Using normative modeling and machine learning for detecting mild traumatic brain injury from magnetoencephalography data.

New biomarkers are urgently needed for many brain disorders; for example, the diagnosis of mild traumatic brain injury (mTBI) is challenging as the clinical symptoms are diverse and nonspecific. EEG and MEG studies have demonstrated several population-level indicators of mTBI that could serve as objective markers of brain injury. However, deriving clinically useful biomarkers for mTBI and other brain disorders from EEG/MEG signals is hampered by the large inter-individual variability even across healthy people. Here, we used a multivariate machine-learning approach to detect mTBI from resting-state MEG measurements. To address the heterogeneity of the condition, we employed a normative modeling approach and modeled MEG signal features of individual mTBI patients as deviations with respect to the normal variation. To this end, a normative dataset comprising 621 healthy participants was used to determine the variation in power spectra across the cortex. In addition, we constructed normative datasets based on age-matched subsets of the full normative data. To discriminate patients from healthy control subjects, we trained support-vector-machine classifiers on the quantitative deviation maps for 25 mTBI patients and 20 controls not included in the normative dataset. The best performing classifier made use of the full normative data across the entire age and frequency ranges. This classifier was able to distinguish patients from controls with an accuracy of 79%. Inspection of the trained model revealed that low-frequency activity in the theta frequency band (4-8 Hz) is a significant indicator of mTBI, consistent with earlier studies. The results demonstrate the feasibility of using normative modeling of MEG data combined with machine learning to advance diagnosis of mTBI and identify patients that would benefit from treatment and rehabilitation. The current approach could be applied to a wide range of brain disorders, thus providing a basis for deriving MEG/EEG-based biomarkers.

Unsupervised representation learning of spontaneous MEG data with nonlinear ICA.

Resting-state magnetoencephalography (MEG) data show complex but structured spatiotemporal patterns. However, the neurophysiological basis of these signal patterns is not fully known and the underlying signal sources are mixed in MEG measurements. Here, we developed a method based on the nonlinear independent component analysis (ICA), a generative model trainable with unsupervised learning, to learn representations from resting-state MEG data. After being trained with a large dataset from the Cam-CAN repository, the model has learned to represent and generate patterns of spontaneous cortical activity using latent nonlinear components, which reflects principal cortical patterns with specific spectral modes. When applied to the downstream classification task of audio-visual MEG, the nonlinear ICA model achieves competitive performance with deep neural networks despite limited access to labels. We further validate the generalizability of the model across different datasets by applying it to an independent neurofeedback dataset for decoding the subject's attentional states, providing a real-time feature extraction and decoding mindfulness and thought-inducing tasks with an accuracy of around 70% at the individual level, which is much higher than obtained by linear ICA or other baseline methods. Our results demonstrate that nonlinear ICA is a valuable addition to existing tools, particularly suited for unsupervised representation learning of spontaneous MEG activity which can then be applied to specific goals or tasks when labelled data are scarce.

Mapping and decoding cortical engagement during motor imagery, mental arithmetic, and silent word generation using MEG.

Accurate quantification of cortical engagement during mental imagery tasks remains a challenging brain-imaging problem with immediate relevance to developing brain-computer interfaces. We analyzed magnetoencephalography (MEG) data from 18 individuals completing cued motor imagery, mental arithmetic, and silent word generation tasks. Participants imagined movements of both hands (HANDS) and both feet (FEET), subtracted two numbers (SUB), and silently generated words (WORD). The task-related cortical engagement was inferred from beta band (17-25 Hz) power decrements estimated using a frequency-resolved beamforming method. In the hands and feet motor imagery tasks, beta power consistently decreased in premotor and motor areas. In the word and subtraction tasks, beta-power decrements showed engagements in language and arithmetic processing within the temporal, parietal, and inferior frontal regions. A support vector machine classification of beta power decrements yielded high accuracy rates of 74 and 68% for classifying motor-imagery (HANDS vs. FEET) and cognitive (WORD vs. SUB) tasks, respectively. From the motor-versus-nonmotor contrasts, excellent accuracy rates of 85 and 80% were observed for hands-versus-word and hands-versus-sub, respectively. A multivariate Gaussian-process classifier provided an accuracy rate of 60% for the four-way (HANDS-FEET-WORD-SUB) classification problem. Individual task performance was revealed by within-subject correlations of beta-decrements. Beta-power decrements are helpful metrics for mapping and decoding cortical engagement during mental processes in the absence of sensory stimuli or overt behavioral outputs. Markers derived based on beta decrements may be suitable for rehabilitation purposes, to characterize motor or cognitive impairments, or to treat patients recovering from a cerebral stroke.

Activity level in left auditory cortex predicts behavioral performance in inhibition tasks in children.

Sensory processing during development is important for the emerging cognitive skills underlying goal-directed behavior. Yet, it is not known how auditory processing in children is related to their cognitive functions. Here, we utilized combined magneto- and electroencephalographic (M/EEG) measurements in school-aged children (6-14y) to show that child auditory cortical activity at ∼250 ms after auditory stimulation predicts the performance in inhibition tasks. While unaffected by task demands, the amplitude of the left-hemisphere activation pattern was significantly correlated with the variability of behavioral response time. Since this activation pattern is typically not present in adults, our results suggest divergent brain mechanisms in adults and children for consistent performance in auditory-based cognitive tasks. This difference can be explained as a shift in cortical resources for cognitive control from sensorimotor associations in the auditory cortex of children to top-down regulated control processes involving (pre)frontal and cingulate areas in adults.

Calibration and Localization of Optically Pumped Magnetometers Using Electromagnetic Coils.

In this paper, we propose a method to estimate the position, orientation, and gain of a magnetic field sensor using a set of (large) electromagnetic coils. We apply the method for calibrating an array of optically pumped magnetometers (OPMs) for magnetoencephalography (MEG). We first measure the magnetic fields of the coils at multiple known positions using a well-calibrated triaxial magnetometer, and model these discreetly sampled fields using vector spherical harmonics (VSH) functions. We then localize and calibrate an OPM by minimizing the sum of squared errors between the model signals and the OPM responses to the coil fields. We show that by using homogeneous and first-order gradient fields, the OPM sensor parameters (gain, position, and orientation) can be obtained from a set of linear equations with pseudo-inverses of two matrices. The currents that should be applied to the coils for approximating these low-order field components can be determined based on the VSH models. Computationally simple initial estimates of the OPM sensor parameters follow. As a first test of the method, we placed a fluxgate magnetometer at multiple positions and estimated the RMS position, orientation, and gain errors of the method to be 1.0 mm, 0.2°, and 0.8%, respectively. Lastly, we calibrated a 48-channel OPM array. The accuracy of the OPM calibration was tested by using the OPM array to localize magnetic dipoles in a phantom, which resulted in an average dipole position error of 3.3 mm. The results demonstrate the feasibility of using electromagnetic coils to calibrate and localize OPMs for MEG.

Spatial sampling of MEG and EEG based on generalized spatial-frequency analysis and optimal design.

In this paper, we analyze spatial sampling of electro- (EEG) and magnetoencephalography (MEG), where the electric or magnetic field is typically sampled on a curved surface such as the scalp. By simulating fields originating from a representative adult-male head, we study the spatial-frequency content in EEG as well as in on- and off-scalp MEG. This analysis suggests that on-scalp MEG, off-scalp MEG and EEG can benefit from up to 280, 90 and 110 spatial samples, respectively. In addition, we suggest a new approach to obtain sensor locations that are optimal with respect to prior assumptions. The approach also allows to control, e.g., the uniformity of the sensor locations. Based on our simulations, we argue that for a low number of spatial samples, model-informed non-uniform sampling can be beneficial. For a large number of samples, uniform sampling grids yield nearly the same total information as the model-informed grids.

Comparison of beamformer implementations for MEG source localization.

Beamformers are applied for estimating spatiotemporal characteristics of neuronal sources underlying measured MEG/EEG signals. Several MEG analysis toolboxes include an implementation of a linearly constrained minimum-variance (LCMV) beamformer. However, differences in implementations and in their results complicate the selection and application of beamformers and may hinder their wider adoption in research and clinical use. Additionally, combinations of different MEG sensor types (such as magnetometers and planar gradiometers) and application of preprocessing methods for interference suppression, such as signal space separation (SSS), can affect the results in different ways for different implementations. So far, a systematic evaluation of the different implementations has not been performed. Here, we compared the localization performance of the LCMV beamformer pipelines in four widely used open-source toolboxes (MNE-Python, FieldTrip, DAiSS (SPM12), and Brainstorm) using datasets both with and without SSS interference suppression. We analyzed MEG data that were i) simulated, ii) recorded from a static and moving phantom, and iii) recorded from a healthy volunteer receiving auditory, visual, and somatosensory stimulation. We also investigated the effects of SSS and the combination of the magnetometer and gradiometer signals. We quantified how localization error and point-spread volume vary with the signal-to-noise ratio (SNR) in all four toolboxes. When applied carefully to MEG data with a typical SNR (3-15 â€‹dB), all four toolboxes localized the sources reliably; however, they differed in their sensitivity to preprocessing parameters. As expected, localizations were highly unreliable at very low SNR, but we found high localization error also at very high SNRs for the first three toolboxes while Brainstorm showed greater robustness but with lower spatial resolution. We also found that the SNR improvement offered by SSS led to more accurate localization.

Potential of on-scalp MEG: Robust detection of human visual gamma-band responses.

Electrophysiological signals recorded intracranially show rich frequency content spanning from near-DC to hundreds of hertz. Noninvasive electromagnetic signals measured with electroencephalography (EEG) or magnetoencephalography (MEG) typically contain less signal power in high frequencies than invasive recordings. Particularly, noninvasive detection of gamma-band activity (>30 Hz) is challenging since coherently active source areas are small at such frequencies and the available imaging methods have limited spatial resolution. Compared to EEG and conventional SQUID-based MEG, on-scalp MEG should provide substantially improved spatial resolution, making it an attractive method for detecting gamma-band activity. Using an on-scalp array comprised of eight optically pumped magnetometers (OPMs) and a conventional whole-head SQUID array, we measured responses to a dynamic visual stimulus known to elicit strong gamma-band responses. OPMs had substantially higher signal power than SQUIDs, and had a slightly larger relative gamma-power increase over the baseline. With only eight OPMs, we could obtain gamma-activity source estimates comparable to those of SQUIDs at the group level. Our results show the feasibility of OPMs to measure gamma-band activity. To further facilitate the noninvasive detection of gamma-band activity, the on-scalp OPM arrays should be optimized with respect to sensor noise, the number of sensors and intersensor spacing.

On-scalp MEG system utilizing an actively shielded array of optically-pumped magnetometers.

The spatial resolution of magnetoencephalography (MEG) can be increased from that of conventional SQUID-based systems by employing on-scalp sensor arrays of e.g. optically-pumped magnetometers (OPMs). However, OPMs reach sufficient sensitivity for neuromagnetic measurements only when operated in a very low absolute magnetic field of few nanoteslas or less, usually not reached in a typical magnetically shielded room constructed for SQUID-based MEG. Moreover, field drifts affect the calibration of OPMs. Static and dynamic suppression of interfering fields is thus necessary for good-quality neuromagnetic measurements with OPMs. Here, we describe an on-scalp MEG system that utilizes OPMs and external compensation coils that provide static and dynamic shielding against ambient fields. In a conventional two-layer magnetically shielded room, our coil system reduced the maximum remanent DC-field component within an 8-channel OPM array from 70 to less than 1 nT, enabling the sensors to operate in the sensitive spin exchange relaxation-free regime. When compensating field drifts below 4 Hz, a low-frequency shielding factor of 22 dB was achieved, which reduced the peak-to-peak drift from 1.3 to 0.4 nT and thereby the standard deviation of the sensor calibration from 1.7% to 0.5%. Without band-limiting the field that was compensated, a low-frequency shielding factor of 43 dB was achieved. We validated the system by measuring brain responses to electric stimulation of the median nerve. With dynamic shielding and digital interference suppression methods, single-trial somatosensory evoked responses could be detected. Our results advance the deployment of OPM-based on-scalp MEG in lighter magnetic shields.

Adaptive neural network classifier for decoding MEG signals.

We introduce two Convolutional Neural Network (CNN) classifiers optimized for inferring brain states from magnetoencephalographic (MEG) measurements. Network design follows a generative model of the electromagnetic (EEG and MEG) brain signals allowing explorative analysis of neural sources informing classification. The proposed networks outperform traditional classifiers as well as more complex neural networks when decoding evoked and induced responses to different stimuli across subjects. Importantly, these models can successfully generalize to new subjects in real-time classification enabling more efficient brain-computer interfaces (BCI).

The impact of improved MEG-MRI co-registration on MEG connectivity analysis.

Co-registration between structural head images and functional MEG data is needed for anatomically-informed MEG data analysis. Despite the efforts to minimize the co-registration error, conventional landmark- and surface-based strategies for co-registering head and MEG device coordinates achieve an accuracy of typically 5-10 mm. Recent advances in instrumentation and technical solutions, such as the development of hybrid ultra-low-field (ULF) MRI-MEG devices or the use of 3D-printed individualized foam head-casts, promise unprecedented co-registration accuracy, i.e., 2 mm or better. In the present study, we assess through simulations the impact of such an improved co-registration on MEG connectivity analysis. We generated synthetic MEG recordings for pairs of connected cortical sources with variable locations. We then assessed the capability to reconstruct source-level connectivity from these recordings for 0-15-mm co-registration error, three levels of head modeling detail (one-, three- and four-compartment models), two source estimation techniques (linearly constrained minimum-variance beamforming and minimum-norm estimation MNE) and five separate connectivity metrics (imaginary coherency, phase-locking value, amplitude-envelope correlation, phase-slope index and frequency-domain Granger causality). We found that beamforming can better take advantage of an accurate co-registration than MNE. Specifically, when the co-registration error was smaller than 3 mm, the relative error in connectivity estimates was down to one-third of that observed with typical co-registration errors. MNE provided stable results for a wide range of co-registration errors, while the performance of beamforming rapidly degraded as the co-registration error increased. Furthermore, we found that even moderate co-registration errors (>6 mm, on average) essentially decrease the difference of four- and three- or one-compartment models. Hence, a precise co-registration is important if one wants to take full advantage of highly accurate head models for connectivity analysis. We conclude that an improved co-registration will be beneficial for reliable connectivity analysis and effective use of highly accurate head models in future MEG connectivity studies.

Non-Linear Dynamical Analysis of Resting Tremor for Demand-Driven Deep Brain Stimulation.

Parkinson's Disease (PD) is currently the second most common neurodegenerative disease. One of the most characteristic symptoms of PD is resting tremor. Local Field Potentials (LFPs) have been widely studied to investigate deviations from the typical patterns of healthy brain activity. However, the inherent dynamics of the Sub-Thalamic Nucleus (STN) LFPs and their spatiotemporal dynamics have not been well characterized. In this work, we study the non-linear dynamical behaviour of STN-LFPs of Parkinsonian patients using ε -recurrence networks. RNs are a non-linear analysis tool that encodes the geometric information of the underlying system, which can be characterised (for example, using graph theoretical measures) to extract information on the geometric properties of the attractor. Results show that the activity of the STN becomes more non-linear during the tremor episodes and that ε -recurrence network analysis is a suitable method to distinguish the transitions between movement conditions, anticipating the onset of the tremor, with the potential for application in a demand-driven deep brain stimulation system.

Consistency and similarity of MEG- and fMRI-signal time courses during movie viewing.

Movie viewing allows human perception and cognition to be studied in complex, real-life-like situations in a brain-imaging laboratory. Previous studies with functional magnetic resonance imaging (fMRI) and with magneto- and electroencephalography (MEG and EEG) have demonstrated consistent temporal dynamics of brain activity across movie viewers. However, little is known about the similarities and differences of fMRI and MEG or EEG dynamics during such naturalistic situations. We thus compared MEG and fMRI responses to the same 15-min black-and-white movie in the same eight subjects who watched the movie twice during both MEG and fMRI recordings. We analyzed intra- and intersubject voxel-wise correlations within each imaging modality as well as the correlation of the MEG envelopes and fMRI signals. The fMRI signals showed voxel-wise within- and between-subjects correlations up to r = 0.66 and r = 0.37, respectively, whereas these correlations were clearly weaker for the envelopes of band-pass filtered (7 frequency bands below 100 Hz) MEG signals (within-subjects correlation r < 0.14 and between-subjects r < 0.05). Direct MEG-fMRI voxel-wise correlations were unreliable. Notably, applying a spatial-filtering approach to the MEG data uncovered consistent canonical variates that showed considerably stronger (up to r = 0.25) between-subjects correlations than the univariate voxel-wise analysis. Furthermore, the envelopes of the time courses of these variates up to about 10 Hz showed association with fMRI signals in a general linear model. Similarities between envelopes of MEG canonical variates and fMRI voxel time-courses were seen mostly in occipital, but also in temporal and frontal brain regions, whereas intra- and intersubject correlations for MEG and fMRI separately were strongest only in the occipital areas. In contrast to the conventional univariate analysis, the spatial-filtering approach was able to uncover associations between the MEG envelopes and fMRI time courses, shedding light on the similarities of hemodynamic and electromagnetic brain activities during movie viewing.

Evidence for a general performance-monitoring system in the human brain.

Adaptive behavior relies on the ability of the brain to form predictions and monitor action outcomes. In the human brain, the same system is thought to monitor action outcomes regardless of whether the information originates from internal (e.g., proprioceptive) and external (e.g., visual) sensory channels. Neural signatures of processing motor errors and action outcomes communicated by external feedback have been studied extensively; however, the existence of such a general action-monitoring system has not been tested directly. Here, we use concurrent EEG-MEG measurements and a probabilistic learning task to demonstrate that event-related responses measured by electroencephalography and magnetoencephalography display spatiotemporal patterns that allow an effective transfer of a multivariate statistical model discriminating the outcomes across the following conditions: (a) erroneous versus correct motor output, (b) negative versus positive feedback, (c) high- versus low-surprise negative feedback, and (d) erroneous versus correct brain-computer-interface output. We further show that these patterns originate from highly-overlapping neural sources in the medial frontal and the medial parietal cortices. We conclude that information about action outcomes arriving from internal or external sensory channels converges to the same neural system in the human brain, that matches this information to the internal predictions.

Requirements for Coregistration Accuracy in On-Scalp MEG.

Recent advances in magnetic sensing has made on-scalp magnetoencephalography (MEG) possible. In particular, optically-pumped magnetometers (OPMs) have reached sensitivity levels that enable their use in MEG. In contrast to the SQUID sensors used in current MEG systems, OPMs do not require cryogenic cooling and can thus be placed within millimetres from the head, enabling the construction of sensor arrays that conform to the shape of an individual's head. To properly estimate the location of neural sources within the brain, one must accurately know the position and orientation of sensors in relation to the head. With the adaptable on-scalp MEG sensor arrays, this coregistration becomes more challenging than in current SQUID-based MEG systems that use rigid sensor arrays. Here, we used simulations to quantify how accurately one needs to know the position and orientation of sensors in an on-scalp MEG system. The effects that different types of localisation errors have on forward modelling and source estimates obtained by minimum-norm estimation, dipole fitting, and beamforming are detailed. We found that sensor position errors generally have a larger effect than orientation errors and that these errors affect the localisation accuracy of superficial sources the most. To obtain similar or higher accuracy than with current SQUID-based MEG systems, RMS sensor position and orientation errors should be [Formula: see text] and [Formula: see text], respectively.

Recovery of the 20 Hz Rebound to Tactile and Proprioceptive Stimulation after Stroke.

Sensorimotor integration is closely linked to changes in motor-cortical excitability, observable in the modulation of the 20 Hz rhythm. After somatosensory stimulation, the rhythm transiently increases as a rebound that reflects motor-cortex inhibition. Stroke-induced alterations in afferent input likely affect motor-cortex excitability and motor recovery. To study the role of somatosensory afferents in motor-cortex excitability after stroke, we employed magnetoencephalographic recordings (MEG) at 1-7 days, one month, and 12 months in 23 patients with stroke in the middle cerebral artery territory and 22 healthy controls. The modulation of the 20 Hz motor-cortical rhythm was evaluated to two different somatosensory stimuli, tactile stimulation, and passive movement of the index fingers. The rebound strengths to both stimuli were diminished in the acute phase compared to the controls and increased significantly during the first month after stroke. However, only the rebound amplitudes to tactile stimuli fully recovered within the follow-up period. The rebound strengths in the affected hemisphere to both stimuli correlated strongly with the clinical scores across the follow-up. The results show that changes in the 20 Hz rebound to both stimuli behave similarly and occur predominantly during the first month. The 20 Hz rebound is a potential marker for predicting motor recovery after stroke.

Measuring MEG closer to the brain: Performance of on-scalp sensor arrays.

Optically-pumped magnetometers (OPMs) have recently reached sensitivity levels required for magnetoencephalography (MEG). OPMs do not need cryogenics and can thus be placed within millimetres from the scalp into an array that adapts to the individual head size and shape, thereby reducing the distance from cortical sources to the sensors. Here, we quantified the improvement in recording MEG with hypothetical on-scalp OPM arrays compared to a 306-channel state-of-the-art SQUID array (102 magnetometers and 204 planar gradiometers). We simulated OPM arrays that measured either normal (nOPM; 102 sensors), tangential (tOPM; 204 sensors), or all components (aOPM; 306 sensors) of the magnetic field. We built forward models based on magnetic resonance images of 10 adult heads; we employed a three-compartment boundary element model and distributed current dipoles evenly across the cortical mantle. Compared to the SQUID magnetometers, nOPM and tOPM yielded 7.5 and 5.3 times higher signal power, while the correlations between the field patterns of source dipoles were reduced by factors of 2.8 and 3.6, respectively. Values of the field-pattern correlations were similar across nOPM, tOPM and SQUID gradiometers. Volume currents reduced the signals of primary currents on average by 10%, 72% and 15% in nOPM, tOPM and SQUID magnetometers, respectively. The information capacities of the OPM arrays were clearly higher than that of the SQUID array. The dipole-localization accuracies of the arrays were similar while the minimum-norm-based point-spread functions were on average 2.4 and 2.5 times more spread for the SQUID array compared to nOPM and tOPM arrays, respectively.

Neural signatures of hand kinematics in leaders vs. followers: A dual-MEG study.

During joint actions, people typically adjust their own actions according to the ongoing actions of the partner, which implies that the interaction modulates the behavior of both participants. However, the neural substrates of such mutual adaptation are still poorly understood. Here, we set out to identify the kinematics-related brain activity of leaders and followers performing hand actions. Sixteen participants as 8 pairs performed continuous, repetitive right-hand opening and closing actions with ~3-s cycles in a leader-follower task. Subjects played each role for 5min. Magnetoencephalographic (MEG) brain signals were recorded simultaneously from both partners with a dual-MEG setup, and hand kinematics was monitored with accelerometers. Modulation index, a cross-frequency coupling measure, was computed between the hand acceleration and the MEG signals in the alpha (7-13Hz) and beta (13-25Hz) bands. Regardless of the participants' role, the strongest alpha and beta modulations occurred bilaterally in the sensorimotor cortices. In the occipital region, beta modulation was stronger in followers than leaders; these oscillations originated, according to beamformer source reconstructions, in early visual cortices. Despite differences in the modulation indices, alpha and beta power did not differ between the conditions. Our results indicate that the beta modulation in the early visual cortices depends on the subject's role as a follower or leader in a joint hand-action task. This finding could reflect the different strategies employed by leaders and followers in integrating kinematics-related visual information to control their own actions.

Towards brain-activity-controlled information retrieval: Decoding image relevance from MEG signals.

We hypothesize that brain activity can be used to control future information retrieval systems. To this end, we conducted a feasibility study on predicting the relevance of visual objects from brain activity. We analyze both magnetoencephalographic (MEG) and gaze signals from nine subjects who were viewing image collages, a subset of which was relevant to a predetermined task. We report three findings: i) the relevance of an image a subject looks at can be decoded from MEG signals with performance significantly better than chance, ii) fusion of gaze-based and MEG-based classifiers significantly improves the prediction performance compared to using either signal alone, and iii) non-linear classification of the MEG signals using Gaussian process classifiers outperforms linear classification. These findings break new ground for building brain-activity-based interactive image retrieval systems, as well as for systems utilizing feedback both from brain activity and eye movements.

Feature-specific information processing precedes concerted activation in human visual cortex.

Current knowledge about the precise timing of visual input to the cortex relies largely on spike timings in monkeys and evoked-response latencies in humans. However, quantifying the activation onset does not unambiguously describe the timing of stimulus-feature-specific information processing. Here, we investigated the information content of the early human visual cortical activity by decoding low-level visual features from single-trial magnetoencephalographic (MEG) responses. MEG was measured from nine healthy subjects as they viewed annular sinusoidal gratings (spanning the visual field from 2 to 10° for a duration of 1 s), characterized by spatial frequency (0.33 cycles/degree or 1.33 cycles/degree) and orientation (45° or 135°); gratings were either static or rotated clockwise or anticlockwise from 0 to 180°. Time-resolved classifiers using a 20 ms moving window exceeded chance level at 51 ms (the later edge of the window) for spatial frequency, 65 ms for orientation, and 98 ms for rotation direction. Decoding accuracies of spatial frequency and orientation peaked at 70 and 90 ms, respectively, coinciding with the peaks of the onset evoked responses. Within-subject time-insensitive pattern classifiers decoded spatial frequency and orientation simultaneously (mean accuracy 64%, chance 25%) and rotation direction (mean 82%, chance 50%). Classifiers trained on data from other subjects decoded the spatial frequency (73%), but not the orientation, nor the rotation direction. Our results indicate that unaveraged brain responses contain decodable information about low-level visual features already at the time of the earliest cortical evoked responses, and that representations of spatial frequency are highly robust across individuals.