The inclusion criteria were human studies, covering all primary studies in English (retrospective, prospective, randomized trials, etc.), with any mention of CCEPs either in the title, abstract or full text. Studies of any disease type, population, patient age and sex were included. Literature reviews, systematic reviews, meta-analyses, guidelines, letters to the editor, commentaries, or trial protocols were excluded.

The Medline, Embase, and Web of Science databases were searched using the following keywords: (“cortico cortical evoked potential” or “CCEP” or “single pulse electrical stimulation”). Searches were conducted from database inception through July 11, 2024, using the terms. Full search syntax for each database is provided in Supplementary Material 1. We also performed a backward and forward citation analysis and used each database’s “similar articles” feature.

After deduplication using Covidence, three reviewers (TA, AW, MW) independently screened titles and abstracts, followed by full-text review. Discrepancies were resolved by consensus. A PRISMA flow diagram is provided in Figure 1. Studies with no explicit mention of methodology for CCEP quantification were excluded. Also, studies employing high frequency stimulation were excluded, since CCEPs are evoked using low frequency stimulation (Matsumoto et al., 2004). Excluded articles and reasons for exclusion are listed in Supplementary Material 2.

Data collected from studies included recording type (sEEG or ECoG), montage, recording system, stimulation amplitude and frequency, epoch limits, whether the data used was open access, and the detection method. Recording type was categorized based on the terminology used by the original studies, and no reclassification was applied based on surgical technique or electrode geometry. Demographic data like study year, sample size, patient sex, age, and disease type were collected. All data was manually extracted by authors, and any uncertainty or discrepancy was resolved via consensus. The data extracted for each included study is available in Supplementary Material 3.

Although this review focused on methodological reporting rather than outcomes, we qualitatively assessed the completeness of methodological descriptions. Studies lacking essential reporting were excluded to maintain interpretive granularity.

In this section, we describe some common practices in CCEP pre-processing. A stimulation trial involves recording the CCEP response on one electrode following the delivery of an electrical stimulus through another electrode. The steps preceding CCEP detection usually involve the epoching of the continuous data into the pre- and post-stimulation data relative to the stimulation onset. The stimulation artifact, which lasts around 1–6 ms, is usually excluded from the analysis by discarding, interpolating, or using more sophisticated filters (e.g. Wiener filters) to mitigate its effect (Prime et al., 2018). The data then undergoes normalization relative to the pre-stimulus data across trials and re-referencing followed by filtering for noise reduction (Levinson et al., 2024).

Cortico-cortical evoked potential detection methodologies were grouped into four categories: (1) peak detection and thresholding (Mouthaan et al., 2016), which identify a characteristic peak of CCEPs and compare its amplitude to a threshold; (2) statistical tests to classify CCEPs by identifying significant differences between the post- and pre-stimulation periods (Dou et al., 2023); (3) frequency-based methods that detect frequency changes or spectral markers (van den Boom et al., 2024); (4) data-driven methods use the data as the basis to determine the shapes of the evoked responses rather than relying on pre-defined assumptions (Miller et al., 2023).

Figure 2 illustrates the typical workflow for CCEP pre-processing and detection. The raw data in Figure 2A is first segmented into pre- (−1500 to −5 ms) and post-stimulation (5–1500 ms) excluding the −5 to 5 ms window to remove stimulation artifacts. The data is then filtered, re-referenced, and normalized, resulting in Figure 2B. Figure 2C highlights the heterogeneity of CCEP detection methods by showcasing four approaches. In the first, a CCEP is identified when the average post-stimulation data exceeds 6 times the pre-stimulation standard deviation (van den Boom et al., 2024). In the second method, the root-mean-square (RMS) of the pre- and post-stimulation data is compared using a Wilcoxon rank-sum test (Boulogne et al., 2022; Almashaikhi et al., 2014). For the frequency-based method, the 20–200 ms response is filtered between 10 and 30 Hz, and 20 Hz power is computed using wavelet transform and compared to the mean overall power for significance (van den Boom et al., 2024). Finally, canonical response parametrization (CRP), a data-driven method for CCEP detection, projects the first trial onto the others to obtain projection magnitudes. Trials are classified as CCEPs based on the significance of these magnitudes at the response duration, with canonical shapes visualized (Miller et al., 2023). Further details on CRP are provided in Section “3.3.4 Data-driven methods.”

Interest in CCEPs has grown, as reflected by the rising number of related studies (Figures 3A, B). A total of 2,855 studies were screened by title and abstract, with 187 studies undergoing full text review (Figure 1). Of the 187 studies, 9.1% (n = 17) utilized visual identification, while 49.7% (n = 93) did not explicitly mention the detection method. After excluding these studies, 72 remained for review, representing 3,424 patients with a mean age of 30.3 ± 8.3 years (mean ± 1 SD) and an age range of 3–69 years, of whom 1,652 were female (48.2%). Across the 72 studies, CCEPs were recorded with sEEG (n = 42, 58.3%), ECoG (n = 22, 30.6%), or both (n = 8, 11.1%). Most studies focused on patients with epilepsy (n = 67 studies, 93.1%), with some including those with brain tumors (n = 2, 2.8%), or both (n = 3, 4.1%) (Table 1). The stimulation protocols in these studies utilized a mean frequency of 0.7 ± 0.4 Hz (range = 0.06–2 Hz) and amplitude of 5.8 ± 3 mA (range = 0.35–15 mA) to elicit CCEPs. Only 13 studies published their data (comprising 18.1% of papers with a detailed detection method description), with an additional 3 papers sharing their analysis code but not the raw data.

The reviewed papers fell into one of four categories based on the methods used for CCEP detection. Some papers employed a mix of these methods or outline multiple implementations. We observe that before 2011 most methods were either not explicitly described or rely on visual inspection. The most common methods (n = 49, 68.1%) involved different thresholding or peak detection techniques to identify CCEPs. The second most common method (n = 12, 16.7%) employs statistical tests to determine the prominence of the responses. Data-driven methods are employed in some papers (n = 3, 4.1%), while others (n = 3, 4.1%) use frequency-based methods. The remaining papers (n = 5, 6.9%) includes a combination of methods. A summary of the different subcategories of each detection method and the corresponding studies that utilize them are presented in Table 2.

The bipolar montage (n = 35, 48.6%) was the most common, followed by several papers where the montage was not explicitly stated (n = 13, 18.1%) and the common average reference (CAR) (n = 4, 5.6%). Some used single-electrode reference with extracranial placements including the skin above the mastoid bone or midline scalp, and intracranial placement using a white matter electrode contact (n = 13, 18.1%). Realizing that bipolar re-referencing affects the polarity of the recording and CAR can introduce bias when many electrodes show an elevated response, some opted for an adjusted CAR (n = 5, 6.8%), where 20% of channels with lowest variance (Valencia et al., 2023; van den Boom et al., 2024) or the mean of the unstimulated channels are utilized (Huang et al., 2023; Miller et al., 2023). Another paper (Campbell et al., 2023) utilized a common median reference (n = 1, 1.4%) that is robust to extreme values introduced by the stimulation artifact (Rolston et al., 2009). One paper utilized both the bipolar and referential montage for comparison (Shahabi et al., 2021).

Although visual identification was not categorized separately in this review, as our emphasis is on rigorous and reproducible methods, we acknowledge its ease of use in low-resource settings. However, the subjectivity and inter-rater variability inherent in visual analysis limit its reliability and generalizability (van den Boom et al., 2024).

The threshold or peak-based detection methods are widely used due to their simplicity and ease. However, these methods rely on assumptions about the waveform of the CCEP, ignoring the vast heterogeneity in CCEPs, especially in sEEG data (Prime et al., 2018). There is variability in the chosen thresholds and low ones may disregard amplitude variations that are evident in CCEPs leading to an increase in false negatives (Frauscher et al., 2018). Statistical approaches improve upon thresholding by offering a more rigorous method to evaluate post-stimulation deviation from baseline. Techniques such as permutation testing, t-tests, and cluster-based inference increase robustness to noise and reduce subjective bias. Nonetheless, statistical tests may still assume pre-defined CCEP shapes and are sensitive to design choices like test type (parametric vs. non-parametric) and correction for multiple comparisons. Moreover, some assumptions required for these tests like stationarity, normality and linearity in data may not hold for complex neural data (Sugimoto et al., 1977). Frequency-based methods provide valuable insight into the spectral dynamics of CCEP responses, including phase locking and power changes across frequencies. They can reveal neural responses not easily observable in the time domain. However, these methods are sensitive to pre-processing choices, lack standardization, and have introduced conflicting findings when considered in studies extracting spectral biomarkers of epileptogenic zones (Qiang et al., 2025). Data-driven methods like CRP represent a shift toward a more standardized evaluation of CCEPs, since it learns from the data directly without assuming response morphology, accommodating CCEP heterogeneity (Miller et al., 2023). Ultimately, the detection method affects what is deemed significant and thus affects clinical decisions such as surgical planning or language mapping or EZ zone localization.

One important factor is the effect of the montage on the processing of CCEP data, since it can introduce bias from the stimulated channels and affect the noise level across electrodes. A recent study compared various montages used to re-reference CCEP data recorded using sEEG to examine which montage better localizes electrodes in gray matter (Dickey et al., 2022). The RMS deviation from baseline was obtained after applying either the common average, bipolar, Laplacian montage, or referential montage using a subgaleal electrode placed near the vertex of the head. The Laplacian and bipolar montages were better at localizing CCEPs in gray matter than the CAR or referential montage as measured by the area under the curve for a receiver operating characteristic (Dickey et al., 2022). Another study including 14 patients with sEEG demonstrated that bipolar and Laplacian montages were able to effectively reduce SPES-related signal deflections at extracortical locations, including outside the brain (Mitsuhashi et al., 2020). The bipolar and Laplacian montages help preserve focal signal features and attenuate more distributed signals that are visible in neighboring electrodes like high-amplitude CCEP deflections (Arnulfo et al., 2015). Other montages that have been used are the common median or the adjusted CAR. If the distributed signals that arise from CCEPs are of interest, local re-referencing methods may not be the optimal choice because they may attenuate the signal, distort temporal profiles and lead to phase reversals (Arnulfo et al., 2015; Huang et al., 2024). Thus, the choice of re-referencing montage should be motivated by the signal features of interest, since different montages influence which signals features are preserved.

Some papers employed an adjusted version of the CAR that approximates the common noise while also minimizing the introduced bias from responsive channels by only including 20% of channels with lowest variance or the mean of unstimulated channels. A recent study introduced CAR by least anticorrelation “CARLA” (Huang et al., 2024), an algorithm that first ranks channels based on increasing covariance across trials, where lower covariance channels are considered less responsive. Channels were then iteratively added to the common average starting with the ones with the lowest covariance, and an anti-correlation statistic was calculated at each step between the common average formed by this subset and the re-referenced signals of the remaining channels. The algorithm sought to optimize the selection of the channels to be included in the common average by minimizing the anti-correlation.

We found that the most common detection methods rely on assumptions about the morphology of the CCEP responses. This can make the analysis process simple and straightforward, but it disregards the variability that is evident in the evoked responses. These methods assume a single morphology for the CCEP response where a sharp deflection called the “N1” peak is expected around 10–30 ms after the stimulus onset and is sometimes followed by a second slower deflection “N2” peak occurring around 80–250 ms from stimulus (Keller et al., 2014). These deflections vary in latency and polarity and are mostly a characteristic of responses recorded using ECoG (Prime et al., 2020), which represent 31.5% of the included studies in this review.

Additionally, we identified differences in the quantification of CCEPs recorded from ECoG compared to sEEG. When using ECoG, a certain morphology, including N1 and N2 peaks, is expected since pyramidal cells are oriented consistently toward the electrode (Prime et al., 2018). However, in sEEG recordings, the depth electrodes sample various anatomical structures and can be located in gray matter, white matter and sometimes cerebrospinal fluid (Prime et al., 2020). This variation in the brain regions sampled and orientation of the cortical layers relative to the electrode likely leads to variation in the observed CCEPs. Indeed, signal amplitude variations have been reported between different anatomical regions during sEEG recordings, along with variations in spectral density distributions (Frauscher et al., 2018). Many studies reported the presence of a wide variety of morphologies and polarities in CCEPs recorded (Prime et al., 2018; Enatsu et al., 2015; Valencia et al., 2023; Miller et al., 2021, 2023; Shine et al., 2017). CCEPs have been used to characterize brain connectivity by quantifying the responses using the strength of the N1 peak or its latency, since it has been tied to direct pathways. However, this approach disregards other response morphologies that are due to indirect propagation or recurrent cortico-subcortico-cortical pathways (Huang et al., 2023; Veit et al., 2021). Thus, it is important to not categorize CCEPs into one expected morphology, since pre-defined forms would fail to capture the significance of various responses and their morphology even after various types of manipulations like temporal scaling or sign-flips (Miller et al., 2023). CCEP recordings from sEEG have been reported to be polymorphic including positive and negative deflections with around three peaks described, the first occurring 10–30 ms after stimulation, the second after 80–250 ms and the third after that (Feys et al., 2024b). The variability in CCEP morphology has also been linked to EZ localization biomarkers.

To properly validate the detected CCEPs, it is recommended to utilize a proper referencing montage that minimizes the noise level and bias from stimulated channels, while also considering the anatomical location of the recording electrode primarily in gray matter, as opposed to white matter or CSF. Ensuring across trial reproducibility and dependence on stimulation parameters like amplitude, which has been used to detect significant responses, are all methods that can help decrease misclassification of CCEPs. The heterogeneity observed in the detection methods highlights a clear need for a standardized way to evaluate CCEP responses that is quantitative, automated, reproducible, and can accurately convey the connectivity from the recorded data. One step closer to this standardized method may be the data-driven method described that provides a structured approach to parametrizing CCEP data by inferring the canonical shape from the trials and defining metrics like the response duration that can be used for comparisons across brain regions or patients (Miller et al., 2023). CRP also introduced a method for automated rejection of artifactual trials based on the cross-projection magnitudes. Moreover, CRP can adapt to various modalities (EEG, fMRI, etc.) making the analysis of data from multimodal studies simpler through the integration of electrophysiological and imaging modalities. This consistent response quantification allows for robust connectivity analysis that can be used to understand connectivity patterns in focal epilepsy. It also offers a way to study brain state transitions during anesthesia by identifying response shifts (Yamao et al., 2021). CRP may also be used in analyzing event-related potentials by identifying these structures and their significances in an automated manner with applications in the study of cognition, emotions, sensory functions and brain injuries (Mazzini, 2004).