Abstract

Background:

Human cognition is derived from functional cortical long-range connectivity, as reflected by phase synchronization between electrode pairs of low-frequency electroencephalographic (EEG) activity <8 Hz related to cognitive tasks.

Methods:

We tested the hypothesis that such an EEG marker, combined with machine learning, can discriminate between Parkinson’s disease (PD) with mild cognitive impairment (MCI) and dementia (D) and those with dementia with Lewy bodies (DLB). Event-related EEG delta (1–3.5 Hz) and theta (4–7 Hz) phase coherence were computed from EEG activity recorded during a visual oddball task in healthy controls (HC, N = 24) and PD-MCI (N = 20), PDD (N = 18), and DLB (N = 11) patients.

Results:

Using delta-band coherence as input, the comparison between HC and PD-MCI yielded an AUC of approximately 0.79 and an accuracy of 86.4%. Higher discriminative performance was observed for HC versus PDD, reaching an AUC near 0.92 with an overall accuracy of 94.6%. In the classification of HC versus DLB participants, the model achieved 83.3% sensitivity and 88.9% specificity, with an AUC around 0.77. Theta-band models showed comparable results, with average AUC values of about 0.75 for HC vs. DLB and slightly above 0.80 for HC vs. PDD, while classification of HC vs. PD-MCI remained in a moderate range.

Conclusion:

These findings suggest that event-related EEG phase coherence at <8 Hz is a promising EEG correlate of cognitive deficits in patients with PDD and DLB, offering insights into disrupted network dynamics of cortical activity related to cognitive processes and potential biomarkers for testing new drugs for cognitive enhancement and disease monitoring.

1 Introduction

Neurodegenerative diseases constitute an increasing global health burden as life expectancy rises, particularly in developed countries. Dementia currently affects approximately 55 million individuals worldwide, with annual global costs surpassing USD 1.3 trillion. Parkinson’s disease (PD) is among the fastest growing neurodegenerative disorders, affecting an estimated 11.77 million individuals globally in 2021 (Li et al., 2025). Cognitive impairment and dementia represent major contributors to disability, reduced quality of life, institutionalization, and mortality in PD. Within the broader dementia spectrum, dementia with Lewy bodies (DLB), characterized by widespread α-synuclein pathology, accounts for approximately 5–8% of all dementia cases.

Despite the substantial clinical and societal impact, dementia in Parkinsonian and Lewy body spectrum disorders remains under-recognized (Alzheimer’s Association, 2025; Ratnavel et al., 2025), and early differentiation between cognitive phenotypes—such as PD with mild cognitive impairment (PD-MCI), Parkinson’s disease dementia (PDD), and DLB—remains challenging in routine clinical practice. Delayed or inaccurate phenotypic classification limits timely intervention, complicates clinical decision-making, and increases caregiver and health-care burden. These challenges highlight the need for accessible, objective, and scalable biomarkers that can be feasibly deployed in outpatient settings to support early diagnosis, phenotypic stratification, and longitudinal monitoring of disease progression and treatment response. Electroencephalography (EEG) is particularly well suited for clinical application, offering a direct and temporally precise assessment of large-scale neural dynamics with minimal patient burden and wide availability.

The human brain orchestrates large-scale information processing across distributed networks through the dynamic synchronization of neural oscillations (Babiloni et al., 2004; Srinivasan et al., 2007). During cognitive task engagement, this coordination creates transient temporal windows that facilitate neural signal integration and is reflected in phase synchronization between electroencephalographic (EEG) signals recorded at spatially distributed electrode pairs, particularly in low-frequency bands (<8 Hz). Event-related (ER) delta (1–3.5/4 Hz) and theta (4–7 Hz) oscillations play a critical role in attention, memory, and executive functions by supporting long-range interregional communication (Başar et al., 1999; Klimesch, 1999). Accordingly, EEG phase coherence was selected as the primary analytic marker in the present study, as it directly quantifies coordination between distributed brain regions and provides interpretable signatures of large-scale neural communication that extend beyond local signal amplitude alone (Haufe et al., 2014).

In neurodegenerative disorders, cognitive decline is increasingly conceptualized as a disconnection syndrome, in which progressive disruption of large-scale network communication underlies cognitive dysfunction. Consistent with this framework, growing evidence indicates that disruptions in delta and theta phase synchronization may reflect cortical functional dysconnectivity underlying cognitive decline in neurodegenerative diseases such as PD at the stage of PD-MCI, PDD, and DLB (Güntekin and Başar, 2016; Yener et al., 2019). For example Güntekin et al. (2019) reported graded reductions in theta inter-trial phase coherence and spectral power during auditory and visual oddball target processing, with the lowest values in PDD, intermediate reductions in PD-MCI, and highest values in cognitively unimpaired PD, most prominently at frontal–central electrode sites with a right-hemisphere predominance.

In another study, Fide et al. (2023) reported reduced delta, theta, and alpha intrahemispheric and midline coherence during visual oddball target processing in patients with Alzheimer’s disease (AD) relative to healthy controls (HC), whereas individuals with amnestic mild cognitive impairment (aMCI) exhibited increased coherence, interpreted as reflecting compensatory or potentially maladaptive network reorganization at an early disease stage. Furthermore, Hünerli-Gündüz et al. (2023) reported reduced delta power and phase-locking values (PLVs) in patients with PD-MCI during an auditory oddball paradigm, and proposed that these reductions are associated with structural atrophy of the thalamus, putamen, and hippocampus, as assessed by magnetic resonance imaging (MRI). In agreement, Güntekin et al. (2021), also found that both delta and theta PLV were significantly reduced during auditory oddball paradigm in patients with PD-D, Alzheimer’s disease with dementia (ADD), DLB, and vascular cognitive impairment (VCI) compared to HC.

Although prior findings are informative, they remain insufficient for clinical assessment of cognitive deficits in PD and DLB. To address this gap, we implemented a structured and interpretable machine-learning pipeline designed for both interpretability and predictive reliability. Phase-locking values (PLVs) in the delta and theta bands were extracted as connectivity features (Ruiz-Gómez et al., 2019; Cao et al., 2025), and Bolasso—a bootstrapped extension of LASSO combining bootstrap resampling with L1 regularization (Bach, 2008; Yang et al., 2023; Tibshirani, 1996)—was applied to enhance feature stability and limit overfitting in high-dimensional EEG spaces with modest sample sizes. Regularized linear models impose sparsity, improve generalization, and enable interpretable feature selection in neural data (Abraham et al., 2014); accordingly, LASSO-based logistic regression provides a robust framework for EEG-based neurological classification, with recent clinical EEG studies demonstrating its predictive utility (Mao et al., 2025). Finally, supervised classification analyses (e.g., HC vs. PDD) were performed using 10-fold cross-validation and regularization to ensure robust and generalizable performance. Additionally, our approach aligns with recent trends favoring explainable and lightweight EEG models, such as xEEGNet (Zanola et al., 2025).

To our knowledge, no prior study has combined wavelet-based EEG phase-connectivity metrics with ML classifiers to differentiate healthy controls, PD-MCI, PDD, and DLB within a single task-based analytic framework. The present work integrates task performance (oddball error scores) and global cognitive status (MMSE) alongside edge- and network-level EEG features, enabling coherence alterations to be interpreted in relation to both behavioral performance and clinical cognitive status within a rigorously defined and diagnostically homogeneous cohort. This multimodal and internally consistent framework strengthens biological interpretability and distinguishes the present study from prior oddball or EEG connectivity investigations, which have typically examined Parkinsonian and Lewy body disorders separately or focused on a single analytic level. At the same time, translation of EEG connectivity measures into clinically meaningful outcomes remains limited. Existing findings in Lewy body– and Parkinsonian-spectrum disorders are difficult to compare due to substantial heterogeneity (di Biase et al., 2023; Kucikova et al., 2024; Law et al., 2020; van der Zande et al., 2018) in acquisition protocols and analysis pipelines, and connectivity metrics remain less consistently standardized and validated than traditional spectral EEG markers in comparative studies. Against this background, the present study provides a systematic and integrative step toward bridging EEG connectivity measures with clinically relevant cognitive phenotypes.

To address this gap, the present study specifically tested the hypothesis that:

(1) wavelet-based event-related delta and theta phase coherence between electrode pairs would progressively decrease with increasing cognitive impairment across PD-MCI, PDD, and DLB relative to HC, with the most pronounced disruptions observed in dementia-stage patients; and

(2) these connectivity measures would support individual-level classification, yielding higher discriminative accuracy for HC vs. PDD and HC vs. DLB comparisons than for HC vs. PD-MCI.

2 Materials and methods2.1 Participants2.1.1 Recruitment

Demographic characteristics for all groups are summarized in Table 1. Groups did not differ significantly in age or years of education (ANOVA, p > 0.05). Mean age (±SD) was 61.3 ± 7.95 years for HC, 67.5 ± 8.59 for PD-MCI, 70.2 ± 7.41 for PDD, and 70.3 ± 9.27 for DLB participants. Mean years of education ranged from 6.7 to 10.0 years across groups.

VariableHCPD-MCIPDDDLBN (Total = 73)24201811Sex (M: F)13:1115:517:18:3Age (Mean ± SD)61.3 ± 7.9567.5 ± 8.5970.2 ± 7.4170.3 ± 9.27Education (Mean ± SD)10.0 ± 4.426.70 ± 2.397.0 ± 3.667.55 ± 3.80

Demographics across groups.

*No significant differences were found across groups for age or education (ANOVA, p > 0.05). Age was included as a covariate in secondary analyses to mitigate potential confounding.

Disease duration was not consistently available across all participants and was therefore not included as a descriptive or analytical variable.

A total of 73 participants (aged 50–80) were recruited from the Neurology Outpatient Clinic of Istanbul Medipol University Hospital and assigned to four groups: PD-MCI (N = 20, 15 male), PDD (N = 18, 17 male), DLB (N = 11, 8 male), and age-matched healthy controls (N = 24, 13 male). All of our patients were on dopaminergic medication; they had their daily medication just 1 h before the EEG recordings. They had normal or corrected vision, no reported hearing impairments, and were naive to the task. The study was approved by the Istanbul Medipol University Ethics Committee (Approval No: 10840098-51, E30217-E30218) and was conducted in accordance with the Declaration of Helsinki. Informed consent was obtained from all participants or their caregivers before the study commenced. Participants did not receive financial compensation for their participation.

2.1.2 Inclusion and exclusion criteria

PDD diagnosis followed Movement Disorder Society criteria (Emre et al., 2007) and UK Brain Bank guidelines (Georgiev et al., 2015), with MMSE < 26 (Güngen et al., 2002), CDR ≥ 0.5, and functional impairment ≥1.5 SD below norms (Huebl et al., 2014). PD-MCI was defined by ≥1.5 SD cognitive impairment in ≥2 tests within one domain. DLB diagnosis followed McKeith et al. (2005) criteria, with MMSE < 26 and functional impairment ≥1.5 SD. Healthy controls scored ≥26 on the MMSE, no neurological/psychiatric history, and no chronic neuroactive medication use. Exclusion criteria included other dementias, cognition-altering drugs, substance abuse, stroke, TBI, epilepsy, psychiatric/neurological conditions, GDS > 13, unstable medical conditions, and MRI abnormalities. Groups were age-matched with no significant differences (p > 0.05) (Podcasy and Epperson, 2016; Walker et al., 2015).

2.2 Task design and procedure

Each participant completed three parts in a single session (Figure 1): (1) neuropsychological evaluation, (2) resting-state EEG, and (3) task-based EEG during a visual oddball paradigm.

End-to-end workflow of the study. Overview of the experimental and analytical pipeline organized into four stages: (1) Clinical and neuropsychological assessment, baseline EEG recording, and visual oddball task structure and stimulus timeline; (2) task-based EEG acquisition during target and non-target processing; (3) EEG preprocessing, wavelet-based time–frequency decomposition, and phase-locking value (PLV) connectivity estimation across inter- and intrahemispheric electrode pairs; and (4) multivariate feature selection, network characterization, low-dimensional embedding, and regularized classification with cross-validated and permutation-based performance evaluation. The numbered layout highlights the integrative and hypothesis-driven nature of the pipeline, spanning task-evoked neurophysiology, network analysis, and predictive modeling across Parkinsonian dementia subtypes.

2.2.1 Neuropsychological evaluation

Cognitive assessment included the Turkish MMSE (max score = 30) (Güngen et al., 2002) to assess global cognition; the Öktem Verbal Memory Processes Test (OVMPT) for verbal memory (Tanör, 2011); semantic and phonemic Verbal Fluency Tests (Crawford et al., 2018); the Clinical Dementia Rating Scale (CDR) for dementia severity (Morris, 1993, 1997); and the 30-item Geriatric Depression Scale (GDS) for mood screening (Yesavage et al., 1982).

2.2.2 Task procedure

The task was implemented in E-Prime (Psychology Software Tools Inc., Pittsburgh, PA) to assess neural responses to target and nontarget stimuli. Stimuli were presented at two luminance levels—10 cd/m2 (nontarget) and 40 cd/m2 (target)—on a 22-inch monitor (75 Hz, 10 ms rise time, 1,000 ms duration). Each trial consisted of 120 stimuli included 40 targets (33.3%) and 80 nontargets (66.6%) in randomized order that was independently generated for each participant, and lasted approximately 12 min, with a variable interstimulus interval of 3–7 s to reduce anticipatory effects. Participants were seated approximately 100 cm from the monitor at eye level in a dimly lit, sound-attenuated room to ensure visual comfort and stable viewing conditions. They were instructed to silently count the target stimuli while ignoring nontargets and to report the total number of targets post-trial. A sample session confirmed task comprehension. Behavioral performance was quantified as the absolute counting error relative to the correct target number (i.e., |40 – response|).

A schematic overview of the task structure and stimulus timeline is provided in Figure 1, illustrating the luminance-defined stimuli, stimulus duration, and interstimulus interval.

2.3 EEG acquisition and preprocessing2.3.1 Recording

EEG signals were recorded using a 32-channel BrainCap with Multitrodes (EasyCap GmbH, Germany), with electrodes placed according to the international 10–20 system. Electrode sites included: Fp1, Fp2, F7, F3, Fz, F4, F8, FT7, FC3, FCz, FC4, FT8, Cz, C3, C4, T7, T8, TP7, CP3, CPz, CP4, TP8, P3, Pz, P4, P7, P8, O1, Oz, and O2. Linked Ag/AgCl electrodes on the earlobes (A1 + A2) served as references, and EOG was recorded from the medial upper and lateral orbital rim of the left eye. Electrode impedance was kept below 10 kΩ. Signals were amplified using a BrainAmp MR Plus 32-channel DC system (Brain Products GmbH, Germany), bandpass filtered between 0.01–250 Hz, and sampled at 500 Hz. Recordings were performed in a dimly lit, shielded room to minimize external interference.

2.3.2 Preprocessing

Event-related EEG data were preprocessed in BrainVision Analyzer (BVA). Signals were downsampled to 256 Hz, band-pass filtered (0.01–60 Hz), and re-referenced to linked mastoids. Ocular artifacts were removed using independent component analysis (ICA). Data were segmented into 6 s epochs (−3 to +3 s) for delta and 2 s epochs (−1 to +1 s) for theta, separately for target and non-target trials. Manual artifact rejection was performed to ensure clean data for time-frequency and connectivity analyses.

After ICA and manual artifact rejection, an average of 17.6 ± 2.9 (HC), 15.1 ± 2.5 (PDD), 14.9 ± 1.8 (PD-MCI), and 14.3 ± 2.3 (DLB) target epochs, as well as 19.7 ± 3.2 (HC), 16.5 ± 2.7 (PDD), 18.6 ± 2.8 (PD-MCI), and 16.5 ± 2.6 (DLB) non-target epochs per participant, were retained for subsequent connectivity analyses.

2.4 EEG analysis2.4.1 Event-related EEG

Time-frequency analyses were conducted using the Gabor-normalized Complex Morlet Wavelet Transform (WT). A fixed wavelet width of three cycles was applied at each center frequency, yielding a temporal standard deviation of σₜ = 3 / (2πf) and a corresponding frequency-domain standard deviation of σ? = f / 3 across 30 logarithmically spaced bins for delta (1–3.5 Hz) and theta (4–7 Hz) bands. Prior to WT, Current Source Density (CSD) transformation was applied to reduce volume conduction (Spline order: 4; Legendre degree: 10; Lambda: 1e–5) (Thatcher et al., 1986). Phase-based connectivity was quantified using PLV, calculated from WT phase outputs, separately for target and non-target conditions without baseline normalization. As defined by Lachaux et al. (1999, 2002), with implementation consistent with contemporary time–frequency approaches (Cohen, 2014). Instantaneous phase estimates were obtained from the complex Morlet wavelet transform (Addison, 2002; Bostanov, 2004), and PLV (PLV; range 0–1) was computed for each electrode pair as the magnitude of the average unit-length phase difference vectors across trials, quantifying functional connectivity between channels. Higher PLV indicates more stable phase relationships across trials. This metric is relatively robust to amplitude fluctuations and volume-conduction artifacts when combined with spatial filtering (i.e., CSD). Delta-band PLV was averaged (E1) over 0–600 ms, and theta over 0–300 ms post-stimulus, across all electrode pairs. Outliers were excluded using the IQR method (values beyond Q1–3*IQR or Q3 + 3*IQR), applied separately to coherence, neuropsychological, and behavioral data.

E1 (Phase Locking Value)where φₓ, trial(f,t) and φᵧ, trial(f,t) are the instantaneous phase values of two signals. The PLV was calculated separately for the delta (1–3.5 Hz) and theta (4–7 Hz) bands.

2.5 Statistical analysis

All statistical and machine learning analyses were conducted in R version 4.2.1 (2022-06-23).

2.5.1 Behavioral and neuropsychological analysis

Task performance was measured by error scores—incorrect responses to target stimuli—reflecting engagement. Although a full neuropsychological battery was administered, MMSE was used as the primary cognitive measure due to its clinical relevance, validation across dementia subtypes, and consistency across groups (Kamarajan and Porjesz, 2015; Rossini et al., 2006; Handayani et al., 2018). Linear regression models were run per group to examine associations between MMSE, error scores, and coherence. To assess whether coherence–behavior relationships were independent of demographic confounds, additional multiple regression models were performed including age as a covariate. Subject-level weighted average coherence during target trials served as single coherence predictor per participant. p-values were computed and corrected for multiple comparisons using the FDR. Standardized regression coefficients (β) reflect effect sizes derived from models fit at the electrode-pair level (Table 2).

MeasureHCPD-MCIPDDDLBANOVA-statisticsPost-hoc comparisonsSensitivity analysis [CI (lower, upper)]Delta phase valuesInterhemispheric0.322 (0.067)0.310 (0.042)0.282 (0.040)0.292 (0.032)F = 2.283, p = 0.087, η2p = 0.090HC–PD-MCI (p = 0.048*, g = 26)HC–PD-MCI [CI (−0.016, 0.044), p = 0.038*]Left hemisphere0.316 (0.060)0.322 (0.042)0.303 (0.041)0.309 (0.028)HC–PDD (p = 0.035*, g = 0.73)HC–PDD [CI (0.007, 0.067), p = 0.029*]Right hemisphere0.346 (0.057)0.313 (0.041)0.282 (0.032)0.292 (0.022)HC–DLB (p = 0.049*, g = 0.54)HC–DLB [CI (−0.005, 0.055), p = 0.032*]MMSERegression-statisticsWithin-group descriptivesInterhemispheric0.358 (0.159)0.322 (0.076)0.315 (0.082)0.298 (0.076)R2 = 0.395, β = 0.061, p < 0.01*HC (R2 = 0.178), PD-MCI (R2 = 0.23), PDD (R2 = 0.56*), DLB (R2 = 0.192)PD-MCI [−5.54, −4.56], PDD [−7.31, −6.30], DLB [−7.17, −5.97]Intrahemispheric0.409 (0.139)0.329 (0.093)0.299 (0.080)0.332 (0.075)R2 = 0.425, β = 0.088, p < 0.001*HC (R2 = 0.101), PD-MCI (R2 = 0.168), PDD (R2 = 0.47*), DLB (R2 = 0.35*)PD-MCI [−5.46, −4.48], PDD [−7.32, −6.32], DLB [−7.08, −5.88]Error scoresInterhemispheric0.359 (0.132)0.313 (0.075)0.299 (0.096)0.319 (0.080)R2 = 0.56, β = −0.080, p < 0.001*HC (R2 = 0.198), PD-MCI (R2 = 0.275*), PDD (R2 = 0.495*), DLB (R2 = 0.315*)PD-MCI [1.63, 2.39], PDD [6.11, 6.88], DLB [7.44, 8.36]Intrahemispheric0.371 (0.115)0.343 (0.102)0.310 (0.086)0.306 (0.066)R2 = 0.68, β = −0.155, p < 0.001*HC (R2 = 0.133), PD-MCI (R2 = 0.330*), PDD (R2 = 0.670*), DLB (R2 = 0.298*)PD-MCI [1.61, 2.36], PDD [6.12, 6.90], DLB [7.44, 8.36]Theta phase valuesInterhemispheric0.292 (0.062)0.269 (0.051)0.288 (0.045)0.279 (0.029)F = 1.188, p = 0.321, η2p = 0.040HC–PD-MCI (p > 0.05, g = 0.22)HC–PD-MCI [CI (−0.006, 0.041), p = 0.048*]Left hemisphere0.295 (0.055)0.301 (0.049)0.282 (0.042)0.308 (0.032)HC–PDD (p = 0.042*, g = 0.43)HC–PDD [CI (−0.004, 0.043), p = 0.029*]Right hemisphere0.292 (0.052)0.269 (0.051)0.288 (0.045)0.279 (0.028)HC–DLB (p = 0.048*, g = 0.30)HC–DLB [CI (−0.045, 0.038), p = 0.032*]MMSERegression-statisticsWithin-group descriptivesInterhemispheric0.324 (0.116)0.275 (0.052)0.280 (0.070)0.310 (0.083)R2 = 0.289, β = 0.057, p < 0.05*HC (R2 = 0.25), PD-MCI (R2 = 0.32*), PDD (R2 = 0.44*), DLB (R2 = 0.23)PD-MCI [−4.56, −3.51], PDD [−6.34, −5.26], DLB [−6.35, −5.07]Intrahemispheric0.350 (0.120)0.310 (0.14)0.290 (0.12)0.310 (0.11)R2 = 0.408, β = 0.068, p < 0.01*HC (R2 = 0.20), PD-MCI (R2 = 0.37*), PDD (R2 = 0.43*), DLB (R2 = 0.30*)PD-MCI [−5.46, −4.49], PDD [−7.32, −6.32], DLB [−7.12, −5.92]Error scoresInterhemispheric0.290 (0.10)0.140 (0.27)0.129 (0.080)0.280 (0.07)R2 = 0.44, β = −0.078, p < 0.05*HC (R2 = 0.18), PD-MCI (R2 = 0.38*), PDD (R2 = 0.42*), DLB (R2 = 0.34*)PD-MCI [1.63, 2.38], PDD [6.11, 6.88], DLB [7.37, 8.29]Intrahemispheric0.310 (0.06)0.290 (0.07)0.310 (0.06)0.290 (0.07)R2 = 0.56, β = −0.86, p < 0.01*HC (R2 = 0.19), PD-MCI (R2 = 0.33*), PDD (R2 = 0.44*), DLB (R2 = 0.32*)PD-MCI [1.61, 2.36], PDD [6.12, 6.90], DLB [7.44, 8.36]

Statistics for the delta and theta band across groups [mean (SD)] and hemispheres, and sensitivity analysis.

The omnibus ANOVA for delta and theta did not reach significance at the conventional threshold [F(3, 69) = 2.283, p = 0.087; F(3, 69) = 1.188, p = 0.321]. Nevertheless, given our a priori expectation of reduced connectivity in patient groups relative to HC, we conducted planned pairwise contrasts (p-values FDR-corrected). These results are consistent with our a priori hypothesis; however, as the omnibus tests were not significant, they should be interpreted cautiously as exploratory. Effect sizes for planned post-hoc contrasts are reported as bias-corrected standardized mean differences (Hedges’ g) computed from model-estimated marginal means (EMMs).

CI = 95% confidence interval (regression coefficients).

2.5.2 Signal analysis

A mixed-design repeated measures ANOVA was conducted to assess differences in EEG coherence across groups (HC, PDD, PD-MCI, DLB), with Group as a between-subjects factor, and Electrode Pair (40 pairs), Frequency Band (Delta, Theta), Stimulus Type (Target, Nontarget), and Hemisphere (Left, Right, Inter) as within-subjects factors (R package: lme4). The 40 electrode pairs included: F3-T7, F3-T8, F4-T7, F4-T8, F3-P3, F3-P4, F4-P3, F4-P4, F3-TP7, F3-TP8, F4-TP7, F4-TP8, F3-P7, F3-P8, F4-P7, F4-P8, F3-O1, F3-O2, F4-O1, F4-O2, C3-T7, C3-T8, C4-T7, C4-T8, C3-P3, C3-P4, C4-P3, C4-P4, C3-TP7, C3-TP8, C4-TP7, C4-TP8, C3-P7, C3-P8, C4-P7, C4-P8, C3-O1, C3-O2, C4-O1, C4-O2. Post-hoc pairwise comparisons were FDR-corrected (p < 0.05), and Greenhouse–Geisser corrections were applied for sphericity violations. Effect sizes for planned post-hoc contrasts are reported as bias-corrected standardized mean differences (Hedges’ g) computed from model-estimated marginal means (EMMs).

To limit the multiple-comparison burden inherent in EEG connectivity analyses, PLV estimates were restricted to a predefined set of 40 electrode pairs selected a priori based on their relevance to large-scale cognitive networks (e.g., fronto–temporal, fronto–parietal, centro–parietal, and temporo–occipital connections). This hypothesis-driven selection substantially reduced dimensionality and avoided mass-univariate testing across thousands of sensor pairs, which can severely reduce statistical power when naïve corrections are applied (Li et al., 2025).

PLV was analyzed at multiple, explicitly defined aggregation levels, depending on the statistical objective. At the edge level, PLV was computed separately for each of the 40 predefined electrode pairs, and these pairwise values were used directly in electrode-pair–level analyses, including bootstrap/permutation testing, feature selection, and supervised classification. At the hemisphere level, PLV values were aggregated within participants by averaging across all eligible electrode pairs belonging to a given category (left intrahemispheric, right intrahemispheric, or interhemispheric), yielding a single summary coherence value per category, frequency band, and participant. These aggregated measures were used in mixed-design ANOVA, regression analyses linking coherence to behavioral and neuropsychological measures, and descriptive group comparisons. At the network level, weighted adjacency matrices constructed from pairwise PLV values were used to compute graph-theoretic metrics (global efficiency, local efficiency, and small-worldness) without prior averaging, preserving the topological structure of functional connectivity. This multi-level strategy allowed us to combine fine-grained edge-level sensitivity with interpretable hemisphere- and network-level summaries while maintaining statistical consistency across analyses.

2.6 Functional network characterization

To assess group-level separability, we constructed a 3D coherence feature embedding using the most discriminative electrode pairs (delta and theta bands) selected via LASSO or regression weighting. Each participant was projected into this space, with axes representing coherence values for the selected pairs. Group distributions were visualized as mean-centered ellipsoids with standard deviation contours (rgl package). A multivariate linear model assessed separability (β coefficients, p < 0.05), serving as a summary view of coherence-based disruption across groups. Feature selection and embedding were performed within the same dataset and were intended for exploratory characterization of coherence. The use of a low-dimensional embedding, fixed-density network construction, and group-level statistical testing was intended to limit model complexity and reduce overfitting in the context of a modest sample size. Network efficiency was evaluated using graph-theoretic metrics implemented in the igraph package. Local efficiency (LE) which measures the efficiency of information transfer within the immediate neighborhood of each node and reflects functional segregation and the resilience of local subnetworks and global efficiency (GE) which quantifies the capacity of the functional network to integrate information across distant nodes were computed from weighted, undirected phase coherence networks in the delta and theta frequency bands. Functional connectivity matrices were thresholded at 10% density (90% sparsity) to retain the strongest connections. Subject-specific graphs were constructed to preserve the physiological structure of coherence-based connectivity. In addition, clustering coefficient (C) and characteristic path length (L) were normalized against 100 degree-matched random networks per subject, and small-worldness (σ) was used to characterize the balance between network integration and segregation and defined as (C/C_rand) / (L/L_rand). Efficiency values are reported as raw measures under fixed edge density, whereas small-worldness values reflect normalized topological characteristics relative to random networks.

These analyses were designed to highlight discriminative coherence patterns, providing a data-driven foundation for subsequent classification.

E2 (Global Efficiency)where N is the number of nodes and dij denotes the shortest path length between nodes i and j. Higher global efficiency reflects more efficient long-range information transfer.

E3 (Local Efficiency)where Vi is the set of neighbors of node i, and djk is the shortest path length between nodes j and k within the subgraph induced by Vi.

E4 (Small-world Coefficient)where C and L are the clustering coefficient and the characteristic path length of the observed network, and Crand, Lrand are the same measures from a random network. A higher σ suggests an optimal network organization that balances segregation and integration.

Global efficiency, local efficiency, and small-worldness were prioritized because they provide complementary and interpretable indices of network integration, segregation, and their balance—core organizational principles of brain networks widely applied in EEG connectivity studies of neurodegeneration (Vecchio et al., 2017; Franciotti et al., 2019; Jalili, 2016; Vecchio et al., 2024). Accordingly, the present study employed fixed-density thresholding and focused on global and local efficiency metrics to enhance comparability across participants and to emphasize stable, network-level markers of cognitive dysfunction.

2.7 Supervised classification of dementia subtypes and performance evaluation

To classify clinical groups, we applied LASSO logistic regression with 10-fold cross-validation (package: glmnet), with the regularization parameter (λ) selected to minimize binomial deviance along the LASSO regularization path enabling simultaneous feature selection and model evaluation. Two standard solutions were retained for evaluation: the model corresponding to the minimum cross-validated deviance (λ_min) and a more conservative solution defined by the one-standard-error criterion (λ_1se). These complementary solutions were used to assess predictive performance as well as model parsimony and stability. In each training fold, the most predictive EEG phase coherence features (delta/theta-band PLV values) were identified by shrinking non-informative coefficients to zero via L1 regularization. For each frequency band (delta and theta), models were constructed using a fixed candidate feature space comprising p = 40 electrode-pair coherence predictors, defined a priori and applied consistently across all group comparisons.

Each observation corresponded to a single participant, ensuring that no participant appeared in more than one fold. Predictor variables were standardized within the modeling procedure using glmnet’s default setting (standardize = TRUE), and regularization parameter selection (λ_min and λ_1se) was performed entirely within the cross-validation framework.

For classification, the dependent variable (DV) was binary diagnostic group membership (HC vs. PD-MCI, HC vs. PDD, or HC vs. DLB), while independent variables (IVs) consisted of participant-level EEG phase coherence features (PLV values) computed for predefined electrode pairs in the delta (1–3.5 Hz) and theta (4–7 Hz) frequency bands.

This ensured that feature selection was performed strictly within the training data of each fold, preventing information leakage from the test set and yielding unbiased performance estimates. The resulting sparse models enhance both generalizability and interpretability of the classification (Bostanov, 2004; Rosset and Zhu, 2006). Regularization strength (λ) was optimized using both the minimum deviance and the 1-SE criteria (E5). The number of retained predictors (K) was defined as the number of non-zero coefficients in the refitted LASSO model at λ_min and λ_1se. Binary classification models were trained separately for each group comparison (HC vs. PDD, HC vs. PD-MCI, HC vs. DLB) and for delta and theta frequency bands. To improve feature selection stability, we additionally applied Bolasso (bootstrap-enhanced LASSO), retaining features selected in ≥85% of bootstrap iterations—a strategy shown to improve reproducibility in EEG-based prediction of cognitive status (Tibshirani, 1996; Fawcett, 2006).

E5where λ is a regularization parameter controlling feature selection. Top features were selected for classification.

Notably, several features consistently appeared across LASSO selections, and were also identified in earlier time–frequency and network-level analyses, supporting their convergence across analytical domains.

Classification perfo