibit HFOs described as indistinguishable from epilepsy models. Public datasets differ sharply in HFO suitability: ADFTD at 500 Hz and BrainLat at 512 Hz support ripple-band analysis below about 250 Hz, PEARL-Neuro at 1000 Hz supports ripple and fast-ripple analysis, whereas 128 Hz datasets such as Pineda et al. and Vicchietti et al. are not HFO-compatible (Yaser et al., 26 Aug 2025).

Intraoperative use has motivated online detectors. A spiking neural network for intraoperative ECoG targeted fast ripples at 250–500 Hz, detected median pre-resection rates of 6.6 HFO/min, and used a residual threshold of at least 1 HFO/min post-resection to predict outcome. In that eight-patient cohort, postsurgical seizure outcome was “predicted” with 100% accuracy, although the small sample size limits generalization (Burelo et al., 2020).

3. Signal processing, machine learning, and event characterization

HFO analysis has driven an unusually broad methodological literature because the events are brief, weak, and often embedded in non-stationary backgrounds. One major line of work uses explicit time–frequency methods. A recent detector based on the Stockwell S-transform combines Teager–Kaiser energy enhancement, Otsu thresholding of time–frequency blobs, multi-domain feature extraction, and unsupervised hierarchical clustering. On a controlled dataset it reported Sensitivity = 97.67%, Precision = 98.57%, and F-score = 97.78%, while patient validation yielded a ratio of 0.73 between HFO rates in resected versus non-resected contacts (Mohammadpour et al., 8 Oct 2025).

A separate direction avoids manual labels. The SS2LD framework begins from legacy detectors such as STE and MNI, converts candidate events to Morlet time–frequency images of size 64×64, learns a latent morphology embedding with a variational autoencoder, derives weak labels by hierarchical $k$-means clustering, and then trains a classifier with VAE-based augmentation. On a multi-institutional interictal iEEG dataset, SS2LD improved patient-level outcome prediction and preserved-region specificity over eHFO and spkHFO, reaching ACC 0.612 ± 0.131, F1 0.464 ± 0.069, and SPEC 0.749 on the open iEEG cohort, and ACC 0.640 ± 0.037, F1 0.583 ± 0.050, and SPEC 0.909 on the Zurich cohort (Zhang et al., 19 Jul 2025).

Deep learning has also been adapted to scalp HFO identification. In one HD-EEG study, 200 ms segments were mapped to STFT images resized to 256×256 and classified by a CNN with three convolutional layers with 16, 32, and 64 filters, kernel size $(3,3)$, followed by a dense layer of 500 neurons. Over 12 runs, the best performance came from RGB images, with Precision 86% ± 2%, Recall 85% ± 3%, Specificity 86% ± 2%, and F1-score 85% ± 1%; red-channel grayscale performed nearly as well with lower computational cost (Milon-Harnois et al., 2023).

Several methods focus on representation rather than direct classification. Gliske et al. analyzed 1.6 million iEEG HFO detections using 33 features and showed that, locally within a channel and state, HFO-feature manifolds are approximately linear: PCA with dimensionality set to the nonlinear local intrinsic dimension captured median 99.8% variance. At the same time, the manifold structure varied strongly across channels, arguing against a single global HFO feature space and favoring channel-specific models (Gliske et al., 2015).

Older transform-based approaches remain important because they expose spectral structure directly. The damped-oscillator oscillator detector (DOOD) models EEG as the driving force on a bank of damped oscillators and was used to identify four rat HFO bands at 80–250 Hz, 250–500 Hz, 600–900 Hz, and 1000–3000 Hz. In human temporal lobe data low-passed at 1000 Hz, HFO-associated peaks occurred at 15 Hz, 85 Hz, 400 Hz, and 700 Hz. The method was explicitly proposed as a way to couple automated HFO detection to wide-band contextual spectra that may help distinguish pathological from normal events (Hsu et al., 2013).

A complementary hardware-oriented literature replaces conventional DSP with event-driven architectures. A mixed-signal neuromorphic chip integrated recording headstages, on-chip filtering into ripple and fast-ripple bands, asynchronous delta-modulator conversion to spikes, and a multi-core SNN detector on the same die. Fabricated in 0.18 $\mu$m CMOS with a die area of 99 mm$^2$, it consumed 614.3 $\mu$W on average during HFO detection and retrospectively predicted postsurgical outcome with accuracy 78%, specificity 100%, and sensitivity 33% in nine temporal lobe epilepsy patients (Sharifshazileh et al., 2020).

4. Solar and laboratory plasma manifestations

In the solar atmosphere above a sunspot umbra, HFOs were identified through simultaneous New Vacuum Solar Telescope and SDO/AIA observations. A distinct band between 10 and 14 mHz, peaking near 12 mHz, was detected at all sampled heights, from the photosphere through the chromosphere and transition region into the corona. Fourier spectra of de-trended umbral mean intensity showed significant peaks above 95% confidence levels in all six channels, while the same 10–14 mHz power was absent in a distant quiet photospheric region. Reconstruction of the 12 mHz component in AIA 171 Å revealed intermittent outwardly propagating disturbances along coronal fan structures, with a measured phase speed of $v_{\rm ph}\approx 49\pm 4\ {\rm km\,s^{-1}}$, well below the coronal sound speed and therefore consistent with upwardly propagating magnetoacoustic slow waves (Wang et al., 2018).

A related but distinct solar result concerns spicules. One-dimensional expanding-flux-tube MHD simulations showed that observed spicule oscillations with periods of 40–50 s can arise by longitudinal-to-transverse mode conversion around the chromospheric equipartition layer where $c_s\approx v_A$. In that interpretation, photospheric longitudinal motions at 5 mHz steepen in the lower chromosphere, generate high-frequency components, and efficiently convert them into transverse waves around $z\approx 1.0$–$1.5$ Mm. The characteristic HFO period corresponds to the sound-crossing time of the mode conversion region, with $k$0 and simulated root-mean-square transverse velocities consistent with observed values (Shoda et al., 2018).

Laboratory plasma work uses the term at still higher absolute frequencies. In a neon direct-current magnetron discharge with a 2 mm cathode–anode gap, low discharge current produced highly coherent 60–65 MHz oscillations, while current above about 13 mA generated additional 5–10 MHz modes and broadened 60–100 MHz fluctuations. The total discharge current oscillated at the same frequencies as the segmented-anode currents, implying axial propagation. A lower-hybrid interpretation was supported by characteristic plasma frequencies and by fitted mode structure: in the single-frequency regime the dominant mode was $k$1, with azimuthal wavelength about 1 cm, axial wavelength about 120 $k$2m, and axial phase velocity about $k$3 (Przybocki et al., 2023).

These solar and plasma studies show that HFOs need not denote localized transients in a neural recording. In wave-bearing media they can instead mark narrow propagating modes, conversion layers, or instability-driven branches. A plausible implication is that the term is best understood as a phenomenological label whose mechanistic content only becomes precise after the propagation geometry, dispersion relation, and forcing are specified.

5. Accretion flows, compact objects, and strong gravity

In X-ray binaries, HFQPOs are narrow peaks in the power-density spectrum at frequencies from about 27 Hz up to about 450 Hz. They have been detected in eight black-hole binaries, can appear singly or in pairs with near-integer ratios such as 2:3 and 3:4, and have Q-factors of order 5–20 with fractional rms amplitudes of a few percent. This phenomenology motivates models in which HFOs reflect non-axisymmetric instabilities, trapped modes, or relativistic orbital frequencies in the inner accretion flow (Varniere et al., 2019).

One model attributes HFQPOs to the Rossby Wave Instability (RWI), triggered at a radius where the disk vortensity has an extremum. In thin barotropic disks, a standard vortensity-like quantity is

$k$4

and the instability is triggered when $k$5 vanishes. GR-AMRVAC plus GYOTO calculations showed that different dominant azimuthal modes can reproduce observed frequency ratios, and that modest changes of about 30% in the parameters of the surface-density bump can switch the power spectrum from a 2:3 to a 3:4 pair for the same black-hole spin and corotation radius. The same study also found that when RWI is triggered near the last stable orbit, the modulation amplitude increases with spin (Varniere et al., 2019).

Another line of work identifies HFOs with trapped $k$6-mode oscillations in thermally unstable Kerr-metric slim disks. There the oscillations arise near the sonic point, propagate outward, and produce a fundamental frequency close to the maximum radial epicyclic frequency together with integer overtones. In model S8, for example, the fundamental was about 74.9 Hz for a 10 $k$7 Schwarzschild case, while in the high-spin model S30 with $k$8 and $k$9 the measured fundamental was about 285.6 Hz. The paper argues that time-dependent ray tracing is required because relativistic boosting, redshift, and time-delay effects reshape the observed power spectrum (Xue et al., 2015).

Observationally, GRS 1915+105 remains the best-studied black-hole source of this kind. A systematic RXTE survey of 1807 observations found 51 detections above 3$(3,3)$0 single-trial significance, with 48 of them between 63.5 and 71.3 Hz. The average centroid frequency was 67.3 ± 2.0 Hz, the average FWHM was 4.4 ± 2.4 Hz, and the HFQPO was detected only in a specific region of the color–color diagram corresponding to state B, when the energy spectrum is dominated by a bright accretion disk component. At the same time, the rms spectrum was very hard and did not flatten up to 40 keV, where the fractional rms reached 11% (Belloni et al., 2013).

A closely related result came from IGR J17091–3624. RXTE observations during its 2011 outburst revealed a 66.5 ± 0.5 Hz HFQPO with quality factor 7.8 ± 1.5, fractional rms 4.9 ± 0.2%, and significance 8.5$(3,3)$1 in a $(3,3)$2-like interval, plus a marginal 164 ± 10 Hz peak at about 4.5$(3,3)$3. The paper explicitly noted that the close match between the 66 Hz QPO in IGR J17091–3624 and the 67 Hz QPO in GRS 1915+105 raises questions about simple mass scaling of QPO frequency (Altamirano et al., 2012).

High-frequency oscillations are not confined to black holes. Global 3D MHD simulations of accretion onto a weakly magnetized, tilted-dipole neutron star found persistent boundary-layer oscillations excited by magnetic braking and an $(3,3)$4 density wave. In the neutron-star example shown, the boundary-layer mode frequency corrected to Schwarzschild geometry was about 1170 Hz, well below the Keplerian frequency at the stellar surface but within the range of upper kHz QPOs in low-mass X-ray binaries. The same mechanism was suggested as relevant to neutron stars, white dwarfs, and classical T Tauri stars (Blinova et al., 2013).

Strong-gravity timing has also been used to test alternative spacetimes. A generalized epicyclic formalism was applied to microquasar, AGN, and Sgr A* HFQPO data under epicyclic resonance and relativistic precession models across many Kerr-like metrics. The comparative result was not a single preferred theory: perfect-fluid dark-matter spacetimes gave the best epicyclic-resonance fits for microquasars, rotating regular black holes in conformal massive gravity gave the best relativistic-precession fits, and Hairy black holes were the only spacetimes fitting Sgr A* in the epicyclic-resonance model (Shahzadi et al., 2023).

6. Interpretation, limitations, and open problems

A persistent misconception in the biomedical literature is that HFOs are interchangeable with high gamma. The reviewed material rejects that equivalence: gamma is generally 30–80 Hz, whereas HFOs begin at at least 80 Hz, and the distinction matters because physiological ripples, pathological fast ripples, and spike-coupled events do not share the same clinical meaning. The same literature also emphasizes that there is still no clear consensus on an exact morphologic definition of pathological HFOs, which explains the low inter-rater reliability noted in supervised labeling studies (Yaser et al., 26 Aug 2025, Zhang et al., 19 Jul 2025).

Methodological limits remain substantial. Some pipelines achieve very high controlled-dataset performance, but labels in patient data are often subjective, sensitivity and specificity against expert ground truth are sometimes absent, and artifact handling varies sharply across studies. In scalp EEG work, STFT parameters, optimizer choice, and loss function were not specified; in big-data EMD work, formal confidence intervals for scaling exponents were not reported; in automated S-transform detection, low SNR remained a clear failure mode, particularly at 0 dB (Milon-Harnois et al., 2023, Huang et al., 2016, Mohammadpour et al., 8 Oct 2025).

Solar and plasma studies face different constraints. Imaging-only observations of sunspot-umbra HFOs lacked direct velocity measurements and magnetic field extrapolations, while broad formation heights and line-of-sight integration complicate propagation-speed estimates. In magnetron discharges, key plasma parameters were partly estimated from simulations and comparable experiments rather than measured locally, so the lower-hybrid interpretation is semi-quantitative rather than fully closed (Wang et al., 2018, Przybocki et al., 2023).

In strong-gravity astrophysics, the principal controversy is model degeneracy. The approximately 66–67 Hz coincidence between IGR J17091–3624 and GRS 1915+105 challenges naive inverse-mass scaling, and comparative spacetime studies show that no single non-Kerr deformation reproduces all HFQPO datasets under a single geodesic model. This suggests that frequency ratios, mode selection, and state dependence cannot be interpreted independently of accretion physics, emission geometry, and the specific timing model used (Altamirano et al., 2012, Shahzadi et al., 2023).

The broad lesson across disciplines is that HFOs are powerful but not self-interpreting observables. In epilepsy they can localize epileptogenic tissue and may become cross-disease markers of hyperexcitability. In solar and laboratory plasmas they resolve wave conversion, instability, and turbulent cascade. In accretion physics they encode the dynamics of inner disks, boundary layers, and strong-field spacetime. Their scientific value therefore lies less in the acronym itself than in the combination of high-frequency selectivity, contextual modeling, and rigorous validation that turns an oscillatory signature into a physically or clinically meaningful marker.