🧠 neuroscience

Disentangling evoked and induced activity from single-trial neurophysiological time series

This paper introduces the Evoked-Ongoing/Induced Separator (EOIS), a novel framework that successfully disentangles evoked, ongoing, and induced neural activity from single-trial MEG and EEG data by leveraging phase-consistency features, thereby enabling subject-agnostic analysis and advancing translational applications like biomarker refinement and closed-loop neuromodulation.

Original authors: Ahola, O., Haxel, L., Ziemann, U., Ilmoniemi, R., Lioumis, P.

Published 2026-10-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ahola, O., Haxel, L., Ziemann, U., Ilmoniemi, R., Lioumis, P.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The human brain is a constant, churning river of electrical activity, even when we are sitting still. This background hum, known as ongoing activity, is always present. But when something happens—a flash of light, a sound, or a gentle tap on the head—the brain reacts with a specific, timed response. Scientists call this the evoked response. For decades, researchers have studied these reactions by recording the brain's electrical signals and averaging them over many repetitions. This averaging acts like a filter: the random background noise washes out, leaving only the clear, repeating signal of the brain's reaction to the event. However, this method throws away a crucial piece of the puzzle. It hides the fact that every single reaction is unique, mixed with the brain's own shifting mood and energy at that exact moment. To understand how the brain truly works in real time, and to develop treatments that can adjust to a patient's immediate state, scientists need to see the reaction and the background noise separately, in every single instance, not just in the average.

A team of researchers has developed a new tool to solve this problem, allowing them to untangle these two mixed signals from a single moment in time. They call their method the Evoked–Ongoing/Induced Separator, or EOIS. Imagine trying to hear a specific conversation in a crowded room where the background chatter is constantly changing; usually, you would need to record the room for a long time and average out the noise to hear the conversation clearly. This new tool, however, can separate the conversation from the chatter instantly, in a single snapshot, without needing to know anything about the specific person speaking or the specific room. The researchers tested this system on data from the brain's magnetic fields and electrical signals, using visual flashes, sounds, and magnetic pulses. They found that the tool successfully splits the mixed signal into two distinct parts: the part that is strictly locked to the event, and the part that is the brain's own fluctuating, non-locked activity.

The power of this separation lies in how the brain responds. The true reaction to a stimulus is consistent in its timing and shape across many trials, while the background activity is random and shifts unpredictably. The new system uses these differences to pull the two apart. When the researchers applied their method to data from hundreds of people, the results were striking. The separated "evoked" signal showed a strong, consistent pattern that matched the brain's reaction to the stimulus, while the "ongoing" signal remained scattered and random, as expected. Crucially, the tool worked without needing to be calibrated for each individual person. It could take data from a new subject it had never seen before and still separate the signals accurately. This is a significant leap forward, as previous methods required extensive, time-consuming setup for every single person to work correctly.

The researchers also discovered that failing to separate these signals can lead to misleading conclusions. When they looked at the mixed, unseparated data, it appeared that the brain's state before a stimulus strongly predicted the reaction after it. However, once they used the new tool to isolate the true reaction, that connection largely disappeared. The apparent link was actually just the background noise carrying over from before the event into the measurement of the reaction. This finding suggests that many previous studies might have been misinterpreting the brain's natural fluctuations as part of the reaction to a stimulus. By cleaning up the data, the researchers showed that the brain's pre-stimulus state is a much better predictor of its ongoing, non-locked activity than it is of the specific, locked reaction to a stimulus.

Beyond correcting past misunderstandings, this tool opens the door for real-time applications. The system is fast enough to process brain signals as they happen, taking less than a quarter of a second to analyze over a second of data. This speed makes it possible to use in closed-loop systems, where a machine could detect a specific brain state and immediately deliver a treatment, such as a magnetic pulse, tailored to that exact moment. The researchers demonstrated that this separation works across different types of brain recording, including visual and auditory tests, as well as magnetic stimulation. While the system is not perfect—especially when the brain signals are very weak or the time window is extremely short—it represents a robust, generalizable way to see the brain's immediate reactions clearly. By providing a clean view of how the brain responds to the world, moment by moment, this work offers a clearer path for developing better diagnostics and therapies for neurological conditions.

Technical Summary: Disentangling Evoked and Induced Activity from Single-Trial Neurophysiological Time Series

Problem Statement
In neurophysiological research, distinguishing between evoked responses (activity time- and phase-locked to a stimulus) and induced/ongoing activity (non-phase-locked but potentially time-locked or modulated by the event) is critical for interpreting single-trial variability. While trial-averaged data naturally suppresses non-phase-locked activity, single-trial recordings contain a dynamic mixture of these components. Existing methods for separating them often rely on subject-specific calibration or constrained filters, limiting their generalizability and real-time applicability. Furthermore, without temporal separation, pre-stimulus dynamics and ongoing background activity can confound inferences regarding post-stimulus variability, leading to misleading attributions of cortical responsiveness. There is a lack of robust, subject-agnostic methods capable of separating these components in real-time without requiring individual priors.

Methodology: The Evoked–Ongoing/Induced Separator (EOIS)
The authors developed EOIS, a temporal signal separation framework designed to decompose single-trial MEG and EEG data into two distinct components:

Evoked component ( $\hat{e}$ ): Phase-locked to the stimulus.
Ongoing/Induced component ( $\hat{u}$ ): Non-phase-locked activity (including pre-stimulus background and post-stimulus induced activity).

Architecture and Training:

Model Structure: EOIS utilizes a deep learning architecture based on dilated residual blocks (inspired by temporal convolutional networks) to capture multi-scale temporal contexts. The model takes a single-trial time series as input and splits it into two parallel "heads" (evoked and ongoing/induced) that output the separated components.
Subject-Agnostic Approach: The model is trained using a Leave-One-Subject-Out (LOSO) cross-validation strategy, ensuring it generalizes across individuals without subject-specific calibration.
Loss Functions: The training is guided by a weighted sum of feature losses that leverage the distinct statistical properties of the two components:
- Reconstruction Loss: Ensures the sum of the separated components reconstructs the original input signal ( $x \approx \hat{e} + \hat{u}$ ).
- Average Reconstruction: Enforces that the across-trial average of the evoked component matches the average of the original data, while the average of the ongoing/induced component converges to zero.
- Phase Consistency (PPC): Uses Pairwise Phase Consistency (PPC) derived from Morlet wavelet transforms. The evoked component is penalized for low phase consistency, while the ongoing/induced component is penalized for high phase consistency.
- Pre-stimulus Constraint: The ongoing/induced component is encouraged to dominate the pre-stimulus period, as evoked activity should be minimal there.
- Power Constraint: Prevents the sum of the component powers from exceeding the power of the original data, mitigating artifact leakage.

Key Results
The framework was evaluated on visual (VIS) and auditory (AUD) MEG data (CamCAN dataset) and TMS–EEG data (REFTEP dataset), including both calibrated and uncalibrated real-time processing scenarios.

Component Recovery: EOIS successfully separated evoked and ongoing/induced components across all modalities. The evoked component exhibited high phase consistency (PPC $\approx$ 0.66–0.80), while the ongoing/induced component showed near-zero PPC, confirming the effective removal of phase-locking from the latter.
Accuracy: The model achieved high reconstruction accuracy ( $R^2 \approx$ 0.96–0.97 for single trials and averaged evoked responses).
Comparison to Alternatives: EOIS significantly outperformed ablated variants (lacking specific loss constraints), Least-Squares Estimation (LSE) scaling, and Dynamic Time Warping (DTW) of mean templates. LSE and DTW showed lower accuracy and higher leakage between components, particularly because they assume fixed timing and waveform shapes across trials.
Bias Reduction: Logistic regression analysis demonstrated that pre-stimulus features (band powers) were more predictive of post-stimulus ongoing/induced activity than evoked activity. This confirms that analyzing unseparated signals can lead to confounded inferences where pre-stimulus dynamics are misattributed to the evoked response.
Real-Time Feasibility: The processing time was approximately 14 ms for 1 second of data on a CPU, with total real-time pipelines (including preprocessing) taking $<0.25$ s per trial, making it suitable for closed-loop neuromodulation.

Significance and Claims
The paper claims that EOIS provides the first subject-agnostic framework capable of separating evoked and ongoing/induced activity from single-trial neurophysiological data without requiring subject-specific calibration.

Scientific Impact: By isolating these components, EOIS enables more accurate attribution of cortical response variability to specific mechanisms (e.g., distinguishing true evoked responsiveness from modulated background noise).
Translational Utility: The method supports the refinement of biomarkers, pharmacodynamic readouts, and closed-loop neuromodulation protocols. It facilitates the deployment of real-time EEG/TMS-EEG mapping across different individuals and care settings.
Limitations: The authors acknowledge limitations regarding phase-estimation uncertainty at short temporal scales (particularly for sub-15 Hz rhythms) and the lack of ground-truth labels for single-trial separation in real data. They also note that the model assumes a purely additive mixture of components.

In conclusion, EOIS offers a robust, generalizable tool for disentangling overlapping neurophysiological signals, thereby enhancing the interpretability of single-trial dynamics in both research and clinical applications.

Technical Summary: Disentangling Evoked and Induced Activity from Single-Trial Neurophysiological Time Series

More like this