PaperlayerView exact revision on arXiv

Source-linked AI summary

Impact of the reference choice on scalp EEG connectivity estimation

Federico Chella, Vittorio Pizzella, Filippo Zappasodi, Laura Marzetti

arXiv:1812.00794v2q-bio.QM

TL;DR

Because EEG measurements depend on a reference that may not be electrically neutral, reference choice can complicate the interpretation and comparison of connectivity studies. The paper compares four reference schemes using simulated and real EEG data, finding that REST reduces simulated connectivity distortion and that connectivity patterns and graph properties vary with reference choice.

  • Problem

    No electrically neutral body location is available as an EEG reference, and the choice can affect measurements and across-study comparability.

    Source: §1 · pp. 3–4

  • Method

    The study compares Cz, DLM, AVE, and REST using imaginary coherency in simulations and eyes-open resting-state EEG, including electrode-density, head-model, and graph analyses.

    Source: §1 · p. 5, §2.3.4 · p. 12, §2.4.4 · p. 15

  • Results

    Simulations showed REST reduced reference bias, while real-data connectivity patterns varied by reference, with distortion largest for Cz and progressively decreasing for DLM, AVE, and REST.

    Source: §4 · p. 27

  • Takeaways & Limitations

    Reference choice biases interpretations of EEG connectivity patterns and graph-derived network topology, so comparisons across different reference schemes warrant caution.

    Source: §5 · p. 29

    For real EEG, potentials referenced to infinity were unavailable, so reference effects were evaluated against REST using a subject-specific realistic head model.

    Source: §3.2.1 · p. 22

Abstract

from arXiv · show

Several scalp EEG functional connectivity studies, mostly clinical, seem to overlook the reference electrode impact. The subsequent interpretation of brain connectivity is thus often biased by the choice a non-neutral reference. This study aims at systematically investigating these effects. As EEG reference, we examined: the vertex electrode (Cz), the digitally linked mastoids (DLM), the average reference (AVE), and the Reference Electrode Standardization Technique (REST). As a connectivity metric, we used the imaginary part of coherency. We tested simulated and real data (eyes open resting state), by evaluating the influence of electrode density, effect of head model accuracy in the REST transformation, and impact on the characterization of the topology of functional networks from graph analysis. Simulations demonstrated that REST significantly reduced the distortion of connectivity patterns when compared to AVE, Cz and DLM references. Moreover, the availability of high-density EEG systems and an accurate knowledge of the head model are crucial elements to improve REST performance. For real data, a systematic change of the spatial pattern of functional connectivity depending on the chosen reference was also observed. The distortion of connectivity patterns was larger for the Cz reference, and progressively decreases when using the DLM, the AVE, the REST. Strikingly, we also showed that network attributes derived from graph analysis, i.e., node degree and local efficiency, are significantly influenced by the EEG reference choice. Overall, this study highlights that significant differences arise in scalp EEG functional connectivity and graph network properties, in dependence of the chosen reference. We hope our study will convey the message that caution should be taken when interpreting and comparing results obtained from different laboratories when using different reference schemes.

1. Introduction

EEG recordings require a reference, but no body location is electrically neutral, so the choice affects measurements and complicates comparisons across studies. This paper investigates how reference choice affects connectivity estimates and graph-based network characterization using simulated and real data.

1. Introduction

  • The reference choice affects EEG measurements because no body location is electrically neutral, and differing schemes impede across-study comparability.The average reference and REST aim to provide relatively neutral references, but can be limited by electrode density, scalp coverage, and, for REST, head-model accuracy.Source: §1 · pp. 3–4
  • Non-neutral reference activity can distort connectivity estimates by creating spurious connections or suppressing existing ones.Prior studies have reported reference-related effects for correlation, coherence, phase coherency, and imaginary coherency.Source: §1 · p. 4
  • Reference-dependent connectivity patterns may lead researchers to different interpretations, while graph-derived network properties may also vary with the reference.Prior graph-theory evidence cited here found changes in network pattern and weighted density across references.Source: §1 · p. 5
  • The study quantitatively compares Cz, DLM, AVE, and REST using imaginary coherency in simulations and real EEG, testing electrode density, head-model accuracy, and graph-network topology.Simulations also compare potential effects with prior studies and include an idealized spherical head model.Source: §1 · p. 5

2.1. The EEG references

The section introduces four reference schemes and explains how each transforms measured EEG potentials. REST estimates potentials referenced to infinity using a transformation constructed from transfer matrices and a head model.

2.1. The EEG references.

2.1.1. Cephalic electrode reference to Cz.

  • Cz re-referencing subtracts the potential measured at Cz from every channel at each time sample.The study uses Cz as its cephalic reference; this transformation can be applied offline to data acquired with another reference.Source: §2.1.1 · p. 6

2.1.2. Digitally linked mastoids reference.

  • DLM re-referencing subtracts the average of the left and right mastoid potentials from every channel at each time sample.The transformation assigns weight 0.5 to each mastoid electrode.Source: §2.1.2 · p. 6

2.1.3. Average reference.

  • AVE re-referencing subtracts the average potential across all electrodes from every channel at each time sample.The corresponding matrix has entries equal to 1/N.Source: §2.1.3 · pp. 6–7

2.1.4. Reference Electrode Standardization Technique.

  • REST uses transfer matrices to estimate EEG potentials referenced to infinity, without explicitly solving for the actual neural sources.Its transformation depends on the head model, electrode montage, original reference, and assumed equivalent source distribution.Source: §2.1.4 · p. 7

2.2. Connectivity estimation by imaginary part of coherency

The paper uses imaginary coherency to quantify frequency-specific phase coupling between EEG signals. This measure is robust to instantaneous volume-conduction mixing, but it cannot detect perfectly synchronous sources or separate changes in coupling magnitude from phase relationship.

2.2. Connectivity estimation by imaginary part of coherency

  • Imaginary coherency is used to measure functional connectivity between EEG signals and reflects phase coupling at a specific frequency.The study estimates its imaginary part between pairs of electrodes.Source: §2.2 · pp. 7–8
  • The imaginary part of coherency is robust to instantaneous volume-conduction mixing, but captures only time-lagged processes.Perfectly synchronous sources do not contribute to the imaginary part.Source: §2.2 · p. 8
  • Imaginary coherency alone cannot distinguish a change in coherency magnitude from a change in phase relationship.Reconstructing both properties requires the real and imaginary parts of complex-valued coherency.Source: §2.2 · p. 8

2.3. Simulations

The simulations compared EEG potentials and imaginary coherency under several reference schemes against an infinity-referenced gold standard. They varied electrode density and the head model used for REST, then assessed errors across repeated simulated datasets.

2.3. Simulations

  • REST and conventional references were compared with infinity-referenced potentials and imaginary coherency, treated as the gold standard.The study evaluated Cz, DLM, AVE, and four REST head-model conditions.Source: §2.3 · p. 9, §2.3.2 · p. 11, §2.3.3 · pp. 11–12

2.3.2. Generation of simulated EEG recordings.

  • Simulated recordings used ten realistic head shapes with brain, skull, and scalp compartments and a cortical source grid sampled at 5 mm.Recordings were generated for five minutes at 500 Hz using two interacting sources, four uncorrelated noise sources, and added instrumentation noise.Source: §2.3.1 · p. 9, §2.3.2 · pp. 9–10

2.3.1. Head model and EEG electrodes.

  • Four nested layouts used 21, 34, 74, or 128 electrodes, allowing electrode density to vary while maintaining scalp coverage.Each denser subset included the sparser subsets, so observed differences could be attributed to density rather than coverage.Source: §2.3.1 · p. 9, Figure 1 · p. 10

2.3.3. Re-referencing the simulated data.

  • REST was applied to Cz-referenced recordings using spherical, standard, perturbed individual, or exact individual head models.The perturbed model shifted mesh vertices by 3 mm; the exact model matched the model used to generate the simulated data.Source: §2.3.3 · pp. 11–12

2.3.4. Coherency analysis.

  • Imaginary coherency at 10 Hz was estimated between each electrode pair from one-second non-overlapping, Hanning-windowed segments.Each reference condition produced an N × N matrix of pairwise values.Source: §2.3.4 · p. 12

2.3.5. Performance criteria.

  • Relative errors against infinity-referenced potentials and coherency were compared across 1,000 simulation repetitions using paired Wilcoxon signed-rank tests.The repetitions randomized source locations and orientations across ten head shapes; statistical significance was set at p < 0.05.Source: §2.3.5 · pp. 12–13, §2.3.6 · p. 13

2.4. Application to real EEG recordings

The real-data analysis used eyes-open resting-state EEG to assess reference-related changes in connectivity patterns and graph attributes. Connectivity was estimated from imaginary coherency, and node degree and local efficiency were compared across reference schemes.

2.4.1. Data acquisition and preprocessing.

  • Ten healthy adults completed ten minutes of eyes-open resting-state EEG with a 128-sensor system recorded at Cz; preprocessing yielded 110 channels for analysis.Artifacts were screened and removed, and missing channels were interpolated where necessary.Source: §2.4.1 · pp. 13–14

2.4.2. Re-referencing the EEG recordings.

  • DLM, AVE, and REST datasets were derived from the Cz-referenced recordings, with REST using spherical, standard, or individual-MRI head models.The real-data analysis used an individual MRI model presumed equivalent to the perturbed model; an exact model was available only in simulations.Source: §2.4.2 · p. 14

2.4.3. Coherency analysis.

  • Imaginary coherency was estimated for electrode pairs in the alpha band, and group matrices were compared using dissimilarity 1 − r².For each subject and reference, analysis focused on the individual alpha-band frequency of maximal imaginary coherency.Source: §2.4.3 · pp. 14–15

2.4.4. Computation of network properties.

  • Connectivity matrices were converted to binary graphs by retaining the 25% strongest positive and negative connections.These unweighted, undirected graphs represented electrodes as nodes and connections as binary links.Source: §2.4.4 · p. 15
  • The analysis compared node degree and local efficiency across references, using paired t-tests with significance assessed by 10,000 label permutations.Degree counts a node’s connections; local efficiency measures communication efficiency among its neighbors when the node is removed.Source: §2.4.4 · pp. 15–16

3. Results

Simulations and real eyes-open recordings show that EEG reference choice changes connectivity estimates and graph-network properties. REST generally reduced simulation bias, while real-data connectivity and network measures varied systematically across references.

3.1.1. Potentials.

  • With an accurate head model, REST real exact reduced potential RE to ≤0.05 on average, reaching 0.02 with 128 electrodes.Cz averaged approximately 0.813 and DLM approximately 0.295; AVE performed better than both, while inaccurate head models weakened REST’s advantage.Source: §3.1.1 · p. 17, Figure 2 · p. 17
  • At SNR=10, added instrumentation noise generally increased potential RE, but REST remained best except that spherical-model REST was similar to AVE.The relative ordering was similar under white Gaussian and iso-spectral noise.Source: §3.1.1 · p. 18, Figure 3 · p. 19

3.1.2. Coherency maps.

  • Across 1,000 simulations, REST reduced imaginary-coherency RE more than other references; Cz and DLM had the largest errors, and iso-spectral noise slightly raised REST error.In an example map, REST patterns resembled the reference-at-infinity map, while the largest visual differences occurred for Cz and DLM.Source: §3.1.2 · p. 20, Figure 4 · p. 21

3.2.1. Coherency maps.

  • In real recordings, connectome dissimilarity from REST real was 0.80 for Cz, 0.33 for DLM, 0.26 for AVE, 0.06 for REST spherical, and 0.01 for REST standard.The authors treat REST real as an approximation to the reference at infinity; AVE maps resembled REST real, unlike Cz maps.Source: §3.2.1 · p. 22, Figure 5 · p. 23

3.2.2. Systematic differences in Node Degree and Local Efficiency.

  • Node degree differed significantly by reference, with spatially specific patterns: differences were more widespread for Cz and DLM than for AVE or REST variants.REST real and REST standard showed higher connectivity over central and occipital electrodes; AVE and REST spherical showed an occipital increase.Source: §3.2.2 · pp. 24–25, Figure 6 · p. 24
  • Local efficiency also showed significant, spatially specific differences across EEG referencing conditions.Source: §3.2.2 · p. 25, Figure 7 · p. 25

4. Discussion

The discussion compares four EEG reference schemes in simulations and real data, finding reference-dependent connectivity patterns and graph properties. It also considers reference-free alternatives, each with its own trade-offs.

4. Discussion

  • Cz substantially altered potentials and imaginary coherency; DLM performed better than Cz but produced greater distortions than AVE and REST.The discussion attributes Cz distortion to non-neutral activity at the reference site and notes that added sensor noise enhanced potential distortion.Source: §4 · p. 26
  • AVE retained bias, and relative error for potentials and imaginary coherency increased with sensor density when scalp coverage remained limited to the upper head.Adding noise did not change AVE performance, particularly for denser arrays.Source: §4 · p. 26
  • REST reduced bias versus other references across 21–128 electrodes and head-model accuracies, with high density and accurate anatomy improving its performance.REST remained better than AVE with noise except when a nonrealistic three-shell spherical model was used; realistic individual anatomy was preferable, especially at high density.Source: §4 · pp. 26–27
  • REST was sensitive to sensor noise because its source-distribution assumption excludes instrumental noise, reducing standardization accuracy.Despite this sensitivity, it generally outperformed AVE under noise, except with a nonrealistic three-shell spherical model.Source: §4 · p. 27
  • In eyes-open resting EEG, connectivity distortion decreased from Cz to DLM, AVE, REST spherical, and REST standard, while node degree and local efficiency depended significantly on reference choice.The authors caution that findings from different reference schemes are not interchangeable.Source: §4 · p. 27

4.1. General comments on reference-free approaches

  • Surface Laplacian avoids reference bias by using spatial second derivatives, but its high-pass spatial filtering suppresses deep and distributed activity.It therefore emphasizes shallow, localized sources.Source: §4.1 · p. 28
  • Bipolar recordings remove the shared reference contribution and can help remove artifacts or identify local activations, but suppress large, distributed activations.The suppression follows from the spatial-derivative effect of the recording scheme.Source: §4.1 · p. 28
  • Source-level connectivity avoids reference effects on inverse localization for noiseless potentials, but depends on source-reconstruction accuracy, for which high-density EEG is preferred.The discussion leaves the advantages and limitations of this approach beyond the paper’s scope.Source: §4.1 · p. 28

5. Conclusions

The conclusion is that EEG reference choice changes connectivity topography and graph-derived network properties, affecting interpretation of brain interactions. The authors recommend REST to reduce these effects and facilitate comparisons across laboratories and databases.

5. Conclusions

  • Different references significantly alter EEG connectivity topography and bias interpretation of brain interactions and graph-derived network topology.The authors report that the findings from imaginary coherency can generalize to other temporal or spectral connectivity metrics, including dynamic functional connectivity.Source: §5 · p. 29
Loading 1812.00794v2…