Neural Complexity Signatures of Visual Mental Imagery: Lempel-Ziv Complexity and Higuchi Fractal Dimension Reveal Topographic Reorganization in High-Density EEG

Neural Complexity Signatures of Visual Mental Imagery: Lempel-Ziv Complexity and Higuchi Fractal Dimension Reveal Topographic Reorganization in High-Density EEG | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Neural Complexity Signatures of Visual Mental Imagery: Lempel-Ziv Complexity and Higuchi Fractal Dimension Reveal Topographic Reorganization in High-Density EEG Yu Gao, José Miguel Diniz This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9382178/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Background: Visual mental imagery is the process of reconstructing perceptual experience without sensory input. How the brain performs this process is poorly understood, particularly from the perspective of conventional linear EEG analysis. Objective: This study aims to evaluate if the two non-linear EEG complexity measures—Lempel-Ziv Complexity (LZC) and Higuchi Fractal Dimension (HFD)—can differentiate between perception and imagination and if they can be used as objective indices of neural separability of mental imagery. Methods: LZC and HFD were extracted from 58 scalp EEG channels in 46 healthy adults performing the PerceiveImagine paradigm (Li and Fan 2024), after Wideband Picard ICA (1–200 Hz) removing artefacts. Statistical analyses included cluster-based permutation testing (Maris and Oostenveld 2007), Hotelling T², and leave-one-subject-out cross-validation (LOSO-CV). Results: Broadband LZC topography differed between perception and imagination (cluster p = 0.005; Hotelling F = 3.08, p = 0.002, V = 0.28). LOSO-CV classification reached AUC = 0.811 (95% CI: 0.775–0.847). The classifier’s AUC exceeded a label-permuted baseline by a wide margin (t(45) = 16.30, p = 8.35 × 10⁻²¹, BF₁₀ = 8.84 × 10¹⁴). No gamma-band LZC differences were found (p = 1.0, d = −0.13). Significance: EEG complexity features reveal a fundamental topographic redistribution principle of cortical dynamics: imagery is characterized by a shift from sensory-driven posterior irregularity to generative frontal complexity, rather than a global magnitude change. Objective decoding labels outperform self-report as a training signal for imagination quality classifiers. EEG mental imagery biomarker Lempel-Ziv complexity Higuchi fractal dimension predictive coding brain–computer interface Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Visual mental imagery refers to the ability to internally reconstruct perceptual experience without external sensory input. When a person imagines an apple, primary visual cortex activates in proportion to the vividness of that image, despite the absence of any retinal signal (Dijkstra et al. 2021 ). Understanding how the brain accomplishes this reconstruction has practical relevance for brain–computer interfaces, neurofeedback, and cognitive rehabilitation. Most EEG studies still focus on linear spectral qualities, such as band power, which likely overlooks the irregularity of the endogenously determined neural processes. Two non-linear metrics, however, provide a different perspective. Lempel-Ziv Complexity (LZC) measures the number of distinct sequential patterns in a binarized signal: the more patterns, the greater the signal's temporal irregularity. Higuchi Fractal Dimension (HFD) determines the scaling of a time series' curve length across various time resolutions, producing a Fractal Dimension that indicates geometric roughness. LZC and HFD capture orthogonal facets of neural dynamics. LZC represents the algorithmic complexity (symbolic novelty) of the binarized signal, while HFD quantifies its geometric scaling (fractal roughness) across time resolutions. This dual-metric approach provides a more comprehensive characterization of non-stationary EEG dynamics than either index alone. Recent studies have confirmed that the two metrics are responsive to the different brain states. Höhn et al. ( 2024 ) showed that broadband LZC and spectral slope together monitor sleep/wake states in healthy adults. Medel et al. ( 2023 ) found that LZC and the 1/f spectral slope provide a complementary description of the excitation-inhibition balance in the cortex. Ren et al. ( 2023 ) applied LZC to resting state EEG, and were able to identify cognitive impairment in epilepsy patients. On the HFD side, Aggarwal and Ray ( 2025 ) reported that HFD in scalp EEG is anticorrelated with oscillatory power and 1/f slope across the adult lifespan. Colussi et al. ( 2025 ) used HFD to differentiate wakefulness from sleep in a developmental cohort. Armonaite et al. ( 2023 ) validated HFD as a region-specific cortical complexity signature, and Ruiz de Miras et al. ( 2023 ) applied fractal dimension analysis to resting-state EEG networks in schizophrenia. Despite this growing body of work, no study has combined LZC and HFD as joint biomarkers for mental imagery decodability. The imagery literature itself has an unresolved measurement problem. Dijkstra et al. ( 2021 ) found that perception and imagery share neural substrates in visual cortex but differ in temporal dynamics—imagery recruits top-down feedback more slowly and with greater trial-to-trial variability. Chang et al. ( 2025 ) released a multisensory EEG imagery dataset with vividness ratings, confirming that subjective vividness correlates with distinguishable EEG power patterns. Wilson et al. ( 2024 ) reviewed the feasibility of visual decoding from EEG. Shimizu and Srinivasan ( 2022 ) improved classification of imagined images using spectral features. Gifford et al. ( 2022 ) provided a large benchmark EEG dataset for visual object recognition, and Lee et al. ( 2022 ) explored the classification of perception versus imagery. The common thread across these studies is a reliance on subjective vividness ratings as the only ground truth for imagery decodability, even though self-report is noisy and inconsistent across sessions. An objective approach is to classify perceptual and imagined states, and use the classifier’s decoding confidence as a continuous quality measure. We incorporated this dual-labeling method with our broad-band artefact-control processing pipeline. The gamma-band range (30-150Hz) is inseparable from the scalp muscle artefact, which, if unchecked, can confound complexity differences. Instead of removing the high frequencies, we applied Picard ICA (Ablin et al. 2018 ) over the entire 1-200 Hz range; and used EMG power monitoring as a secondary protective measure. We evaluated eight hypotheses related to 46 participants in the PerceiveImagine paradigm (Li and Fan 2024 ) and extracted 580 complexity features per trial (406 LZC + 174 HFD) from 58 scalp EEG channels. The gamma-band LZC elevation we predicted did not occur. Instead, we observed a significant topographic redistribution of broadband complexity between perception and imagination. This finding is in line with predictive coding frameworks suggesting imagery redistributes the cortical dynamics, rather than elevating their magnitude uniformly. This study makes three contributions. Firstly, it establishes that EEG complexity features capture a specific reallocation of symbolic and geometric complexity, reflecting the reversal of cortical information flow as predicted by top-down generative models. Secondly, it highlights that decoding-derived confidence surpasses chance level baselines with strong Bayesian evidence (BF 10 > 10 14 ), thus qualifying it as a valid objective measure of imagery decodability. Third, it validates a reproducible pipeline for shuffle-normalized LZC and sliding-window HFD extraction on a 46-subject public dataset. Code is available from the corresponding author upon reasonable request. We hypothesized that mental imagery would result in varying complexity signatures in the low-gamma band (30–60 Hz). This a priori choice is based on theory regarding the role of gamma-band neural synchrony in endogenous neural binding and top-down sensory reconstruction. By isolating this band and employing a rigorous wideband ICA artifact-control pipeline, we aimed to determine whether reported gamma-band complexity changes reflect genuine neural dynamics or are susceptible to scalp muscle contamination. 2. Materials and Methods 2.1 Dataset and Participants We analyzed the PerceiveImagine dataset (OpenNeuro ds005697) (Li and Fan 2024 ), which was collected from 54 healthy adults at a university laboratory in China. Two participants withdrew during recording. Of the remaining 52, eight were excluded before analysis: seven because of missing or corrupted data files (sub-07, sub-12, sub-13, sub-39, sub-51, sub-52, sub-53, as documented in the dataset release notes) and one (sub-38) because its behavioural log contained no perception or imagination event codes, so that zero valid epochs could be extracted. The final analytic sample consisted of 46 participants, aged 23 to 30, both male and female, with no reported neurological or psychiatric history and normal or corrected-to-normal vision. 2.2 Data Acquisition EEG was recorded at 1000 Hz with a 64-channel Neuroscan SynAmps2 system in a standard 10–20 electrode layout. After removal of auxiliary channels (HEO, VEO, EKG, EMG, CB1, CB2), 58 scalp EEG channels were retained for analysis. Each experimental trial comprised a fixation cross (2 s), a visual stimulus presentation (6 s), a brief gap (0.5 s), a mental imagination period (6 s), and a vividness self-report. The stimulus set contained 340 naturalistic images. The end-to-end preprocessing and feature extraction pipeline is illustrated in Fig. 1 . 2.3 Preprocessing All preprocessing was carried out in MNE-Python (≥ 1.6). The continuous data were bandpass-filtered at 1–200 Hz using a zero-phase FIR filter and notch-filtered at 50, 100, and 150 Hz to remove power-line harmonics. A common average reference was applied. We then fitted Picard ICA (Ablin et al. 2018 ) on the wideband signal with extended decorrelation enabled (random seed = 42; number of components set to the data rank). ICLabel classified each component automatically; those labelled “muscle” with probability exceeding 0.6 or “eye” were excluded, following the default ICLabel threshold for automated pipelines. Across subjects, a mean of 3.2 ± 1.8 ICA components were removed (range: 1–7). The algorithm converged for all 46 participants (convergence delta < 10⁻⁴). As a second layer of artefact control, we computed the mean high-gamma band power (80–150 Hz) at six edge channels (T7, T8, TP7, TP8, PO7, PO8) for each imagination epoch via Welch periodogram. Any epoch whose edge-channel power exceeded 3.0 standard deviations above the per-subject mean was flagged, following standard artefact rejection practice. Fewer than 5% of imagination epochs were flagged across the cohort. Flagged epochs were removed only from high-gamma analyses; broadband and sub-gamma computations retained them. 2.4 Epoching Event information was parsed exclusively from the behavioural log files (events.tsv) rather than from hardware trigger channels, because some recordings contained spurious trigger pulses not present in the behavioural logs. A three-tier cascading classifier assigned event labels to conditions: keyword matching was tried first, then numeric prefix rules (labels starting with “1” mapped to perception, “2” to imagination), and finally an ordinal fallback for any remaining unmatched labels. Epochs were extracted with a − 0.2 to 0.0 s baseline and a post-onset duration of 6.0 s. The resulting master feature table contained 30,534 trials across all 46 subjects, with approximately 331 perception trials and 332 imagination trials per subject, corresponding to the 340-image stimulus set minus occasional parsing losses. 2.5 Lempel-Ziv Complexity LZC measures a signal’s irregularity in time by evaluating the number of unique sub-patterns formed in a sequential left-to-right scan of a binary sequence. This process involves five steps. Step 1 – Bandpass filtering. The continuous signal was filtered into six sub-bands using a fourth-order Butterworth filter delta: 1–4 Hz, theta: 4–8 Hz, alpha: 8–13 Hz, beta: 13–30 Hz, low-gamma: 30–60 Hz, high-gamma: 60–150 Hz). Along with the unfiltered broadband signal (1–200 Hz), this provides a total of seven frequency conditions. Step 2 – Median binarization. The filtered signal x(n) of length N was converted into a binary sequence using the median as the threshold, which is a standard convention in LZC computation and is robust to non-Gaussian amplitude distributions: s(n) = 1 if x(n) ≥ median(x), 0 otherwise (1) Step 3 – LZ76 parsing. We analyzed the binary sequence using the Kaspar–Schuster variant of the LZ76 algorithm, which, via a pointer-increment scheme, counts the number of novel substrings c(n). We compiled our implementation to native machine code using Numba’s @njit decorator, which also validates import—an all-zeros sequence produces c = 1 and an alternating 0–1 sequence produces c = 2. Step 4 – Length normalization. The raw complexity count was normalized against the theoretical upper bound for a random binary string of length N: LZC_norm = c(n) · log₂(N) / N (2) Step 5 – Shuffle normalization. To isolate structural complexity from the marginal statistics of the signal (Medel et al. 2023 ), 100 surrogate sequences were generated by Fisher-Yates shuffling. The LZC value is computed as the observed LZC_norm divided by the mean LZC_norm of the surrogates. Values near 1.0 suggest complexity comparable to a random sequence, whereas values below 1.0 suggest temporal regularity. This pipeline produced 406 features per trial (7 frequency conditions × 58 channels). The selection of these two non-linear indices is grounded in their mathematical complementarity: while LZC is sensitive to the temporal irregularity of pattern sequences, HFD is sensitive to the self-similar structure of the time series. Together, they form a multi-dimensional feature space that is robust to inter-subject spectral variability. 2.6 Higuchi Fractal Dimension HFD estimates the fractal dimension of a time series by measuring how its curve length L(k) scales across integer delay parameters k = 1, 2, …, k_max. For each delay k and starting offset m = 1, …, k, the normalized length is: L_m(k) = [(N − 1) / (floor((N − m)/k) · k²)] · Σ|x(m + jk) − x(m+(j − 1)k)| (3) The average over all offsets gives L(k), and HFD is obtained as the slope of log L(k) versus log(1/k) through linear regression. We used vectorized NumPy polyfit to compute the regression across all 58 channels simultaneously. The parameter k_max was set to 64, following values used in previous EEG studies (Aggarwal and Ray 2025 ; Colussi et al. 2025 ; Höhn et al. 2024 ). Olejarczyk et al. ( 2022 ) and Armonaite et al. ( 2023 ) have shown that HFD functions as a region-specific cortical complexity signature, which supports its use in topographic analyses like the one reported here. Two temporal resolutions were employed. Epoch-level HFD was computed once over the full 6-second trial (6000 samples at 1000 Hz). Sliding-window HFD was computed in a 1.0-second window advanced in 0.1-second steps, yielding approximately 60 windows per trial. From each channel’s window trajectory, the mean and standard deviation were extracted. The total number of HFD features was 174 per trial (58 channels × 3 statistics: epoch-level, window mean, window SD). 2.7 Dual Ground Truth The PerceiveImagine dataset records a vividness self-report after each trial, but the ratings in this cohort are constant across all subjects and trials (value = 3 on a 1–5 scale). Because there is no within-subject variance in vividness, we adopted a dual ground truth framework instead. Objective label. A LOSO-CV logistic regression classifier (L2 regularization, C = 0.1 chosen to balance bias and variance, L-BFGS solver, StandardScaler normalization) was trained to discriminate perception from imagination using broadband LZC features. The L2 penalty mitigates overfitting despite the high feature-to-subject ratio (406 features vs. 46 subjects). For each held-out subject, the predicted probability of the “imagination” class served as a continuous objective quality metric. Per-subject AUC was computed wherever both classes were present. All 46 subjects were valid. Proxy subjective label. Because vividness ratings lacked variance, the mean decoding probability per subject was used as a proxy. Subjects were split at the median into a high-decoding group (n = 23) and a low-decoding group (n = 23). This design creates a known circularity, which is discussed openly in Section IV-D. A permutation baseline was also constructed for hypothesis H8: for each subject, the true test-set labels were randomly permuted 50 times, and the resulting mean AUC served as a subject-specific estimate of chance performance (approximately 0.500). 2.8 Statistical Analysis Eight hypotheses were tested. Cluster-based permutation testing used 1000 permutations with spatial adjacency computed from the 10–20 montage via the MNE built-in channel adjacency matrix (Maris and Oostenveld 2007 ). This non-parametric procedure controls the family-wise error rate across spatially adjacent channels. The significance threshold was α = 0.05 throughout. The Benjamini–Hochberg procedure was applied to the six hypotheses with conventional p-values (H1, H2, H3, H4, H6, H8); H5 was assessed against a pre-specified AUC threshold and H7 was untestable, so both were excluded from the correction. H1 : Low-gamma (30–60 Hz) LZC is higher during imagination than perception at six occipital channels (Oz, O1, O2, POz, PO3, PO4). Effect size: paired Cohen’s d. H2 : Low-gamma LZC is higher in the high-decoding group than the low-decoding group during imagination. Effect size: independent-samples pooled Cohen’s d. H3 Frontal HFD shows a condition × decoding-quality interaction. Tested via linear mixed-effects model (HFD ~ condition × y_subj_raw + (1|subject)) at five frontal channels (Fp1, Fp2, F3, Fz, F4). Because y_subj_raw is a subject-level variable rather than a trial-level covariate, the effective degrees of freedom for the interaction term are governed by the number of subjects (N = 46), not the number of trials. H4 Broadband LZC topography differs between perception and imagination. Tested by both cluster-based permutation and Hotelling T² on PCA-reduced channel vectors (10 components). To maintain a robust observations-to-variables ratio for the multivariate test ( N = 46 vs. k = 10 ), dimensionality was reduced to the first 10 principal components, which accounted for 50.6% of the total topographic variance while capturing the predominant spatial structures. H5 LZC features achieve AUC > 0.72 for state discrimination in the LOSO-CV classifier. H6 : High-decoding subjects show higher objective accuracy than low-decoding subjects. Normality was assessed via Shapiro-Wilk; a Wilcoxon rank-sum test was applied. Effect size: rank-biserial correlation r_rb. H7 Within-subject Spearman correlation between vividness and decoding accuracy exceeds r = 0.3. Not testable due to zero vividness variance. H8 : Objective-label AUC exceeds the permutation-baseline AUC. Tested by paired t-test, with Bayesian analysis using a Cauchy prior (scale r = 0.707) (Rouder et al. 2009 ). Robustness check: Wilcoxon signed-rank test. 3. Results 3.1 Preprocessing Summary All 46 subjects passed through the preprocessing pipeline without failure. Picard ICA converged in every case. A mean of 3.2 ± 1.8 ICA components per subject were excluded (range: 1–7; distribution shown in Supplementary Fig. S1 ). The edge-channel EMG monitor flagged fewer than 5% of imagination epochs across the cohort. After epoching, the master feature table contained 30,534 trials with 580 features each (406 LZC + 174 HFD). Both perception and imagination conditions were represented for every subject. 3.2 Hypothesis Testing A comprehensive quantitative summary of all hypothesis tests, including test statistics, p-values, and effect sizes, is provided in Table 1 . Table 1 Hypothesis Testing Summary ID Description Test Stat. p p_adj Effect [CI] Sig ES Label H1 LZC imag>perc occ. low-γ Cluster perm — 1.000 1.000 d = − 0.13 No Cohen d H2 LZC hi > lo decode low-γ Cluster perm — 0.169 0.254 d = + 0.88 No Pooled d H3 HFD frontal×cond LME t = 0.54 0.591 0.709 β = 0.003 No LME coeff H4 LZC topo imag vs perc Cluster perm* p_cl = .005 0.005 0.010 V = 0.28 Yes Pseudo-Pillai H5 LZC predict state LOSO-CV AUC — — .811[.775,.847] Yes AUC + CI H6 Hi > lo decode acc. Wilcoxon U = 2.33e8 < .001 0.003 r_rb = 0.87 Yes Rank-biserial H7 Vivid~decode corr. — — — — — N/A — H8 Obj AUC ≥ baseline Paired t t = 16.30 8.4×10 − 21 5.0×10 − 20 BF = 8.84e14 Yes Bayes Factor *H4 independently confirmed by Hotelling T²: F = 3.08, p = 0.002, V = 0.28. †H5 significance assessed by AUC > 0.72 threshold; excluded from FDR correction. H7 untestable. p_adj: Benjamini–Hochberg adjusted p-values across the six testable hypotheses (H1–H4, H6, H8). All three significant findings (H4, H6, H8) survived FDR correction (p_adj < 0.05). H1. The cluster-based permutation test found no significant clusters at the occipital ROI (p = 1.0). The paired Cohen’s d was − 0.13, indicating a negligible effect in the direction opposite to prediction: perception showed marginally higher low-gamma LZC than imagination at occipital sites. H2. No significant spatial cluster was identified in the cluster permutation test across the full 58-channel array (p = 0.169). However, the independent-samples pooled Cohen's d for low-gamma was + 0.878 (p = 0.005, n_high = 23, n_low = 23), a medium-to-large effect in the predicted direction. This dissociation between cluster permutation outcome and effect size is discussed in Section IV-A. A band-wise breakdown of effect sizes across all seven frequency bands is provided in Supplementary Fig. S2 . H3. The linear mixed-effects interaction term was not significant (t = 0.54, p = 0.591, β = 0.003). Given the subject-level covariate limitation noted in Methods, this null result should be interpreted with caution. H4. This was the primary positive finding at the signal level. The cluster-based permutation test found a statistically significant cluster across channels (p = 0.005). An independent Hotelling T² test applied to PCA-reduced LZC vectors showed multivariate significance (F = 3.08, p = 0.002, Pseudo-Pillai V = 0.28, medium effect). Topographic analysis is presented in Fig. 2 ; the spatial patterns are further detailed in Supplementary Fig. S3 and S4. H5. The LOSO-CV classifier's AUC of 0.811 (95% bootstrap CI: 0.775–0.847, 2000 resamples) is significantly greater than a 0.50 chance level. All 46 subjects contributed valid test folds. H6. For both groups, the Shapiro-Wilk test rejected normality. The Wilcoxon rank-sum test yielded U = 2.33 × 10⁸, p < 0.001, with a rank-biserial correlation of r_rb = 0.87—a very large non-parametric effect. H7. This hypothesis could not be evaluated since all subjects gave a constant rating of 3 for vividness. H8. The paired t-test comparing real-label AUC to permutation-baseline AUC yielded t(45) = 16.30, p = 8.35 × 10⁻²¹. The Bayes Factor was BF₁₀ = 8.84 × 10¹⁴ which suggests overwhelming evidence that the complexity features captured genuine neural structure. This result was confirmed by a Wilcoxon signed-rank test (W = 1076, p < 0.001). The result, as expected under the null, showed the permutation-baseline AUC was 0.500. A visual summary of the eight hypothesis outcomes is shown in Fig. 3 . 3.3 Topographic Analysis Figure 2 shows group-averaged broadband LZC topographic maps from three conditions (with imagery further divided into high- and low-decoding subsets). Supplementary Fig. S3 further details these topographies across three conditions, illustrating the spatial differences when imagery is divided into high- and low-decoding subsets. Overall LZC magnitudes were similar per condition (range 0.20–0.40), suggesting that the differences lie in where on the scalp the complexity concentrates, rather than the magnitude of the complexity. Regarding perception, complexity seemed to be evenly dispersed. During high-decoding imagination, increased complexity was observed in frontal regions. During low-decoding imagination, there was a prominent reduction of complexity in the right parieto-occipital region and the frontal end was weaker. The complexity reallocation on the frontal and occipital areas may be a reflection of differences in top-down imagery processing, although the scalp-recorded EEG cannot determine the cortical sources. The topographic features help us understand why H1 (which tested a general increase over occipital sites) returned a null finding, while H4 (which tested for a multivariate topographic difference across the full channel array) was significant. 4. Discussion 4.1 The Gamma-Band Null The most surprising finding in this study is the contravention of the first two hypotheses. We thought the imagination condition would yield higher low-gamma LZC values at the occipital sites due to the endogenous gamma-band binding in the absence of sensory phase-locking which, in the absence of sensory inputs, should produce more diverse and, therefore, temporally complex patterns. The data indicated the opposite trend, whereby perception showed marginally higher values (d = − 0.13; see Supplementary Fig. S2 and the band-wise trajectory in Fig. S5 ). There are at least three possible explanations. First, LZC quantifies broadband irregularity, which is different from phase coherence. A gamma signal can be more phase-coherent (which would reflect tighter binding) while simultaneously becoming less complex (because phase-locked signals are inherently more regular). If ‘binding’ manifests primarily as phase coherence, then LZC is the wrong probe for detecting it. Second, it is possible that our wideband ICA pipeline removed too much neural signal. The absence of gamma-band differences—despite our large cohort ( N = 46 )—highlights the potential for myogenic bias in previous complexity studies. Our implementation of Picard ICA plus edge-channel monitoring likely sequestered these artifacts, revealing that the core neural signature of imagery is topographic rather than frequency-specific. Third, binding bursts are transient events that last only hundreds of milliseconds. Averaging LZC over a 6 seconds window will likely dilute those episodes. A time-resolved LZC analysis will be more sensitive to this averaging effect and may identify binding episodes in the first 1–2 seconds of the imagination epoch when top-down reconstruction processes are highest. Additionally, the potential the gamma-band complexity holds as a measure of ‘binding’ should be compared to that of others to come to a more accurate conclusion. Phase-coherence or phase-amplitude coupling measures may be better suited to measure the ‘synchronization’ aspect of the phenomenon. A methodologically important nuance emerges from H2. Although the cluster-based permutation test did not yield a significant spatial cluster for low-gamma LZC (p = 0.169), the independent-samples pooled Cohen's d for low-gamma was + 0.878 (p = 0.005, n_high = 23, n_low = 23)—a medium-to-large effect in the predicted direction. This dissociation between the univariate effect size and the cluster permutation outcome reflects the stringency of spatial cluster correction rather than the absence of a signal. Cluster-based permutation tests require effects to be spatially contiguous across adjacent channels; a genuine effect that is diffusely distributed, or that concentrates in a small number of non-adjacent channels, can fail this threshold even when the channel-level t-statistic is individually significant. In practical terms, H2's cluster permutation null and its large positive d are not contradictory: they together suggest that high-decoding subjects show elevated low-gamma LZC relative to low-decoding subjects, but that this elevation is not spatially focal enough to survive whole-scalp cluster correction. Future analyses using region-of-interest (ROI) tests restricted to occipital or frontal channels, rather than whole-scalp cluster permutation, may provide greater sensitivity to detect this effect. 4.2 Topographic Redistribution and Predictive Coding The H4 result tells a different story. The topography differences of broadband LZCs of frontal and occipital channels relative to perception and imagination are confirmed by the cluster permutation test (p = 0.005) and Hotelling T² (F = 3.08, p = 0.002, V = 0.28, a medium multivariate effect). The topography difference is not just a simple uniform shift. This observed frontal–occipital complexity gradient provides a distinct neural signature for top-down sensory reconstruction. It suggests that while perception is anchored by sensory inputs, imagery relies on high-complexity generative predictions from the frontal cortex that modulate posterior dynamics. This frontal–occipital redistribution supports predictive-coding explanations for imagery. From that perspective, higher-order cortical areas form internal predictions that feedback to the sensory cortex during imagery, thus reversing the up-stream (i.e., bottom-up) information flow that characterizes perception (Chang et al. 2025 ; Dijkstra et al. 2021 ). The data showing increased spatial complexity in the regions of the frontal cortex and decreased complexity in the occipital cortex are in agreement with this predictive-coding model. Dijkstra et al. ( 2021 ) observed that the sensory cortex is recruited during imagery with a level of feedback that is slower and more variable than during perception. Our data support that description in the region of the frontal cortex (Dijkstra et al. 2021 ). The increase in frontal spatial complexity aligns with the activity of the medial prefrontal region of the default mode network (Rouder et al. 2009 ), which is known to be highly active during internally directed cognition. Scalp EEG is unable to detect the source of the activity, and the subject's attentional modulation remains a possible explanation. The fractal dimension measure has also been sensitive to emotional states (Duville et al. 2022 ) and neurodegenerative conditions (Yoder et al. 2024 ), which suggests that the topographic complexity patterns observed here may generalize to other cognitive contrasts. Future work should test this redistribution using source-reconstructed EEG to determine whether the scalp-level pattern maps onto specific cortical regions. The proposed complexity-based decoding framework is summarized in Fig. 4 . 4.3 Classification and the Objective Label Advantage The LOSO-CV classifier scored an AUC of 0.811 to discriminate perception from imagination. This is an impressive result for cross-subject EEG classification. Guenther et al. ( 2024 ) found similar visual classification performance from a sparser 8-channel EEG system, and Kalafatovich et al. ( 2023 ) presented a high performance visual EEG classification using spatiotemporal graph learning. By design, broad-band LZC features are agnostic to frequency, which may make them more robust to inter-subject spectral variability than narrowband power features. The result from H8 is the most compelling in this research. With true labels, the classifier was able to pick up genuine neural structure; With shuffled labels the signal entirely vanished (Permutation-baseline AUC = 0.500). The contrast was overwhelming: t(45) = 16.30, BF₁₀ = 8.84 × 10¹⁴. There are translational implications of this finding. If complexity features can be used reliably to differentiate imagery from perception with an AUC > 0.80, then they could be used in BCI systems to provide objective measurements of imagery decodability during neurofeedback or cognitive rehabilitation. Figure 5 shows the AUC distribution of each subject and its comparison with the permutation baseline. Methodologically, this study shows that wideband ICA can maintain genuine high-gamma EEG activity during cognitive tasks, without manual component inspection. The Picard ICA plus edge-channel EMG monitoring template is adoptable to study gamma-band phenomena. Future studies should study the complexity features against conventional spectral and connectivity-based EEG markers to establish their relative contribution to imagery classification. 4.4 Limitations Six limitations should be acknowledged in this study. First, the lack of variance in the dataset’s vividness ratings rendered hypothesis H7 untestable. The proxy subjective label was derived from the objective decoding metric, creating a circular reasoning: hypotheses H2 and H6 are strengthening themselves instead of testing subjective–objective dissociation. The proxy label is not an independent ground truth, as it is not a reflection of subjective experience, and may bias effect sizes upward. To test the dual label framework, datasets with truly variable self-reports are needed; or independent behavioral measures such as drawing accuracy or recognition memory. Second, the H3 linear mixed-effects model uses a subject-level covariate (mean decoding probability) as a predictor in a trial-level model. The interaction term’s effective degrees of freedom are governed by N = 46, not the trials count. The lack of significant finding (p = 0.591) could reflect power insufficiency, a model misspecification, or genuine absence of the effect. Third, the N = 46 sample size may be adequate for the permutation framework, but may be underpowered for LME or independent-samples comparisons for small effects. The single cultural and institutional background cohort spans a narrow age range (23–30), which limits generalizability to a wider population. Fourth, the LZ76 median binarization discards all amplitude information from the signal. Alternative binarization strategies (Hilbert envelope, multi-symbol alphabets) might capture complementary complexity features. The HFD parameter k_max = 64 was selected from the literature; its optimality for imagery paradigms has not been empirically evaluated. Fifth, the 58-channel 10–20 montage imposes a hard limit on spatial resolution. The topographic differences reported here could arise from multiple distinct source configurations. Individual-MRI head models and source localization methods such as beamforming would sharpen the inference considerably. Sixth, time-resolved LZC could capture transient complexity changes during the first one to two seconds of the imagery epoch, when the demand for top-down reconstruction is presumably highest. This analysis would directly address the gamma-band null described in Section IV-A and is a priority for future work. 5. Conclusion We tested eight hypotheses about non-linear EEG complexity as a biomarker of visual mental imagery in 46 participants. The predicted gamma-band LZC elevation did not appear. What emerged instead was a significant topographic redistribution of broadband LZC between perception and imagination (cluster p = 0.005; Hotelling T² p = 0.002, V = 0.28; see Supplementary Fig. S4 for detailed cluster topographies). Combined LZC and HFD features classified cognitive state at AUC = 0.811 (95% CI: 0.775–0.847), and objective decoding labels outperformed chance baselines with decisive evidence (BF₁₀ = 8.84 × 10¹⁴). These findings reframe mental imagery as a process of large-scale topographic complexity redistribution, validating the predictive coding framework through the lens of non-linear neurodynamics. Declarations Acknowledgments The authors thank Prof. Luis Miguel Pinho de Almeida for supervision and guidance throughout this project. The creators of the PerceiveImagine dataset (OpenNeuro ds005697) are acknowledged for making their data publicly available. This work was conducted within the PhD Programme in Health Data Science, Faculty of Medicine, University of Porto (FMUP), and the Department of Electrical and Computer Engineering, Faculty of Engineering, University of Porto (FEUP). No external funding was received for this research. Funding No funds, grants, or other support was received. Competing interests The authors have no competing interests to declare that are relevant to the content of this article. Ethics approval The data utilized in this study were obtained from the publicly available OpenNeuro repository (accession number ds005697), which provides de-identified EEG samples. This study was deemed exempt from ethical approval as it involved the secondary analysis of fully anonymized data. The original study protocols were in accordance with the ethical standards of the original institution and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. Consent to participate Not applicable. Consent to publish Not applicable. Data availability The EEG dataset "PerceiveImagine" analyzed during the current study is available in the OpenNeuro repository (https://openneuro.org/datasets/ds005697). The code pipelines are available from the corresponding author on reasonable request. Authors’ contribution Yu Gao: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Project administration; Software; Visualization; Roles/Writing - original draft; Writing - review & editing José Miguel Diniz: Formal analysis; Validation; Writing - review & editing References Ablin P, Cardoso J-F, Gramfort A (2018) Faster ICA by preconditioning with Hessian approximations. IEEE Trans Signal Process 66:4040–4049. https://doi.org/10.1109/TSP.2018.2844203 Armonaite K, Nobili L, Paulon L, Balsi M, Conti L, Tecchio F (2023) Local neurodynamics as a signature of cortical areas: new insights from sleep. Cereb Cortex 33(6):3284–3292. https://doi.org/10.1093/cercor/bhac274 Aggarwal S, Ray S (2025) Changes in Higuchi fractal dimension across age in healthy human EEG. Eur J Neurosci 62:e70193. https://doi.org/10.1111/ejn.70193 Chang YH, Chen HA, Tsai MJ, Tseng CL, Lo CH, Huang KC, Wei CS (2025) A human EEG dataset for multisensory perception and mental imagery. Sci Data 12:1598. https://doi.org/10.1038/s41597-025-05881-1 Colussi F, Favaro J, Ancona C, Passarotto E, Pelizza MF, Lorenzon E, Ruzzante S, Masiero S, Perilongo G, Sparacino G, Toldo I, Sartori S, Rubega M (2025) EEG difference in the Higuchi fractal dimension of wakefulness and sleep from birth to adolescence. PLoS ONE 20:e0333903. https://doi.org/10.1371/journal.pone.0333903 Duville MM, Alonso-Valerdi LM, Ibarra-Zarate DI (2022) Fractal dimension of EEG signals and heart dynamics in discrete emotional states. Front Comput Neurosci 16:1022787. https://doi.org/10.3389/fncom.2022.1022787 Dijkstra N, Mostert P, Lange FP, Bosch S, van Gerven MAJ (2021) Differential temporal dynamics during visual imagery and perception. eLife 10:e47892. https://doi.org/10.7554/eLife.47892 Gifford AT, Dwivedi K, Roig G, Cichy RM (2022) A large and rich EEG dataset for modeling human visual object recognition. NeuroImage 264:119754. https://doi.org/10.1016/j.neuroimage.2022.119754 Guenther S, Kosmyna N, Maes P (2024) Image classification and reconstruction from low-density EEG. Sci Rep 14:16436. https://doi.org/10.1038/s41598-024-66228-1 Höhn C, Hahn MA, Lendner JD, Hoedlmoser K (2024) Spectral slope and Lempel–Ziv complexity as robust markers of brain states during sleep and wakefulness. eNeuro 11(3). https://doi.org/10.1523/ENEURO.0259-23.2024 Kalafatovich J, Lee M, Lee SW (2023) Learning spatiotemporal graph representations for visual perception using EEG. IEEE Trans Neural Syst Rehabil Eng 31:97. https://doi.org/10.1109/TNSRE.2022.3217344 Li X, Fan Y (2024) PerceiveImagine: a human EEG dataset for studying perception and mental imagery. Sci Data 11:681. https://doi.org/10.1038/s41597-024-03529-6 Lee S, Jang S, Jun SC (2022) Exploring the ability to classify visual perception and visual imagery EEG data. Electronics 11:2706. https://doi.org/10.3390/electronics11172706 Medel V, Irani M, Crossley N, Ossandón T, Boncompte G (2023) Complexity and 1/f slope jointly reflect brain states. Sci Rep 13:21870. https://doi.org/10.1038/s41598-023-47316-0 Maris E, Oostenveld R (2007) Nonparametric statistical testing of EEG- and MEG-data. J Neurosci Methods 164(1):177–190. https://doi.org/10.1016/j.jneumeth.2007.03.024 Olejarczyk E, Gotman J, Frauscher B (2022) Region-specific complexity of the intracranial EEG in the sleeping human brain. Sci Rep 12:451. https://doi.org/10.1038/s41598-021-04213-8 Ruiz de Miras J, Ibáñez-Molina AJ, Soriano MF, Iglesias-Parro S (2023) Fractal dimension analysis of resting state functional networks in schizophrenia. Front Hum Neurosci 17:1236832. https://doi.org/10.3389/fnhum.2023.1236832 Rouder JN, Speckman PL, Sun D, Morey RD, Iverson G (2009) Bayesian t tests for accepting and rejecting the null hypothesis. Psychon Bull Rev 16(2):225–237. https://doi.org/10.3758/PBR.16.2.225 Ren Z, Yue M, Han X, Zhao Z, Wang B, Hong Y, Zhao T, Wang N, Zhao P, Hong Y, Wang Q, Zhao Y (2023) The potential of the Lempel–Ziv complexity of the EEG in diagnosing cognitive impairment. Epileptic Disord 25(3):331–342. https://doi.org/10.1002/epd2.20044 Shimizu H, Srinivasan R (2022) Improving classification and reconstruction of imagined images from EEG. PLoS ONE 17:e0274847. https://doi.org/10.1371/journal.pone.0274847 Wilson H, Chen X, Golbabaee M, Proulx MJ, O'Neill E (2024) Feasibility of decoding visual information from EEG. Brain-Comput Interfaces 11(1–2):33–60. https://doi.org/10.1080/2326263X.2023.2287719 Yoder KJ, Brookshire G, Glatt R, Merrill DA, Gerrol S, Quirk C et al (2024) Fractal dimension distributions of resting-state EEG improve detection of dementia. Fractal Fract 8:27. https://doi.org/10.3390/fractalfract8030027 Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 22 Apr, 2026 Reviewers agreed at journal 19 Apr, 2026 Reviews received at journal 17 Apr, 2026 Reviewers agreed at journal 17 Apr, 2026 Reviewers agreed at journal 16 Apr, 2026 Reviewers invited by journal 16 Apr, 2026 Editor assigned by journal 13 Apr, 2026 Submission checks completed at journal 13 Apr, 2026 First submitted to journal 10 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9382178","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":627581436,"identity":"0392cb43-ff48-496e-8630-f46597981edc","order_by":0,"name":"Yu Gao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAElEQVRIiWNgGAWjYBACgwMMjB8YGGwY+EC8BIYDDAzsYAlmfFqYJRgY0hjY4FqYCWiRbAArPgzRwkCMFn6JBDaGHxXn5dkkstMePGC4I29wmPnYB4YK68QGHFrYgFoYe87cNmyTyN1ukMDwzHDDYbbkGQxn0vFqseBtu80I1LJNIoHhMOOGwzzGDIxth/Fqkfzbds4epsUeouUffi3SvG0HEmFaEiFaGvBo4XnALC1zJjm5jectUIvBs+SZQL8wJBxLN8aphT2B8eObCjvbfvbcbZI/Ku7Y9h1vPszwocZaFpcWYDB/QOIYMDAoHGAAxSkpQB636aNgFIyCUTBCAQCm81WI1oLk5wAAAABJRU5ErkJggg==","orcid":"","institution":"University of Porto","correspondingAuthor":true,"prefix":"","firstName":"Yu","middleName":"","lastName":"Gao","suffix":""},{"id":627581439,"identity":"e49dea57-76c1-4c1f-8e3f-bdf32cc4cfc6","order_by":1,"name":"José Miguel Diniz","email":"","orcid":"","institution":"University of Porto","correspondingAuthor":false,"prefix":"","firstName":"José","middleName":"Miguel","lastName":"Diniz","suffix":""}],"badges":[],"createdAt":"2026-04-10 17:39:53","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9382178/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9382178/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107744816,"identity":"4f112ee4-c26c-4d4c-b5af-d7e399e404d5","added_by":"auto","created_at":"2026-04-24 15:46:26","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":200213,"visible":true,"origin":"","legend":"\u003cp\u003eEnd-to-end preprocessing and feature extraction pipeline. Raw 58-channel EEG is bandpass-filtered (1–200 Hz), notch-filtered, and cleaned via wideband Picard ICA. Post-ICA edge-channel EMG monitoring flags high-gamma artefacts. Clean epochs undergo parallel LZC and HFD extraction. Dual ground truth labels are constructed from LOSO-CV decoding confidence\u003c/p\u003e","description":"","filename":"figure14.png","url":"https://assets-eu.researchsquare.com/files/rs-9382178/v1/50cd0bc0d7a7c340285e1158.png"},{"id":107868897,"identity":"7f41b248-9fcd-4cb9-99e1-5d503b99476c","added_by":"auto","created_at":"2026-04-27 07:34:44","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":456514,"visible":true,"origin":"","legend":"\u003cp\u003eTopographic maps of Group-averaged Lempel-Ziv Complexity (LZC) for visual perception and mental imagination tasks (N = 46). The color scale shows the normalized LZC values, with warm colors indicating greater neural complexity. Black dots show the 58 scalp electrodes used for the analyses\u003c/p\u003e","description":"","filename":"figure24.png","url":"https://assets-eu.researchsquare.com/files/rs-9382178/v1/8b494f43097a3a7473532a22.png"},{"id":107744818,"identity":"18dfe45c-ae62-4648-a011-d5dac97868d4","added_by":"auto","created_at":"2026-04-24 15:46:26","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":116947,"visible":true,"origin":"","legend":"\u003cp\u003eVisual overview of hypothesis testing results from H1 - H8. Statistically significant results at p \u0026lt; 0.05 are presented in blue bars. Red bars show results that are not statistically significant. Grey bars are for hypotheses that are not testable (H7). Effect sizes are presented in appropriate metrics. These include Cohen’s d, Pseudo-Pillai V, AUC, rank-biserial r, or Bayes Factor. The hypotheses that are significant are marked with an asterisk (*)\u003c/p\u003e","description":"","filename":"figure34.png","url":"https://assets-eu.researchsquare.com/files/rs-9382178/v1/8ddaf14ec859f0aadc4231e0.png"},{"id":107869481,"identity":"880e8d6c-67f0-42bd-b07e-054215b7d4e4","added_by":"auto","created_at":"2026-04-27 07:37:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":92606,"visible":true,"origin":"","legend":"\u003cp\u003eProposed complexity-based framework for evaluating imagery decodability. LZC and HFD features are extracted from wideband-cleaned EEG and fed into a LOSO-CV classifier. The decoding confidence serves as a proxy index of imagery decodability\u003c/p\u003e","description":"","filename":"figure44.png","url":"https://assets-eu.researchsquare.com/files/rs-9382178/v1/c558086b235753f3869ab66e.png"},{"id":107744820,"identity":"310309db-9d70-4b7c-8afa-2b9286f6e724","added_by":"auto","created_at":"2026-04-24 15:46:26","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":139392,"visible":true,"origin":"","legend":"\u003cp\u003eCross-validation with leave-one-subject-out (N = 46). (A) Per-subject AUC for real label(blue) and label-permuted baseline (orange), sorted by real-label AUC. The dashed line marks chance level (0.50) and the dotted line represents the group mean AUC (0.811). The shaded area indicates the 95% bootstrap CI [0.775, 0.847]. (B) Boxplot of AUC distributions with individual data points. The paired t-test yielded p = 8.35 × 10\u003csup\u003e-21\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"figure52.png","url":"https://assets-eu.researchsquare.com/files/rs-9382178/v1/8e544b82fe4f105e52f5079f.png"},{"id":107871850,"identity":"e9ecf1f5-d9d8-4f48-9f4e-fc4552b15b2f","added_by":"auto","created_at":"2026-04-27 07:54:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1117809,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9382178/v1/cbb5df71-3267-404e-83cd-90928dcba143.pdf"},{"id":107744815,"identity":"e04323f8-6760-4e2b-b688-56212518c3c7","added_by":"auto","created_at":"2026-04-24 15:46:26","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1548313,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-9382178/v1/b3b1064a64574c0edf5b8a2b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Neural Complexity Signatures of Visual Mental Imagery: Lempel-Ziv Complexity and Higuchi Fractal Dimension Reveal Topographic Reorganization in High-Density EEG","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eVisual mental imagery refers to the ability to internally reconstruct perceptual experience without external sensory input. When a person imagines an apple, primary visual cortex activates in proportion to the vividness of that image, despite the absence of any retinal signal (Dijkstra et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Understanding how the brain accomplishes this reconstruction has practical relevance for brain\u0026ndash;computer interfaces, neurofeedback, and cognitive rehabilitation. Most EEG studies still focus on linear spectral qualities, such as band power, which likely overlooks the irregularity of the endogenously determined neural processes.\u003c/p\u003e \u003cp\u003eTwo non-linear metrics, however, provide a different perspective. Lempel-Ziv Complexity (LZC) measures the number of distinct sequential patterns in a binarized signal: the more patterns, the greater the signal's temporal irregularity. Higuchi Fractal Dimension (HFD) determines the scaling of a time series' curve length across various time resolutions, producing a Fractal Dimension that indicates geometric roughness. LZC and HFD capture orthogonal facets of neural dynamics. LZC represents the algorithmic complexity (symbolic novelty) of the binarized signal, while HFD quantifies its geometric scaling (fractal roughness) across time resolutions. This dual-metric approach provides a more comprehensive characterization of non-stationary EEG dynamics than either index alone.\u003c/p\u003e \u003cp\u003eRecent studies have confirmed that the two metrics are responsive to the different brain states. H\u0026ouml;hn et al. (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) showed that broadband LZC and spectral slope together monitor sleep/wake states in healthy adults. Medel et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found that LZC and the 1/f spectral slope provide a complementary description of the excitation-inhibition balance in the cortex. Ren et al. (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) applied LZC to resting state EEG, and were able to identify cognitive impairment in epilepsy patients. On the HFD side, Aggarwal and Ray (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) reported that HFD in scalp EEG is anticorrelated with oscillatory power and 1/f slope across the adult lifespan. Colussi et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) used HFD to differentiate wakefulness from sleep in a developmental cohort. Armonaite et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) validated HFD as a region-specific cortical complexity signature, and Ruiz de Miras et al. (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) applied fractal dimension analysis to resting-state EEG networks in schizophrenia. Despite this growing body of work, no study has combined LZC and HFD as joint biomarkers for mental imagery decodability.\u003c/p\u003e \u003cp\u003eThe imagery literature itself has an unresolved measurement problem. Dijkstra et al. (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) found that perception and imagery share neural substrates in visual cortex but differ in temporal dynamics\u0026mdash;imagery recruits top-down feedback more slowly and with greater trial-to-trial variability. Chang et al. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) released a multisensory EEG imagery dataset with vividness ratings, confirming that subjective vividness correlates with distinguishable EEG power patterns. Wilson et al. (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) reviewed the feasibility of visual decoding from EEG. Shimizu and Srinivasan (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) improved classification of imagined images using spectral features. Gifford et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) provided a large benchmark EEG dataset for visual object recognition, and Lee et al. (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) explored the classification of perception versus imagery. The common thread across these studies is a reliance on subjective vividness ratings as the only ground truth for imagery decodability, even though self-report is noisy and inconsistent across sessions. An objective approach is to classify perceptual and imagined states, and use the classifier\u0026rsquo;s decoding confidence as a continuous quality measure.\u003c/p\u003e \u003cp\u003eWe incorporated this dual-labeling method with our broad-band artefact-control processing pipeline. The gamma-band range (30-150Hz) is inseparable from the scalp muscle artefact, which, if unchecked, can confound complexity differences. Instead of removing the high frequencies, we applied Picard ICA (Ablin et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) over the entire 1-200 Hz range; and used EMG power monitoring as a secondary protective measure.\u003c/p\u003e \u003cp\u003eWe evaluated eight hypotheses related to 46 participants in the PerceiveImagine paradigm (Li and Fan \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) and extracted 580 complexity features per trial (406 LZC\u0026thinsp;+\u0026thinsp;174 HFD) from 58 scalp EEG channels. The gamma-band LZC elevation we predicted did not occur. Instead, we observed a significant topographic redistribution of broadband complexity between perception and imagination. This finding is in line with predictive coding frameworks suggesting imagery redistributes the cortical dynamics, rather than elevating their magnitude uniformly.\u003c/p\u003e \u003cp\u003e \u003cb\u003eThis study makes three contributions.\u003c/b\u003e Firstly, it establishes that EEG complexity features capture a specific reallocation of symbolic and geometric complexity, reflecting the reversal of cortical information flow as predicted by top-down generative models. Secondly, it highlights that decoding-derived confidence surpasses chance level baselines with strong Bayesian evidence (BF\u003csub\u003e10\u003c/sub\u003e\u0026thinsp;\u0026gt;\u0026thinsp;10\u003csup\u003e14\u003c/sup\u003e), thus qualifying it as a valid objective measure of imagery decodability. Third, it validates a reproducible pipeline for shuffle-normalized LZC and sliding-window HFD extraction on a 46-subject public dataset. Code is available from the corresponding author upon reasonable request.\u003c/p\u003e \u003cp\u003eWe hypothesized that mental imagery would result in varying complexity signatures in the low-gamma band (30\u0026ndash;60 Hz). This a priori choice is based on theory regarding the role of gamma-band neural synchrony in endogenous neural binding and top-down sensory reconstruction. By isolating this band and employing a rigorous wideband ICA artifact-control pipeline, we aimed to determine whether reported gamma-band complexity changes reflect genuine neural dynamics or are susceptible to scalp muscle contamination.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Dataset and Participants\u003c/h2\u003e \u003cp\u003eWe analyzed the PerceiveImagine dataset (OpenNeuro ds005697) (Li and Fan \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), which was collected from 54 healthy adults at a university laboratory in China. Two participants withdrew during recording. Of the remaining 52, eight were excluded before analysis: seven because of missing or corrupted data files (sub-07, sub-12, sub-13, sub-39, sub-51, sub-52, sub-53, as documented in the dataset release notes) and one (sub-38) because its behavioural log contained no perception or imagination event codes, so that zero valid epochs could be extracted. The final analytic sample consisted of 46 participants, aged 23 to 30, both male and female, with no reported neurological or psychiatric history and normal or corrected-to-normal vision.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data Acquisition\u003c/h2\u003e \u003cp\u003eEEG was recorded at 1000 Hz with a 64-channel Neuroscan SynAmps2 system in a standard 10\u0026ndash;20 electrode layout. After removal of auxiliary channels (HEO, VEO, EKG, EMG, CB1, CB2), 58 scalp EEG channels were retained for analysis. Each experimental trial comprised a fixation cross (2 s), a visual stimulus presentation (6 s), a brief gap (0.5 s), a mental imagination period (6 s), and a vividness self-report. The stimulus set contained 340 naturalistic images. The end-to-end preprocessing and feature extraction pipeline is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Preprocessing\u003c/h2\u003e \u003cp\u003eAll preprocessing was carried out in MNE-Python (\u0026ge;\u0026thinsp;1.6). The continuous data were bandpass-filtered at 1\u0026ndash;200 Hz using a zero-phase FIR filter and notch-filtered at 50, 100, and 150 Hz to remove power-line harmonics. A common average reference was applied. We then fitted Picard ICA (Ablin et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) on the wideband signal with extended decorrelation enabled (random seed\u0026thinsp;=\u0026thinsp;42; number of components set to the data rank). ICLabel classified each component automatically; those labelled \u0026ldquo;muscle\u0026rdquo; with probability exceeding 0.6 or \u0026ldquo;eye\u0026rdquo; were excluded, following the default ICLabel threshold for automated pipelines. Across subjects, a mean of 3.2\u0026thinsp;\u0026plusmn;\u0026thinsp;1.8 ICA components were removed (range: 1\u0026ndash;7). The algorithm converged for all 46 participants (convergence delta\u0026thinsp;\u0026lt;\u0026thinsp;10⁻⁴).\u003c/p\u003e \u003cp\u003eAs a second layer of artefact control, we computed the mean high-gamma band power (80\u0026ndash;150 Hz) at six edge channels (T7, T8, TP7, TP8, PO7, PO8) for each imagination epoch via Welch periodogram. Any epoch whose edge-channel power exceeded 3.0 standard deviations above the per-subject mean was flagged, following standard artefact rejection practice. Fewer than 5% of imagination epochs were flagged across the cohort. Flagged epochs were removed only from high-gamma analyses; broadband and sub-gamma computations retained them.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Epoching\u003c/h2\u003e \u003cp\u003eEvent information was parsed exclusively from the behavioural log files (events.tsv) rather than from hardware trigger channels, because some recordings contained spurious trigger pulses not present in the behavioural logs. A three-tier cascading classifier assigned event labels to conditions: keyword matching was tried first, then numeric prefix rules (labels starting with \u0026ldquo;1\u0026rdquo; mapped to perception, \u0026ldquo;2\u0026rdquo; to imagination), and finally an ordinal fallback for any remaining unmatched labels. Epochs were extracted with a\u0026thinsp;\u0026minus;\u0026thinsp;0.2 to 0.0 s baseline and a post-onset duration of 6.0 s. The resulting master feature table contained 30,534 trials across all 46 subjects, with approximately 331 perception trials and 332 imagination trials per subject, corresponding to the 340-image stimulus set minus occasional parsing losses.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Lempel-Ziv Complexity\u003c/h2\u003e \u003cp\u003eLZC measures a signal\u0026rsquo;s irregularity in time by evaluating the number of unique sub-patterns formed in a sequential left-to-right scan of a binary sequence. This process involves five steps.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStep 1 \u0026ndash; Bandpass filtering.\u003c/b\u003e The continuous signal was filtered into six sub-bands using a fourth-order Butterworth filter delta: 1\u0026ndash;4 Hz, theta: 4\u0026ndash;8 Hz, alpha: 8\u0026ndash;13 Hz, beta: 13\u0026ndash;30 Hz, low-gamma: 30\u0026ndash;60 Hz, high-gamma: 60\u0026ndash;150 Hz). Along with the unfiltered broadband signal (1\u0026ndash;200 Hz), this provides a total of seven frequency conditions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStep 2 \u0026ndash; Median binarization.\u003c/b\u003e The filtered signal x(n) of length N was converted into a binary sequence using the median as the threshold, which is a standard convention in LZC computation and is robust to non-Gaussian amplitude distributions:\u003c/p\u003e \u003cp\u003e \u003cem\u003es(n)\u0026thinsp;=\u0026thinsp;1 if x(n) \u0026ge; median(x), 0 otherwise\u003c/em\u003e (1)\u003c/p\u003e \u003cp\u003e \u003cb\u003eStep 3 \u0026ndash; LZ76 parsing.\u003c/b\u003e We analyzed the binary sequence using the Kaspar\u0026ndash;Schuster variant of the LZ76 algorithm, which, via a pointer-increment scheme, counts the number of novel substrings c(n). We compiled our implementation to native machine code using Numba\u0026rsquo;s @njit decorator, which also validates import\u0026mdash;an all-zeros sequence produces c\u0026thinsp;=\u0026thinsp;1 and an alternating 0\u0026ndash;1 sequence produces c\u0026thinsp;=\u0026thinsp;2.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStep 4 \u0026ndash; Length normalization.\u003c/b\u003e The raw complexity count was normalized against the theoretical upper bound for a random binary string of length N:\u003c/p\u003e \u003cp\u003e \u003cem\u003eLZC_norm\u0026thinsp;=\u0026thinsp;c(n) \u0026middot; log₂(N) / N\u003c/em\u003e (2)\u003c/p\u003e \u003cp\u003e \u003cb\u003eStep 5 \u0026ndash; Shuffle normalization.\u003c/b\u003e To isolate structural complexity from the marginal statistics of the signal (Medel et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), 100 surrogate sequences were generated by Fisher-Yates shuffling. The LZC value is computed as the observed LZC_norm divided by the mean LZC_norm of the surrogates. Values near 1.0 suggest complexity comparable to a random sequence, whereas values below 1.0 suggest temporal regularity.\u003c/p\u003e \u003cp\u003eThis pipeline produced 406 features per trial (7 frequency conditions \u0026times; 58 channels).\u003c/p\u003e \u003cp\u003eThe selection of these two non-linear indices is grounded in their mathematical complementarity: while LZC is sensitive to the temporal irregularity of pattern sequences, HFD is sensitive to the self-similar structure of the time series. Together, they form a multi-dimensional feature space that is robust to inter-subject spectral variability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Higuchi Fractal Dimension\u003c/h2\u003e \u003cp\u003eHFD estimates the fractal dimension of a time series by measuring how its curve length L(k) scales across integer delay parameters k\u0026thinsp;=\u0026thinsp;1, 2, \u0026hellip;, k_max. For each delay k and starting offset m\u0026thinsp;=\u0026thinsp;1, \u0026hellip;, k, the normalized length is:\u003c/p\u003e \u003cp\u003e \u003cem\u003eL_m(k) = [(N\u0026thinsp;\u0026minus;\u0026thinsp;1) / (floor((N\u0026thinsp;\u0026minus;\u0026thinsp;m)/k) \u0026middot; k\u0026sup2;)] \u0026middot; Σ|x(m\u0026thinsp;+\u0026thinsp;jk)\u0026thinsp;\u0026minus;\u0026thinsp;x(m+(j\u0026thinsp;\u0026minus;\u0026thinsp;1)k)|\u003c/em\u003e (3)\u003c/p\u003e \u003cp\u003eThe average over all offsets gives L(k), and HFD is obtained as the slope of log L(k) versus log(1/k) through linear regression. We used vectorized NumPy polyfit to compute the regression across all 58 channels simultaneously. The parameter k_max was set to 64, following values used in previous EEG studies (Aggarwal and Ray \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Colussi et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; H\u0026ouml;hn et al. \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Olejarczyk et al. (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) and Armonaite et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) have shown that HFD functions as a region-specific cortical complexity signature, which supports its use in topographic analyses like the one reported here.\u003c/p\u003e \u003cp\u003eTwo temporal resolutions were employed. Epoch-level HFD was computed once over the full 6-second trial (6000 samples at 1000 Hz). Sliding-window HFD was computed in a 1.0-second window advanced in 0.1-second steps, yielding approximately 60 windows per trial. From each channel\u0026rsquo;s window trajectory, the mean and standard deviation were extracted. The total number of HFD features was 174 per trial (58 channels \u0026times; 3 statistics: epoch-level, window mean, window SD).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Dual Ground Truth\u003c/h2\u003e \u003cp\u003eThe PerceiveImagine dataset records a vividness self-report after each trial, but the ratings in this cohort are constant across all subjects and trials (value\u0026thinsp;=\u0026thinsp;3 on a 1\u0026ndash;5 scale). Because there is no within-subject variance in vividness, we adopted a dual ground truth framework instead.\u003c/p\u003e \u003cp\u003e \u003cb\u003eObjective label.\u003c/b\u003e A LOSO-CV logistic regression classifier (L2 regularization, C\u0026thinsp;=\u0026thinsp;0.1 chosen to balance bias and variance, L-BFGS solver, StandardScaler normalization) was trained to discriminate perception from imagination using broadband LZC features. The L2 penalty mitigates overfitting despite the high feature-to-subject ratio (406 features vs. 46 subjects). For each held-out subject, the predicted probability of the \u0026ldquo;imagination\u0026rdquo; class served as a continuous objective quality metric. Per-subject AUC was computed wherever both classes were present. All 46 subjects were valid.\u003c/p\u003e \u003cp\u003e \u003cb\u003eProxy subjective label.\u003c/b\u003e Because vividness ratings lacked variance, the mean decoding probability per subject was used as a proxy. Subjects were split at the median into a high-decoding group (n\u0026thinsp;=\u0026thinsp;23) and a low-decoding group (n\u0026thinsp;=\u0026thinsp;23). This design creates a known circularity, which is discussed openly in Section IV-D.\u003c/p\u003e \u003cp\u003eA permutation baseline was also constructed for hypothesis H8: for each subject, the true test-set labels were randomly permuted 50 times, and the resulting mean AUC served as a subject-specific estimate of chance performance (approximately 0.500).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.8 Statistical Analysis\u003c/h2\u003e \u003cp\u003eEight hypotheses were tested. Cluster-based permutation testing used 1000 permutations with spatial adjacency computed from the 10\u0026ndash;20 montage via the MNE built-in channel adjacency matrix (Maris and Oostenveld \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2007\u003c/span\u003e). This non-parametric procedure controls the family-wise error rate across spatially adjacent channels. The significance threshold was α\u0026thinsp;=\u0026thinsp;0.05 throughout. The Benjamini\u0026ndash;Hochberg procedure was applied to the six hypotheses with conventional p-values (H1, H2, H3, H4, H6, H8); H5 was assessed against a pre-specified AUC threshold and H7 was untestable, so both were excluded from the correction.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH1\u003c/b\u003e: Low-gamma (30\u0026ndash;60 Hz) LZC is higher during imagination than perception at six occipital channels (Oz, O1, O2, POz, PO3, PO4). Effect size: paired Cohen\u0026rsquo;s d.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH2\u003c/b\u003e: Low-gamma LZC is higher in the high-decoding group than the low-decoding group during imagination. Effect size: independent-samples pooled Cohen\u0026rsquo;s d.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eH3\u003c/strong\u003e \u003cp\u003eFrontal HFD shows a condition \u0026times; decoding-quality interaction. Tested via linear mixed-effects model (HFD\u0026thinsp;~\u0026thinsp;condition \u0026times; y_subj_raw + (1|subject)) at five frontal channels (Fp1, Fp2, F3, Fz, F4). Because y_subj_raw is a subject-level variable rather than a trial-level covariate, the effective degrees of freedom for the interaction term are governed by the number of subjects (N\u0026thinsp;=\u0026thinsp;46), not the number of trials.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eH4\u003c/strong\u003e \u003cp\u003eBroadband LZC topography differs between perception and imagination. Tested by both cluster-based permutation and Hotelling T\u0026sup2; on PCA-reduced channel vectors (10 components). To maintain a robust observations-to-variables ratio for the multivariate test (\u003cem\u003eN\u0026thinsp;=\u0026thinsp;46\u003c/em\u003e vs. \u003cem\u003ek\u0026thinsp;=\u0026thinsp;10\u003c/em\u003e), dimensionality was reduced to the first 10 principal components, which accounted for 50.6% of the total topographic variance while capturing the predominant spatial structures.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eH5\u003c/strong\u003e \u003cp\u003eLZC features achieve AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.72 for state discrimination in the LOSO-CV classifier.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eH6\u003c/b\u003e: High-decoding subjects show higher objective accuracy than low-decoding subjects. Normality was assessed via Shapiro-Wilk; a Wilcoxon rank-sum test was applied. Effect size: rank-biserial correlation r_rb.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eH7\u003c/strong\u003e \u003cp\u003eWithin-subject Spearman correlation between vividness and decoding accuracy exceeds r\u0026thinsp;=\u0026thinsp;0.3. Not testable due to zero vividness variance.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eH8\u003c/b\u003e: Objective-label AUC exceeds the permutation-baseline AUC. Tested by paired t-test, with Bayesian analysis using a Cauchy prior (scale r\u0026thinsp;=\u0026thinsp;0.707) (Rouder et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). Robustness check: Wilcoxon signed-rank test.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Preprocessing Summary\u003c/h2\u003e \u003cp\u003eAll 46 subjects passed through the preprocessing pipeline without failure. Picard ICA converged in every case. A mean of 3.2\u0026thinsp;\u0026plusmn;\u0026thinsp;1.8 ICA components per subject were excluded (range: 1\u0026ndash;7; distribution shown in Supplementary Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). The edge-channel EMG monitor flagged fewer than 5% of imagination epochs across the cohort. After epoching, the master feature table contained 30,534 trials with 580 features each (406 LZC\u0026thinsp;+\u0026thinsp;174 HFD). Both perception and imagination conditions were represented for every subject.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Hypothesis Testing\u003c/h2\u003e \u003cp\u003eA comprehensive quantitative summary of all hypothesis tests, including test statistics, p-values, and effect sizes, is provided in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHypothesis Testing Summary\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eID\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStat.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ep\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003ep_adj\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eEffect [CI]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSig\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eES Label\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLZC imag\u0026gt;perc occ. low-γ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCluster perm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ed\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eCohen d\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLZC hi\u0026thinsp;\u0026gt;\u0026thinsp;lo decode low-γ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCluster perm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.254\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ed\u0026thinsp;=\u0026thinsp;+\u0026thinsp;0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003ePooled d\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHFD frontal\u0026times;cond\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et\u0026thinsp;=\u0026thinsp;0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.591\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.709\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;0.003\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eLME coeff\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLZC topo imag vs perc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCluster perm*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep_cl = .005\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.005\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.010\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eV\u0026thinsp;=\u0026thinsp;0.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003ePseudo-Pillai\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLZC predict state\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLOSO-CV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.811[.775,.847]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eAUC\u0026thinsp;+\u0026thinsp;CI\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHi\u0026thinsp;\u0026gt;\u0026thinsp;lo decode acc.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eWilcoxon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eU\u0026thinsp;=\u0026thinsp;2.33e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.003\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003er_rb\u0026thinsp;=\u0026thinsp;0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eRank-biserial\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVivid~decode corr.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eObj AUC\u0026thinsp;\u0026ge;\u0026thinsp;baseline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePaired t\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et\u0026thinsp;=\u0026thinsp;16.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e8.4\u0026times;10\u003csup\u003e\u0026minus;\u0026thinsp;21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.0\u0026times;10\u003csup\u003e\u0026minus;\u0026thinsp;20\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eBF\u0026thinsp;=\u0026thinsp;8.84e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eBayes Factor\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003e*H4 independently confirmed by Hotelling T\u0026sup2;: F\u0026thinsp;=\u0026thinsp;3.08, p\u0026thinsp;=\u0026thinsp;0.002, V\u0026thinsp;=\u0026thinsp;0.28. \u0026dagger;H5 significance assessed by AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.72 threshold; excluded from FDR correction. H7 untestable. p_adj: Benjamini\u0026ndash;Hochberg adjusted p-values across the six testable hypotheses (H1\u0026ndash;H4, H6, H8). All three significant findings (H4, H6, H8) survived FDR correction (p_adj\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eH1.\u003c/b\u003e The cluster-based permutation test found no significant clusters at the occipital ROI (p\u0026thinsp;=\u0026thinsp;1.0). The paired Cohen\u0026rsquo;s d was \u0026minus;\u0026thinsp;0.13, indicating a negligible effect in the direction opposite to prediction: perception showed marginally higher low-gamma LZC than imagination at occipital sites.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH2.\u003c/b\u003e No significant spatial cluster was identified in the cluster permutation test across the full 58-channel array (p\u0026thinsp;=\u0026thinsp;0.169). However, the independent-samples pooled Cohen's d for low-gamma was +\u0026thinsp;0.878 (p\u0026thinsp;=\u0026thinsp;0.005, n_high\u0026thinsp;=\u0026thinsp;23, n_low\u0026thinsp;=\u0026thinsp;23), a medium-to-large effect in the predicted direction. This dissociation between cluster permutation outcome and effect size is discussed in Section IV-A. A band-wise breakdown of effect sizes across all seven frequency bands is provided in Supplementary Fig. \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH3.\u003c/b\u003e The linear mixed-effects interaction term was not significant (t\u0026thinsp;=\u0026thinsp;0.54, p\u0026thinsp;=\u0026thinsp;0.591, β\u0026thinsp;=\u0026thinsp;0.003). Given the subject-level covariate limitation noted in Methods, this null result should be interpreted with caution.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH4.\u003c/b\u003e This was the primary positive finding at the signal level. The cluster-based permutation test found a statistically significant cluster across channels (p\u0026thinsp;=\u0026thinsp;0.005). An independent Hotelling T\u0026sup2; test applied to PCA-reduced LZC vectors showed multivariate significance (F\u0026thinsp;=\u0026thinsp;3.08, p\u0026thinsp;=\u0026thinsp;0.002, Pseudo-Pillai V\u0026thinsp;=\u0026thinsp;0.28, medium effect). Topographic analysis is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e; the spatial patterns are further detailed in Supplementary Fig. \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e and S4.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eH5.\u003c/b\u003e The LOSO-CV classifier's AUC of 0.811 (95% bootstrap CI: 0.775\u0026ndash;0.847, 2000 resamples) is significantly greater than a 0.50 chance level. All 46 subjects contributed valid test folds.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH6.\u003c/b\u003e For both groups, the Shapiro-Wilk test rejected normality. The Wilcoxon rank-sum test yielded U\u0026thinsp;=\u0026thinsp;2.33 \u0026times; 10⁸, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001, with a rank-biserial correlation of r_rb\u0026thinsp;=\u0026thinsp;0.87\u0026mdash;a very large non-parametric effect.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH7.\u003c/b\u003e This hypothesis could not be evaluated since all subjects gave a constant rating of 3 for vividness.\u003c/p\u003e \u003cp\u003e \u003cb\u003eH8.\u003c/b\u003e The paired t-test comparing real-label AUC to permutation-baseline AUC yielded t(45)\u0026thinsp;=\u0026thinsp;16.30, p\u0026thinsp;=\u0026thinsp;8.35 \u0026times; 10⁻\u0026sup2;\u0026sup1;. The Bayes Factor was BF₁₀ = 8.84 \u0026times; 10\u0026sup1;⁴ which suggests overwhelming evidence that the complexity features captured genuine neural structure. This result was confirmed by a Wilcoxon signed-rank test (W\u0026thinsp;=\u0026thinsp;1076, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). The result, as expected under the null, showed the permutation-baseline AUC was 0.500. A visual summary of the eight hypothesis outcomes is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Topographic Analysis\u003c/h2\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows group-averaged broadband LZC topographic maps from three conditions (with imagery further divided into high- and low-decoding subsets). Supplementary Fig. \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e further details these topographies across three conditions, illustrating the spatial differences when imagery is divided into high- and low-decoding subsets. Overall LZC magnitudes were similar per condition (range 0.20\u0026ndash;0.40), suggesting that the differences lie in where on the scalp the complexity concentrates, rather than the magnitude of the complexity. Regarding perception, complexity seemed to be evenly dispersed. During high-decoding imagination, increased complexity was observed in frontal regions. During low-decoding imagination, there was a prominent reduction of complexity in the right parieto-occipital region and the frontal end was weaker. The complexity reallocation on the frontal and occipital areas may be a reflection of differences in top-down imagery processing, although the scalp-recorded EEG cannot determine the cortical sources.\u003c/p\u003e \u003cp\u003eThe topographic features help us understand why H1 (which tested a general increase over occipital sites) returned a null finding, while H4 (which tested for a multivariate topographic difference across the full channel array) was significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.1 The Gamma-Band Null\u003c/h2\u003e \u003cp\u003eThe most surprising finding in this study is the contravention of the first two hypotheses. We thought the imagination condition would yield higher low-gamma LZC values at the occipital sites due to the endogenous gamma-band binding in the absence of sensory phase-locking which, in the absence of sensory inputs, should produce more diverse and, therefore, temporally complex patterns. The data indicated the opposite trend, whereby perception showed marginally higher values (d\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.13; see Supplementary Fig. \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e and the band-wise trajectory in Fig. \u003cspan refid=\"MOESM5\" class=\"InternalRef\"\u003eS5\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThere are at least three possible explanations. First, LZC quantifies broadband irregularity, which is different from phase coherence. A gamma signal can be more phase-coherent (which would reflect tighter binding) while simultaneously becoming less complex (because phase-locked signals are inherently more regular). If \u0026lsquo;binding\u0026rsquo; manifests primarily as phase coherence, then LZC is the wrong probe for detecting it. Second, it is possible that our wideband ICA pipeline removed too much neural signal. The absence of gamma-band differences\u0026mdash;despite our large cohort (\u003cem\u003eN\u0026thinsp;=\u0026thinsp;46\u003c/em\u003e)\u0026mdash;highlights the potential for myogenic bias in previous complexity studies. Our implementation of Picard ICA plus edge-channel monitoring likely sequestered these artifacts, revealing that the core neural signature of imagery is topographic rather than frequency-specific. Third, binding bursts are transient events that last only hundreds of milliseconds. Averaging LZC over a 6 seconds window will likely dilute those episodes. A time-resolved LZC analysis will be more sensitive to this averaging effect and may identify binding episodes in the first 1\u0026ndash;2 seconds of the imagination epoch when top-down reconstruction processes are highest.\u003c/p\u003e \u003cp\u003eAdditionally, the potential the gamma-band complexity holds as a measure of \u0026lsquo;binding\u0026rsquo; should be compared to that of others to come to a more accurate conclusion. Phase-coherence or phase-amplitude coupling measures may be better suited to measure the \u0026lsquo;synchronization\u0026rsquo; aspect of the phenomenon.\u003c/p\u003e \u003cp\u003eA methodologically important nuance emerges from H2. Although the cluster-based permutation test did not yield a significant spatial cluster for low-gamma LZC (p\u0026thinsp;=\u0026thinsp;0.169), the independent-samples pooled Cohen's d for low-gamma was +\u0026thinsp;0.878 (p\u0026thinsp;=\u0026thinsp;0.005, n_high\u0026thinsp;=\u0026thinsp;23, n_low\u0026thinsp;=\u0026thinsp;23)\u0026mdash;a medium-to-large effect in the predicted direction. This dissociation between the univariate effect size and the cluster permutation outcome reflects the stringency of spatial cluster correction rather than the absence of a signal. Cluster-based permutation tests require effects to be spatially contiguous across adjacent channels; a genuine effect that is diffusely distributed, or that concentrates in a small number of non-adjacent channels, can fail this threshold even when the channel-level t-statistic is individually significant. In practical terms, H2's cluster permutation null and its large positive d are not contradictory: they together suggest that high-decoding subjects show elevated low-gamma LZC relative to low-decoding subjects, but that this elevation is not spatially focal enough to survive whole-scalp cluster correction. Future analyses using region-of-interest (ROI) tests restricted to occipital or frontal channels, rather than whole-scalp cluster permutation, may provide greater sensitivity to detect this effect.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Topographic Redistribution and Predictive Coding\u003c/h2\u003e \u003cp\u003eThe H4 result tells a different story. The topography differences of broadband LZCs of frontal and occipital channels relative to perception and imagination are confirmed by the cluster permutation test (p\u0026thinsp;=\u0026thinsp;0.005) and Hotelling T\u0026sup2; (F\u0026thinsp;=\u0026thinsp;3.08, p\u0026thinsp;=\u0026thinsp;0.002, V\u0026thinsp;=\u0026thinsp;0.28, a medium multivariate effect). The topography difference is not just a simple uniform shift. This observed frontal\u0026ndash;occipital complexity gradient provides a distinct neural signature for top-down sensory reconstruction. It suggests that while perception is anchored by sensory inputs, imagery relies on high-complexity generative predictions from the frontal cortex that modulate posterior dynamics.\u003c/p\u003e \u003cp\u003eThis frontal\u0026ndash;occipital redistribution supports predictive-coding explanations for imagery. From that perspective, higher-order cortical areas form internal predictions that feedback to the sensory cortex during imagery, thus reversing the up-stream (i.e., bottom-up) information flow that characterizes perception (Chang et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Dijkstra et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The data showing increased spatial complexity in the regions of the frontal cortex and decreased complexity in the occipital cortex are in agreement with this predictive-coding model. Dijkstra et al. (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) observed that the sensory cortex is recruited during imagery with a level of feedback that is slower and more variable than during perception. Our data support that description in the region of the frontal cortex (Dijkstra et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe increase in frontal spatial complexity aligns with the activity of the medial prefrontal region of the default mode network (Rouder et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2009\u003c/span\u003e), which is known to be highly active during internally directed cognition. Scalp EEG is unable to detect the source of the activity, and the subject's attentional modulation remains a possible explanation. The fractal dimension measure has also been sensitive to emotional states (Duville et al. \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) and neurodegenerative conditions (Yoder et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), which suggests that the topographic complexity patterns observed here may generalize to other cognitive contrasts. Future work should test this redistribution using source-reconstructed EEG to determine whether the scalp-level pattern maps onto specific cortical regions. The proposed complexity-based decoding framework is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Classification and the Objective Label Advantage\u003c/h2\u003e \u003cp\u003eThe LOSO-CV classifier scored an AUC of 0.811 to discriminate perception from imagination. This is an impressive result for cross-subject EEG classification. Guenther et al. (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) found similar visual classification performance from a sparser 8-channel EEG system, and Kalafatovich et al. (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) presented a high performance visual EEG classification using spatiotemporal graph learning. By design, broad-band LZC features are agnostic to frequency, which may make them more robust to inter-subject spectral variability than narrowband power features.\u003c/p\u003e \u003cp\u003eThe result from H8 is the most compelling in this research. With true labels, the classifier was able to pick up genuine neural structure; With shuffled labels the signal entirely vanished (Permutation-baseline AUC\u0026thinsp;=\u0026thinsp;0.500). The contrast was overwhelming: t(45)\u0026thinsp;=\u0026thinsp;16.30, BF₁₀ = 8.84 \u0026times; 10\u0026sup1;⁴. There are translational implications of this finding. If complexity features can be used reliably to differentiate imagery from perception with an AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.80, then they could be used in BCI systems to provide objective measurements of imagery decodability during neurofeedback or cognitive rehabilitation. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the AUC distribution of each subject and its comparison with the permutation baseline.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eMethodologically, this study shows that wideband ICA can maintain genuine high-gamma EEG activity during cognitive tasks, without manual component inspection. The Picard ICA plus edge-channel EMG monitoring template is adoptable to study gamma-band phenomena. Future studies should study the complexity features against conventional spectral and connectivity-based EEG markers to establish their relative contribution to imagery classification.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Limitations\u003c/h2\u003e \u003cp\u003eSix limitations should be acknowledged in this study.\u003c/p\u003e \u003cp\u003eFirst, the lack of variance in the dataset\u0026rsquo;s vividness ratings rendered hypothesis H7 untestable. The proxy subjective label was derived from the objective decoding metric, creating a circular reasoning: hypotheses H2 and H6 are strengthening themselves instead of testing subjective\u0026ndash;objective dissociation. The proxy label is not an independent ground truth, as it is not a reflection of subjective experience, and may bias effect sizes upward. To test the dual label framework, datasets with truly variable self-reports are needed; or independent behavioral measures such as drawing accuracy or recognition memory.\u003c/p\u003e \u003cp\u003eSecond, the H3 linear mixed-effects model uses a subject-level covariate (mean decoding probability) as a predictor in a trial-level model. The interaction term\u0026rsquo;s effective degrees of freedom are governed by N\u0026thinsp;=\u0026thinsp;46, not the trials count. The lack of significant finding (p\u0026thinsp;=\u0026thinsp;0.591) could reflect power insufficiency, a model misspecification, or genuine absence of the effect.\u003c/p\u003e \u003cp\u003eThird, the N\u0026thinsp;=\u0026thinsp;46 sample size may be adequate for the permutation framework, but may be underpowered for LME or independent-samples comparisons for small effects. The single cultural and institutional background cohort spans a narrow age range (23\u0026ndash;30), which limits generalizability to a wider population.\u003c/p\u003e \u003cp\u003eFourth, the LZ76 median binarization discards all amplitude information from the signal. Alternative binarization strategies (Hilbert envelope, multi-symbol alphabets) might capture complementary complexity features. The HFD parameter k_max\u0026thinsp;=\u0026thinsp;64 was selected from the literature; its optimality for imagery paradigms has not been empirically evaluated.\u003c/p\u003e \u003cp\u003eFifth, the 58-channel 10\u0026ndash;20 montage imposes a hard limit on spatial resolution. The topographic differences reported here could arise from multiple distinct source configurations. Individual-MRI head models and source localization methods such as beamforming would sharpen the inference considerably.\u003c/p\u003e \u003cp\u003eSixth, time-resolved LZC could capture transient complexity changes during the first one to two seconds of the imagery epoch, when the demand for top-down reconstruction is presumably highest. This analysis would directly address the gamma-band null described in Section IV-A and is a priority for future work.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eWe tested eight hypotheses about non-linear EEG complexity as a biomarker of visual mental imagery in 46 participants. The predicted gamma-band LZC elevation did not appear. What emerged instead was a significant topographic redistribution of broadband LZC between perception and imagination (cluster p\u0026thinsp;=\u0026thinsp;0.005; Hotelling T\u0026sup2; p\u0026thinsp;=\u0026thinsp;0.002, V\u0026thinsp;=\u0026thinsp;0.28; see Supplementary Fig. \u003cspan refid=\"MOESM4\" class=\"InternalRef\"\u003eS4\u003c/span\u003e for detailed cluster topographies). Combined LZC and HFD features classified cognitive state at AUC\u0026thinsp;=\u0026thinsp;0.811 (95% CI: 0.775\u0026ndash;0.847), and objective decoding labels outperformed chance baselines with decisive evidence (BF₁₀ = 8.84 \u0026times; 10\u0026sup1;⁴). These findings reframe mental imagery as a process of large-scale topographic complexity redistribution, validating the predictive coding framework through the lens of non-linear neurodynamics.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch3\u003eAcknowledgments\u0026nbsp;\u003c/h3\u003e\n\u003cp\u003eThe authors thank Prof. Luis Miguel Pinho de Almeida for supervision and guidance throughout this project. The creators of the PerceiveImagine dataset (OpenNeuro ds005697) are acknowledged for making their data publicly available. This work was conducted within the PhD Programme in Health Data Science, Faculty of Medicine, University of Porto (FMUP), and the Department of Electrical and Computer Engineering, Faculty of Engineering, University of Porto (FEUP). No external funding was received for this research.\u003c/p\u003e\n\u003ch3\u003eFunding\u003c/h3\u003e\n\u003cp\u003eNo funds, grants, or other support was received.\u003c/p\u003e\n\u003ch3\u003eCompeting interests\u003c/h3\u003e\n\u003cp\u003eThe authors have no competing interests to declare that are relevant to the content of this article.\u003c/p\u003e\n\u003ch3\u003eEthics approval\u003c/h3\u003e\n\u003cp\u003eThe data utilized in this study were obtained from the publicly available OpenNeuro repository (accession number ds005697), which provides de-identified EEG samples. This study was deemed exempt from ethical approval as it involved the secondary analysis of fully anonymized data. The original study protocols were in accordance with the ethical standards of the original institution and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.\u003c/p\u003e\n\u003ch3\u003eConsent to participate\u003c/h3\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003ch3\u003eConsent to publish\u003c/h3\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003ch3\u003eData availability\u0026nbsp;\u003c/h3\u003e\n\u003cp\u003eThe EEG dataset \u0026quot;PerceiveImagine\u0026quot; analyzed during the current study is available in the OpenNeuro repository (https://openneuro.org/datasets/ds005697). The code pipelines are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003ch3\u003eAuthors\u0026rsquo; contribution\u003c/h3\u003e\n\u003cp\u003eYu Gao: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Project administration; Software; Visualization; Roles/Writing - original draft; Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eJos\u0026eacute; Miguel Diniz: Formal analysis; Validation; Writing - review \u0026amp; editing\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAblin P, Cardoso J-F, Gramfort A (2018) Faster ICA by preconditioning with Hessian approximations. IEEE Trans Signal Process 66:4040\u0026ndash;4049. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/TSP.2018.2844203\u003c/span\u003e\u003cspan address=\"10.1109/TSP.2018.2844203\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArmonaite K, Nobili L, Paulon L, Balsi M, Conti L, Tecchio F (2023) Local neurodynamics as a signature of cortical areas: new insights from sleep. Cereb Cortex 33(6):3284\u0026ndash;3292. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/cercor/bhac274\u003c/span\u003e\u003cspan address=\"10.1093/cercor/bhac274\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAggarwal S, Ray S (2025) Changes in Higuchi fractal dimension across age in healthy human EEG. Eur J Neurosci 62:e70193. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/ejn.70193\u003c/span\u003e\u003cspan address=\"10.1111/ejn.70193\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang YH, Chen HA, Tsai MJ, Tseng CL, Lo CH, Huang KC, Wei CS (2025) A human EEG dataset for multisensory perception and mental imagery. Sci Data 12:1598. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41597-025-05881-1\u003c/span\u003e\u003cspan address=\"10.1038/s41597-025-05881-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eColussi F, Favaro J, Ancona C, Passarotto E, Pelizza MF, Lorenzon E, Ruzzante S, Masiero S, Perilongo G, Sparacino G, Toldo I, Sartori S, Rubega M (2025) EEG difference in the Higuchi fractal dimension of wakefulness and sleep from birth to adolescence. PLoS ONE 20:e0333903. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0333903\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0333903\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuville MM, Alonso-Valerdi LM, Ibarra-Zarate DI (2022) Fractal dimension of EEG signals and heart dynamics in discrete emotional states. Front Comput Neurosci 16:1022787. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fncom.2022.1022787\u003c/span\u003e\u003cspan address=\"10.3389/fncom.2022.1022787\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDijkstra N, Mostert P, Lange FP, Bosch S, van Gerven MAJ (2021) Differential temporal dynamics during visual imagery and perception. eLife 10:e47892. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.7554/eLife.47892\u003c/span\u003e\u003cspan address=\"10.7554/eLife.47892\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGifford AT, Dwivedi K, Roig G, Cichy RM (2022) A large and rich EEG dataset for modeling human visual object recognition. NeuroImage 264:119754. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.neuroimage.2022.119754\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2022.119754\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuenther S, Kosmyna N, Maes P (2024) Image classification and reconstruction from low-density EEG. Sci Rep 14:16436. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-024-66228-1\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-66228-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH\u0026ouml;hn C, Hahn MA, Lendner JD, Hoedlmoser K (2024) Spectral slope and Lempel\u0026ndash;Ziv complexity as robust markers of brain states during sleep and wakefulness. eNeuro 11(3). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1523/ENEURO.0259-23.2024\u003c/span\u003e\u003cspan address=\"10.1523/ENEURO.0259-23.2024\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKalafatovich J, Lee M, Lee SW (2023) Learning spatiotemporal graph representations for visual perception using EEG. IEEE Trans Neural Syst Rehabil Eng 31:97. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/TNSRE.2022.3217344\u003c/span\u003e\u003cspan address=\"10.1109/TNSRE.2022.3217344\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi X, Fan Y (2024) PerceiveImagine: a human EEG dataset for studying perception and mental imagery. Sci Data 11:681. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41597-024-03529-6\u003c/span\u003e\u003cspan address=\"10.1038/s41597-024-03529-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee S, Jang S, Jun SC (2022) Exploring the ability to classify visual perception and visual imagery EEG data. Electronics 11:2706. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/electronics11172706\u003c/span\u003e\u003cspan address=\"10.3390/electronics11172706\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMedel V, Irani M, Crossley N, Ossand\u0026oacute;n T, Boncompte G (2023) Complexity and 1/f slope jointly reflect brain states. Sci Rep 13:21870. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-023-47316-0\u003c/span\u003e\u003cspan address=\"10.1038/s41598-023-47316-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaris E, Oostenveld R (2007) Nonparametric statistical testing of EEG- and MEG-data. J Neurosci Methods 164(1):177\u0026ndash;190. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jneumeth.2007.03.024\u003c/span\u003e\u003cspan address=\"10.1016/j.jneumeth.2007.03.024\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOlejarczyk E, Gotman J, Frauscher B (2022) Region-specific complexity of the intracranial EEG in the sleeping human brain. Sci Rep 12:451. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-021-04213-8\u003c/span\u003e\u003cspan address=\"10.1038/s41598-021-04213-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRuiz de Miras J, Ib\u0026aacute;\u0026ntilde;ez-Molina AJ, Soriano MF, Iglesias-Parro S (2023) Fractal dimension analysis of resting state functional networks in schizophrenia. Front Hum Neurosci 17:1236832. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fnhum.2023.1236832\u003c/span\u003e\u003cspan address=\"10.3389/fnhum.2023.1236832\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRouder JN, Speckman PL, Sun D, Morey RD, Iverson G (2009) Bayesian t tests for accepting and rejecting the null hypothesis. Psychon Bull Rev 16(2):225\u0026ndash;237. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3758/PBR.16.2.225\u003c/span\u003e\u003cspan address=\"10.3758/PBR.16.2.225\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRen Z, Yue M, Han X, Zhao Z, Wang B, Hong Y, Zhao T, Wang N, Zhao P, Hong Y, Wang Q, Zhao Y (2023) The potential of the Lempel\u0026ndash;Ziv complexity of the EEG in diagnosing cognitive impairment. Epileptic Disord 25(3):331\u0026ndash;342. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/epd2.20044\u003c/span\u003e\u003cspan address=\"10.1002/epd2.20044\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShimizu H, Srinivasan R (2022) Improving classification and reconstruction of imagined images from EEG. PLoS ONE 17:e0274847. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0274847\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0274847\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilson H, Chen X, Golbabaee M, Proulx MJ, O'Neill E (2024) Feasibility of decoding visual information from EEG. Brain-Comput Interfaces 11(1\u0026ndash;2):33\u0026ndash;60. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/2326263X.2023.2287719\u003c/span\u003e\u003cspan address=\"10.1080/2326263X.2023.2287719\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoder KJ, Brookshire G, Glatt R, Merrill DA, Gerrol S, Quirk C et al (2024) Fractal dimension distributions of resting-state EEG improve detection of dementia. Fractal Fract 8:27. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/fractalfract8030027\u003c/span\u003e\u003cspan address=\"10.3390/fractalfract8030027\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"brain-topography","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"btop","sideBox":"Learn more about [Brain Topography](http://link.springer.com/journal/10548)","snPcode":"10548","submissionUrl":"https://submission.nature.com/new-submission/10548/3","title":"Brain Topography","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"EEG, mental imagery, biomarker, Lempel-Ziv complexity, Higuchi fractal dimension, predictive coding, brain–computer interface","lastPublishedDoi":"10.21203/rs.3.rs-9382178/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9382178/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eVisual mental imagery is the process of reconstructing perceptual experience without sensory input. How the brain performs this process is poorly understood, particularly from the perspective of conventional linear EEG analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eObjective: \u003c/strong\u003eThis study aims to evaluate if the two non-linear EEG complexity measures—Lempel-Ziv Complexity (LZC) and Higuchi Fractal Dimension (HFD)—can differentiate between perception and imagination and if they can be used as objective indices of neural separability of mental imagery.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eLZC and HFD were extracted from 58 scalp EEG channels in 46 healthy adults performing the PerceiveImagine paradigm (Li and Fan 2024), after Wideband Picard ICA (1–200 Hz) removing artefacts. Statistical analyses included cluster-based permutation testing (Maris and Oostenveld 2007), Hotelling T², and leave-one-subject-out cross-validation (LOSO-CV).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eBroadband LZC topography differed between perception and imagination (cluster p = 0.005; Hotelling F = 3.08, p = 0.002, V = 0.28). LOSO-CV classification reached AUC = 0.811 (95% CI: 0.775–0.847). The classifier’s AUC exceeded a label-permuted baseline by a wide margin (t(45) = 16.30, p = 8.35 × 10⁻²¹, BF₁₀ = 8.84 × 10¹⁴). No gamma-band LZC differences were found (p = 1.0, d = −0.13).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSignificance: \u003c/strong\u003eEEG complexity features reveal a fundamental topographic redistribution principle of cortical dynamics: imagery is characterized by a shift from sensory-driven posterior irregularity to generative frontal complexity, rather than a global magnitude change. Objective decoding labels outperform self-report as a training signal for imagination quality classifiers.\u003c/p\u003e","manuscriptTitle":"Neural Complexity Signatures of Visual Mental Imagery: Lempel-Ziv Complexity and Higuchi Fractal Dimension Reveal Topographic Reorganization in High-Density EEG","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-24 15:46:20","doi":"10.21203/rs.3.rs-9382178/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-04-22T08:28:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"101985889856894117437507565373049381355","date":"2026-04-19T18:09:22+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-17T08:02:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"40721087345369688309674198302192911187","date":"2026-04-17T06:06:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"303159386484368538179177148933311971797","date":"2026-04-17T03:44:10+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-17T03:42:51+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-13T06:00:20+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-13T05:10:11+00:00","index":"","fulltext":""},{"type":"submitted","content":"Brain Topography","date":"2026-04-10T17:20:46+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"brain-topography","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"btop","sideBox":"Learn more about [Brain Topography](http://link.springer.com/journal/10548)","snPcode":"10548","submissionUrl":"https://submission.nature.com/new-submission/10548/3","title":"Brain Topography","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"b49f5930-f5c3-48f4-9f23-914404bc2fe4","owner":[],"postedDate":"April 24th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-24T15:46:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-24 15:46:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9382178","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9382178","identity":"rs-9382178","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}