A Bayesian Classification Approach for Exploring EEG Band Power Reduction and P300 Attenuation as Candidate Group-Associated Markers in Alcohol Use Disorder: A Preliminary Study

Abstract

Alcohol Use Disorder (AUD) is associated with widespread neurophysiological dysregulation, yet accessible and objective biomarkers for early risk identification remain limited. This preliminary study investigates electroencephalographic (EEG) frequency-domain and event-related potential features as candidate group-associated markers in AUD, with a focus on band power reduction and P300 event-related potential attenuation. Using the Begleiter EEG Database (UCI Machine Learning Repository, 1995), comprising 77 individuals with AUD and 45 healthy controls and recorded from 64 channels, we performed power spectral density estimation via Welch’s method and extracted band power across five frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz). Band power was computed at the trial level and averaged across trials at the subject level. Statistical comparisons using the Mann–Whitney U test with Benjamini–Hochberg FDR correction revealed significant band power reductions in the AUD group across frontal, central, and parietal channels, with the largest effects observed in parietal delta (P3: Hedges’ $g=-0.97$, $q<0.0001$; P4: $g=-0.95$, $q<0.0001$) and central theta (C4: $g=-0.97$, $q<0.0001$). P300 amplitude was significantly attenuated at parietal sites in S2 match (target) trials (Pz: $g=-0.75$, $p=0.0002$), consistent with hypothesised dopaminergic dysregulation of attentional processing networks. A Gaussian Naive Bayes classifier using the eight analysis channels selected for classification achieved a cross-validated ROC AUC = $0.713\pm 0.088$ [$95\%$ CI: $0.526$, $0.833$] under repeated stratified cross-validation, confirmed above chance by a permutation test ($p=0.001$). The complete analytical pipeline was executed on consumer-grade hardware without GPU acceleration, in line with Green AI principles. Topographic analysis indicated a spatially consistent parietal-dominant pattern of band power reduction. These findings suggest that task-evoked EEG frequency-domain and P300 features are systematically associated with AUD group membership and may warrant further investigation as candidate markers. The cross-sectional design of this study precludes conclusions regarding early risk identification, screening utility, or causal dopaminergic mechanisms; longitudinal validation in larger cohorts is required.

1. Introduction

Alcohol Use Disorder (AUD) represents one of the most prevalent and socially burdensome neuropsychiatric conditions worldwide, affecting an estimated 280 million individuals and contributing disproportionately to premature mortality, disability-adjusted life years, and healthcare expenditure [1,2,3].

Despite decades of clinical research, early identification of at-risk individuals remains a significant challenge, as diagnosis relies predominantly on behavioural self-report criteria and clinical interview, both of which are subject to underreporting and recall bias [4,5]. The absence of objective, neurophysiological biomarkers capable of detecting AUD vulnerability prior to disorder onset represents a critical gap in both research and clinical practice [5,6,7,8].

At the neurobiological level, AUD involves marked dysregulation of the dopaminergic reward system [9,10,11]. Chronic alcohol exposure induces progressive downregulation of D2 dopamine receptors in the mesolimbic pathway, disrupting the motivational salience circuitry that governs reward anticipation, impulse control, and stimulus evaluation [12,13,14]. This neurochemical dysregulation does not remain confined to subcortical reward structures but propagates to cortical networks, producing measurable alterations in the amplitude, synchrony, and spectral composition of neural oscillations detectable at the scalp surface via electroencephalography (EEG) [15,16,17].

EEG offers several properties that make it particularly well-suited for neurophysiological group-difference research in AUD [16,17,18,19]. It is non-invasive, low-cost, widely available in clinical settings, and sensitive to neuromodulatory tone, including dopaminergic drive, at a temporal resolution of milliseconds [18,19,20].

Frequency-domain analysis of task-evoked EEG decomposes the neural signal into canonical bands—each reflecting distinct neural processes and circuit-level dynamics [20]:

• Delta (1–4 Hz);
• Theta (4–8 Hz);
• Alpha (8–13 Hz);
• Beta (13–30 Hz);
• Gamma (30–50 Hz).

Alterations in the above-mentioned bands have been reported across multiple psychiatric and neurological conditions, and their sensitivity to dopaminergic modulation is well-established in both pharmacological and clinical EEG research [18,19,21].

Among various EEG-derived measures, the P300 event-related potential (ERP) component has accumulated the most robust evidence as a neurophysiological correlate of AUD vulnerability [22,23]. The P300 is a positive deflection occurring approximately 300 ms after a task-relevant target stimulus, is maximal over parietal scalp regions, and reflects the allocation of attentional and cognitive resources to stimulus evaluation [23,24]. Its amplitude is modulated by dopaminergic signalling, such that higher dopaminergic tone facilitates larger P300 responses, while dopaminergic dysregulation, as in AUD, attenuates the component [22,23]. The seminal finding by Begleiter and colleagues in 1984 demonstrated that P300 amplitude was significantly reduced not only in individuals with established AUD but also in their biological sons who had never consumed alcohol, establishing P300 reduction as a heritable endophenotype of AUD vulnerability rather than a consequence of drinking per se [25]. This finding fundamentally repositioned P300 as a prospective risk marker with potential screening applications.

Subsequent research has confirmed and extended these observations. Reduced P300 amplitude in AUD has been replicated across numerous independent cohorts, associated with structural and functional alterations in the prefrontal and parietal cortex, and linked to deficits in working memory, response inhibition, and reward processing [26,27].

Parallel work on task-evoked EEG band power has demonstrated that AUD is associated with broad reductions in slow-wave oscillatory power, particularly in delta and theta bands, reflecting impaired thalamocortical gating and reduced inhibitory tone in frontoparietal networks [28,29].

Despite this rich empirical foundation, relatively few studies have applied probabilistic classification frameworks to EEG frequency features for AUD risk stratification [29]. Most existing classification work has employed support vector machines, linear discriminant analysis, or convolutional neural networks, which, while achieving reasonable accuracy, offer limited interpretability and do not naturally express uncertainty in terms of posterior probabilities—a property of particular importance in clinical risk assessment contexts. Bayesian classifiers, and specifically Gaussian Naive Bayes, provide a probabilistic output that directly estimates the posterior probability of class membership given observed features, offer transparent likelihood modelling per feature, and remain tractable in small- to medium-sample settings where deep learning approaches may overfit.

Our study constitutes a preliminary investigation of EEG frequency-domain and P300 features as candidate group-associated markers in AUD using the publicly available Begleiter EEG Database [25,30] from the UCI Machine Learning Repository. We perform power spectral density estimation and band power extraction across eight clinically relevant channels spanning frontal, central, and parietal scalp regions; characterise group differences between AUD and healthy control subjects using both effect size and inferential statistics; analyse P300 amplitude attenuation across the same channel set; construct topographic maps of the full scalp distribution of band power differences; and apply a Gaussian Naive Bayes classifier with cross-validated ROC AUC evaluation to assess the discriminative value of spectral features.

Given the sample size constraints inherent to the dataset and the exploratory nature of the classification approach, this work is explicitly positioned as a preliminary study providing a methodological and empirical foundation for larger-scale longitudinal validation efforts. While prior studies have reported EEG band power reductions and P300 attenuation in AUD using the Begleiter dataset, the present work contributes a methodologically refined reanalysis incorporating trial-type separation, rigorous effect size reporting with bootstrap confidence intervals, FDR correction, probabilistic classification with calibration, and a computationally accessible pipeline, rather than claiming empirical novelty in the observed group differences themselves. Our study makes the following contributions:

• It provides a systematic, multichannel quantification of task-evoked EEG band power reduction in AUD with Hedges’ g effect sizes and bootstrap confidence intervals reported across a standardised eight-channel montage.
• It characterises P300 amplitude attenuation across the same channel set within a unified analytical framework.
• It applies a Gaussian Naive Bayes probabilistic classifier with log-transformed band power features, calibration analysis, permutation testing, and logistic regression comparison under repeated stratified cross-validation ($10\times 5$ folds, $n=122$).
• It demonstrates that the complete pipeline is executable on consumer-grade hardware, providing a computationally accessible baseline that avoids GPU acceleration and cloud computing resources, completing in $4.3$ min. No formal energy consumption or carbon footprint estimate was performed; quantitative Green AI benchmarking is identified as a direction for future work.
• It provides a methodologically refined reanalysis of the Begleiter dataset combining: explicit trial-type separation, Hedges’ g with bootstrap confidence intervals under FDR correction, log-transformed probabilistic classification with calibration analysis and permutation testing, and a computationally accessible pipeline—none of which have been applied in combination to this dataset in prior published work.

More recently, deep learning approaches, including convolutional neural networks, long short-term memory recurrent networks, graph neural networks, and transfer learning, have been applied to raw or minimally preprocessed EEG signals, reporting ROC AUC values in the range of $0.85$–$0.98$ across various datasets [31]. These gains come at the cost of substantially reduced interpretability and greater sensitivity to dataset-specific overfitting. Such methods also require GPU acceleration and substantial hyperparameter optimisation, and they require large training datasets to generalise reliably. Both constraints limit their applicability to the relatively small samples characteristic of most clinical EEG studies in AUD and to resource-constrained settings more generally. The present study positions itself as a computationally accessible, interpretable baseline rather than a competitor to these state-of-the-art approaches.

2. Materials and Methods

This section describes the dataset, preprocessing pipeline, feature extraction procedures, statistical analysis framework, and classification methodology employed in this study. All analyses were designed to run on standard consumer-grade hardware without specialised computational resources.

2.1. Applied Dataset

Our study uses the EEG Database contributed by Begleiter and colleagues to the UCI Machine Learning Repository [30], a publicly available neurophysiological dataset, which has served as a benchmark resource for EEG-based AUD research since its initial release. The dataset comprises recordings from 122 subjects divided into two groups:

• A total of 77 individuals diagnosed with Alcohol Use Disorder (AUD group): Subjects in the AUD group met diagnostic criteria for alcohol dependence as defined by the Diagnostic and Statistical Manual of Mental Disorders at the time of data collection;
• A total of 45 healthy controls (control group): Control subjects had no personal or first-degree family history of a psychiatric or neurological disorder.

EEG was recorded from 64 scalp electrodes placed according to the extended international 10–20 system, with a sampling frequency of 256 Hz. Recordings were obtained during a visual delayed matching-to-sample paradigm. In each trial a first stimulus (S1) was presented, followed after an interstimulus interval by a second stimulus (S2) that either matched S1 or did not. Subjects judged whether the two stimuli matched. Trials are accordingly labelled S1, S2 match, and S2 no-match. The P300 analysed in this study was elicited by S2 match stimuli, which required comparison of the current stimulus against the S1 representation maintained in the working memory. Each subject completed multiple trials, with each trial yielding a 1000 ms epoch of 256 samples per channel. The dataset is organised such that each subject’s data are stored as a compressed archive containing individual trial files, each representing a single stimulus presentation across all 64 channels.

2.2. Data Loading and Preprocessing

Raw data were extracted from the nested compressed archive structure using a custom Python pipeline. Each subject-level archive contained multiple trial files in a proprietary format in which each data record comprised four white-space-separated fields: trial index, channel label, sample index, and amplitude value in microvolts. Given that the raw file content was stored as a single continuous token stream without explicit line breaks between records, a token-based parsing algorithm was implemented to reconstruct the multichannel time series from the sequential token stream. Records were identified by matching the pattern of integers, alphabetic channel labels, integers, and floating-point values across consecutive tokens, with non-conforming tokens corresponding to header metadata discarded.

For each subject, valid trials were extracted and the multichannel time series from each trial were aligned to a common length by truncating to the minimum trial duration across all trials for that subject. The mean number of usable S2 match trials per subject was $23.4$ overall (AUD group: $23.4$; control group: $22.6$), confirming balanced trial counts across groups. No further artefact rejection was applied. For band power analysis, PSD was computed separately for each trial and averaged across all trials at the subject level. For P300 analysis, trial-averaged ERPs were computed separately for S2 match (target) and S1 trials. Trial averaging constitutes the standard ERP analysis approach and serves to attenuate trial-to-trial noise while preserving stimulus-locked neural responses.

Channel labels were normalised to conform to the MNE-Python standard 10–20 montage naming convention, in which midline electrodes were designated with lowercase z suffixes (e.g., Fz, Cz, Pz) rather than the uppercase Z convention used in the source dataset. A mapping dictionary was applied to perform this normalisation prior to topographic analysis.

2.3. Channel Selection

Eight channels were selected for primary analysis based on their established relevance to AUD neurophysiology and their coverage of the major scalp regions of interest:

• F3 and F4 (left and right frontal);
• C3 and C4 (left and right central);
• Cz (central midline);
• P3 and P4 (left and right parietal);
• Pz (parietal midline).

The above selection covers the frontal executive network, the central sensorimotor region, and the parietal attention network: the three cortical systems most consistently implicated in AUD-related EEG alterations in the literature [32]. All 59 channels with valid standard 10–20 montage positions were retained for topographic mapping analyses.

2.4. Power Spectral Density Estimation

Power spectral density (PSD) was estimated for each channel of each subject using Welch’s method, as implemented in the SciPy signal processing library [33]. The segment length parameter nperseg was set equal to the full epoch length of 256 samples, yielding a frequency resolution of 1 Hz, which was sufficient to resolve all five canonical EEG frequency bands. A Hann window was applied to each segment. The PSD was estimated at frequencies from 0 to 128 Hz, and subsequent analyses were restricted to the $1.0$ to 50 Hz range to exclude DC offset components and high-frequency noise. It is acknowledged that band power computed from task-evoked epochs reflects both stimulus-locked and non-stimulus-locked neural activity; these features should therefore be interpreted as task-evoked spectral measures rather than resting-state power estimates.

Band power was computed for each of the five canonical EEG frequency bands by integrating the PSD over the corresponding frequency range using the trapezoidal rule, as given in Equation (1):

$$BP(f_1,f_2)={\int }_{f_1}^{f_2}S\left(f\right)df\approx {\displaystyle \sum _{k=1}^{K-1}}{\displaystyle \frac{1}{2}}\left(S\left(f_k\right)+S\left(f_{k+1}\right)\right)\left(f_{k+1}-f_k\right),$$

(1)

where $S\left(f\right)$ denotes the power spectral density at frequency f, $f_1$ and $f_2$ are the lower and upper band boundaries, K is the number of frequency bins within the band $[f_1,f_2]$, and band power $BP$ is expressed in ${\mathsf{\mu }\mathrm{V}}^2$.

The five frequency bands were defined as: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz).

It is acknowledged that setting nperseg equal to the full epoch length of 256 samples results in a single-segment Welch estimate, mathematically equivalent to a windowed periodogram. This choice is deliberate: shorter segments would reduce frequency resolution below 1 Hz, further compromising representation of the delta band. Variance reduction is achieved through averaging across trials rather than within-epoch segmentation, consistent with standard practice for short task-evoked epochs. The delta band lower boundary was set to $1.0$ Hz to match the frequency resolution of the analysis. As an additional validation, band power was recomputed using multitaper spectral estimation (DPSS tapers, $NW=4$, $K_t=7$) on the same trial-level data. The direction and relative magnitude of group differences were fully consistent between Welch and multitaper estimates across all 40 channel–band combinations. For example, at channel Fz, both methods identified delta and theta as the bands with the largest group differences, with control group power exceeding AUD group power in both cases. Absolute values differed due to normalisation conventions, but the group contrast was preserved, confirming that the primary findings were not an artefact of the spectral estimation method.

2.5. P300 Amplitude Extraction

P300 amplitude was extracted from the trial-averaged ERP waveform for each channel and subject. Following standard practice in the ERP literature [23,24], the P300 window was defined as 250 to 400 ms post-stimulus onset. Within this window, peak amplitude was identified as the maximum value of the averaged ERP waveform. Time was computed from the sample index as $t=n/f_s\times 1000$ ms, where n is the sample index and $f_s=256$ Hz is the sampling frequency.

Since the paradigm is a delayed matching task rather than an oddball, the S1 versus S2 match contrast reflects stimulus-comparison and working memory demand rather than stimulus probability. The reported attenuation is therefore interpreted as reduced resource allocation during stimulus evaluation, without invoking a probability-based account.

2.6. Conducted Statistical Analysis

Group differences in band power and P300 amplitude between AUD and control subjects were assessed as follows: The assumption of normality was assessed formally for each channel–band combination using the Shapiro–Wilk test. The normality assumption was violated for all 40 channel–band combinations, and the Mann–Whitney U test was therefore applied throughout, as implemented in the SciPy statistics module. To control for Type I error inflation across multiple comparisons, Benjamini–Hochberg false discovery rate (FDR) correction was applied across all 40 channel–band tests, with corrected q-values reported alongside nominal p-values in Table 1. Effect sizes were quantified using Hedges’ g, preferred over Cohen’s d because of its correction for small-sample bias, computed as in Equation (2):

$$g=J·{\displaystyle \frac{{\mu }_{AUD}-{\mu }_{Control}}{s_{pooled}}},s_{pooled}=\sqrt{{\displaystyle \frac{(n_{AUD}-1)s_{AUD}^2+(n_{Control}-1)s_{Control}^2}{n_{AUD}+n_{Control}-2}}},$$

(2)

where $n_{AUD}=77$ and $n_{Control}=45$ are the group sample sizes, $s^2$ denotes the sample variance within each group, and J is the small-sample correction factor:

$$J=1-{\displaystyle \frac{3}{4(n_{AUD}+n_{Control}-2)-1}}.$$

Statistical significance was evaluated on FDR-corrected q-values at three thresholds:

• q < 0.05 (*);
• q < 0.01 (**);
• q < 0.001 (***).

2.7. Gaussian Naive Bayes Classification

All stochastic procedures used a fixed random seed of 42, including repeated stratified cross-validation, permutation testing (1000 permutations), and bootstrap confidence interval estimation (1000 resamples).

Prior to classification, band power features were log-transformed using the natural logarithm with a unit offset ($\log (1+x)$) to reduce the positive skew characteristic of power spectral density distributions and to better approximate the Gaussian class-conditional likelihood assumption of the GNB classifier. Log-transformation reduced skewness substantially across all channel–band combinations (e.g., PZ delta: skewness reduced from $5.10$ to $1.06$). Features were subsequently standardised using z-score normalisation within each training fold to prevent data leakage.

The Gaussian Naive Bayes (GNB) classifier models the class-conditional likelihood of each feature as a univariate Gaussian distribution, as given in Equation (3):

$$P(x_i\mid y=c)={\displaystyle \frac{1}{\sqrt{2\pi {\sigma }_{ic}^2}}}\exp \left(-{\displaystyle \frac{{(x_i-{\mu }_{ic})}^2}{2{\sigma }_{ic}^2}}\right)$$

(3)

where ${\mu }_{ic}$ and ${\sigma }_{ic}^2$ are the mean and variance of feature i within class c, estimated from the training data via maximum likelihood. The posterior probability of class membership is then computed via Bayes’ theorem, as given in Equation (4):

$$P(y=c\mid \mathbf{x})={\displaystyle \frac{P(y=c){\prod }_{i}P(x_i\mid y=c)}{{\sum }_{c^{\prime }}P(y=c^{\prime }){\prod }_{i}P(x_i\mid y=c^{\prime })}}$$

(4)

with a uniform prior $P(y=c)=0.5$ assigned to both classes to ensure that classification performance reflects discriminative capacity independently of the empirical class imbalance in the dataset.

Feature values were standardised prior to classification using z-score normalisation, with mean and standard deviation estimated from the training fold to prevent data leakage.

The primary classification analysis used the full 40-feature vector comprising band power from all eight analysis channels (F3, F4, C3, C4, Cz, P3, P4, Pz), evaluated on the full dataset ($n=122$) without subsampling. Repeated stratified cross-validation ($10\times 5$ folds) was applied, with the area under the receiver operating characteristic curve (ROC AUC) as the primary performance metric. ROC AUC was selected in preference to accuracy given the class imbalance between AUD ($n=77$) and control ($n=45$) subjects. The analysis yielded ROC AUC = $0.713\pm 0.088$ [$95\%$ CI: $0.526$, $0.833$] for GNB, with a fold score range of $0.526$ to $0.833$ across the 50 folds. A logistic regression classifier was included as a comparison baseline. A permutation test (1000 permutations) was conducted to assess whether the observed AUC exceeded chance level.

Two additional analyses were performed as reference points. First, a balanced evaluation set was constructed by sampling 40 subjects from each group, yielding a classification dataset of 80 subjects, evaluated under stratified five-fold cross-validation. This analysis is reported to permit comparison with prior work on this dataset that used balanced subsamples. Second, band power features were extracted from a single channel, Fz, which lay outside the eight-channel analysis set and was used to obtain a minimal feature reference value. The single-channel Fz result (AUC = $0.591$) is retained as a lower-bound baseline only; no substantive conclusions are drawn from this minimal feature set. A systematic channel selection analysis using maximum relevance minimum redundancy (mRMR) or similar criteria is identified as a priority for future work.

Calibration was assessed using isotonic regression post hoc calibration under 5-fold cross-validation. The uncalibrated GNB yielded a Brier score of $0.272$, i.e., worse than the no-skill value of $0.233$ obtained from the empirical class base rate, while the calibrated model yielded a Brier score of $0.199$, which improved on that value. Since the calibration model was fitted and evaluated within the same cross-validation procedure, the calibrated score was likely to be optimistic; nested cross-validation would be required for an unbiased estimate. The uniform prior $P\left(\mathrm{AUD}\right)=0.5$ was used to evaluate discriminative capacity independently of prevalence assumptions and should not be interpreted as a real-world AUD risk probability. Clinical application would require prior adjustment to match target population prevalence and formal prospective calibration validation.

2.8. Topographic Mapping

Topographic scalp maps of band power were generated using the MNE-Python library [34]. Channel positions were assigned according to the standard 10–20 montage as implemented in MNE-Python’s built-in montage database. For each frequency band, mean band power was computed across all subjects within each group for each of the 59 channels with valid montage positions. Scalp distributions were interpolated using the spherical spline method implemented in MNE-Python’s plot_topomap function, with six contour lines overlaid to indicate isolines of equal power. Separate maps were generated for the AUD group, the control group, and the difference (AUD minus control), with the difference map displayed using a diverging colourmap centred at zero to facilitate visualisation of the direction and spatial distribution of group differences.

All analyses were implemented in Python 3.12 using the following libraries: NumPy 1.26 [35], SciPy 1.13 [33], pandas 2.2 [36], MNE-Python 1.8 [34], scikit-learn 1.5 [37,38], and Matplotlib 3.10 [39].

3. Results

This section presents the results of the spectral, topographic, event-related potential, and Bayesian classification analyses. Results are reported in order of increasing spatial and temporal specificity, beginning with group-level power spectral density comparisons and proceeding through band power statistics, topographic distributions, P300 amplitude characterisation, and classification performance.

3.1. Power Spectral Density

Power spectral density curves computed via Welch’s method for the AUD and control groups at channel Fz are presented in Figure 1. Across the full frequency range from $1.0$ to 50 Hz, the control group exhibited consistently higher spectral power than the AUD group, with the between-group separation most pronounced in the low-frequency range corresponding to the delta and theta bands.

The spectral profiles of both groups followed the expected $1/f$ pattern characteristic of task-evoked EEG, with power declining monotonically as a function of frequency. The between-group power ratio was largest in the delta band, where control group power exceeded AUD group power by a factor of approximately $2.0$ at channel Fz and decreased progressively across theta, alpha, and beta, reaching approximate parity in the gamma range.

3.2. Band Power Comparison

Mean band power values, standard deviations, Hedges’ g effect sizes with bootstrap confidence intervals, and FDR-corrected q-values for all eight analysis channels and five frequency bands are presented in Table 1.

Delta band power was significantly reduced in the AUD group at all eight analysis channels. The largest delta effects were observed at parietal channels P3 (Hedges’ $g=-0.97$, $q<0.0001$) and P4 ($g=-0.95$, $q<0.0001$). C4 ($g=-0.80$), F3 ($g=-0.71$), C3 ($g=-0.67$), Pz ($g=-0.65$), Cz ($g=-0.57$), and F4 ($g=-0.52$) showed moderate to large effects, all significant at $q\le 0.0006$.

This pattern is visualised in Figure 2, which presents grouped band power comparisons with significance annotations for channel Fz.

Theta band power was significantly reduced at all eight channels ($q<0.0001$). Effect sizes were largest at central channels C4 ($g=-0.97$), Cz ($g=-0.94$), and C3 ($g=-0.92$), followed by parietal and frontal channels (g range: $-0.93$ to $-0.79$).

Alpha band power was significantly reduced at all eight channels, with the largest effects at P4 ($g=-0.85$, $q<0.0001$), P3 ($g=-0.79$, $q=0.0001$), and C4 ($g=-0.78$, $q=0.0003$). Frontal channels F3 ($g=-0.60$, $q=0.0006$) and F4 ($g=-0.61$, $q=0.0018$) showed the smallest alpha effects.

Beta band power reached significance only at the parietal channels: P4 ($g=-0.44$, $q=0.0090$), P3 ($g=-0.37$, $q=0.0310$), and Pz ($g=-0.30$, $q=0.0399$). Frontal and central channels did not reach significance (q range: $0.1053$–$0.2397$).

Gamma band power showed no significant group differences at any channel (q range: $0.4657$–$0.9155$; g range: $-0.13$ to $+0.28$).

The complete channel-by-band power matrix for both groups is summarised in the heatmap presented in Figure 3, which illustrates the spatial and spectral distribution of group differences across all eight analysis channels simultaneously.

Table 1. Mean band power (${\mathsf{\mu }\mathrm{V}}^2$), standard deviation (SD), Hedges’ g with $95\%$ bootstrap confidence interval (1000 resamples, percentile method), and FDR-corrected q-value (Benjamini–Hochberg) for AUD vs. control groups across eight analysis channels and five frequency bands. Band power is computed as the trial-level PSD (Welch periodogram) averaged across trials at the subject level. Mann–Whitney U test is used for all comparisons. Significance after FDR correction: * $q<0.05$, ** $q<0.01$, *** $q<0.001$, and ns = not significant.

Ch.	Band	AUD	AUD SD	Ctrl	Ctrl SD	g [95% CI]	q
		(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)	(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)	(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)	(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)
F3	delta	4.691	3.225	7.106	3.602	$-0.71$ [$-1.27$, $-0.32$]	<0.0001 ***
F3	theta	4.337	2.713	7.664	4.787	$-0.91$ [$-1.33$, $-0.57$]	<0.0001 ***
F3	alpha	3.455	3.848	6.001	4.745	$-0.60$ [$-1.10$, $-0.21$]	$0.0006$ ***
F3	beta	4.019	3.418	4.499	2.998	$-0.15$ [$-0.55$, $0.19$]	$0.1488$ ns
F3	gamma	1.247	1.285	1.169	1.038	$0.06$ [$-0.30$, $0.37$]	$0.9155$ ns
F4	delta	5.120	4.156	7.200	3.534	$-0.52$ [$-1.14$, $-0.15$]	<0.0001 ***
F4	theta	4.617	2.830	7.672	4.767	$-0.83$ [$-1.21$, $-0.50$]	<0.0001 ***
F4	alpha	3.459	3.482	5.875	4.637	$-0.61$ [$-1.09$, $-0.23$]	$0.0018$ **
F4	beta	4.328	3.535	4.549	2.730	$-0.07$ [$-0.45$, $0.26$]	$0.2393$ ns
F4	gamma	1.447	1.454	1.118	1.130	$0.24$ [$-0.12$, $0.56$]	$0.4657$ ns
C3	delta	1.513	0.943	2.447	1.895	$-0.67$ [$-1.00$, $-0.36$]	$0.0006$ ***
C3	theta	1.795	1.149	3.477	2.586	$-0.92$ [$-1.35$, $-0.63$]	<0.0001 ***
C3	alpha	2.157	2.178	3.922	3.542	$-0.63$ [$-1.07$, $-0.25$]	$0.0016$ **
C3	beta	2.238	1.835	2.573	2.055	$-0.17$ [$-0.52$, $0.19$]	$0.2397$ ns
C3	gamma	0.777	0.900	0.566	0.399	$0.28$ [$-0.01$, $0.53$]	$0.8543$ ns
C4	delta	1.378	0.858	2.173	1.184	$-0.80$ [$-1.24$, $-0.43$]	<0.0001 ***
C4	theta	1.687	1.141	3.433	2.532	$-0.97$ [$-1.35$, $-0.66$]	<0.0001 ***
C4	alpha	1.938	1.952	4.073	3.696	$-0.78$ [$-1.20$, $-0.42$]	$0.0003$ ***
C4	beta	2.367	2.761	2.554	1.881	$-0.07$ [$-0.52$, $0.23$]	$0.1053$ ns
C4	gamma	0.981	1.870	0.594	0.750	$0.25$ [$-0.06$, $0.45$]	$0.8574$ ns
P3	delta	3.813	1.927	6.728	4.220	$-0.97$ [$-1.39$, $-0.67$]	<0.0001 ***
P3	theta	4.853	3.090	9.288	6.734	$-0.92$ [$-1.33$, $-0.64$]	<0.0001 ***
P3	alpha	4.337	3.840	8.869	7.866	$-0.79$ [$-1.16$, $-0.46$]	$0.0001$ ***
P3	beta	3.722	2.769	4.992	4.314	$-0.37$ [$-0.72$, $0.01$]	$0.0310$ *
P3	gamma	0.867	0.574	0.983	1.220	$-0.13$ [$-0.47$, $0.38$]	$0.8637$ ns
P4	delta	3.731	1.841	7.326	5.670	$-0.95$ [$-1.52$, $-0.72$]	<0.0001 ***
P4	theta	4.684	3.015	9.382	7.237	$-0.93$ [$-1.34$, $-0.65$]	<0.0001 ***
P4	alpha	3.970	3.715	9.051	8.486	$-0.85$ [$-1.22$, $-0.52$]	<0.0001 ***
P4	beta	3.346	2.453	4.801	4.284	$-0.44$ [$-0.78$, $-0.07$]	$0.0090$ **
P4	gamma	0.815	0.528	0.840	0.680	$-0.04$ [$-0.43$, $0.38$]	$0.9155$ ns
Pz	delta	3.068	2.461	4.623	2.170	$-0.65$ [$-1.40$, $-0.25$]	<0.0001 ***
Pz	theta	3.403	2.438	5.750	3.693	$-0.79$ [$-1.25$, $-0.42$]	<0.0001 ***
Pz	alpha	2.899	2.970	6.147	6.404	$-0.71$ [$-1.10$, $-0.40$]	<0.0001 ***
Pz	beta	2.519	1.992	3.222	2.814	$-0.30$ [$-0.65$, $0.07$]	$0.0399$ *
Pz	gamma	0.547	0.414	0.520	0.313	$0.07$ [$-0.31$, $0.38$]	$0.8605$ ns
Cz	delta	24.343	11.009	36.748	32.306	$-0.57$ [$-0.85$, $-0.33$]	$0.0006$ ***
Cz	theta	17.872	7.289	27.721	14.174	$-0.94$ [$-1.34$, $-0.59$]	<0.0001 ***
Cz	alpha	9.383	7.380	15.572	11.617	$-0.67$ [$-1.11$, $-0.31$]	$0.0012$ **
Cz	beta	10.435	5.688	10.659	3.948	$-0.04$ [$-0.45$, $0.27$]	$0.2397$ ns
Cz	gamma	4.803	3.755	4.473	2.689	$0.10$ [$-0.27$, $0.40$]	$0.9155$ ns

Table 1. Mean band power (${\mathsf{\mu }\mathrm{V}}^2$), standard deviation (SD), Hedges’ g with $95\%$ bootstrap confidence interval (1000 resamples, percentile method), and FDR-corrected q-value (Benjamini–Hochberg) for AUD vs. control groups across eight analysis channels and five frequency bands. Band power is computed as the trial-level PSD (Welch periodogram) averaged across trials at the subject level. Mann–Whitney U test is used for all comparisons. Significance after FDR correction: * $q<0.05$, ** $q<0.01$, *** $q<0.001$, and ns = not significant.

Ch.	Band	AUD	AUD SD	Ctrl	Ctrl SD	g [95% CI]	q
		(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)	(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)	(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)	(${\mathsf{\mu }\mathbf{V}}^{\mathbf{2}}$)
F3	delta	4.691	3.225	7.106	3.602	$-0.71$ [$-1.27$, $-0.32$]	<0.0001 ***
F3	theta	4.337	2.713	7.664	4.787	$-0.91$ [$-1.33$, $-0.57$]	<0.0001 ***
F3	alpha	3.455	3.848	6.001	4.745	$-0.60$ [$-1.10$, $-0.21$]	$0.0006$ ***
F3	beta	4.019	3.418	4.499	2.998	$-0.15$ [$-0.55$, $0.19$]	$0.1488$ ns
F3	gamma	1.247	1.285	1.169	1.038	$0.06$ [$-0.30$, $0.37$]	$0.9155$ ns
F4	delta	5.120	4.156	7.200	3.534	$-0.52$ [$-1.14$, $-0.15$]	<0.0001 ***
F4	theta	4.617	2.830	7.672	4.767	$-0.83$ [$-1.21$, $-0.50$]	<0.0001 ***
F4	alpha	3.459	3.482	5.875	4.637	$-0.61$ [$-1.09$, $-0.23$]	$0.0018$ **
F4	beta	4.328	3.535	4.549	2.730	$-0.07$ [$-0.45$, $0.26$]	$0.2393$ ns
F4	gamma	1.447	1.454	1.118	1.130	$0.24$ [$-0.12$, $0.56$]	$0.4657$ ns
C3	delta	1.513	0.943	2.447	1.895	$-0.67$ [$-1.00$, $-0.36$]	$0.0006$ ***
C3	theta	1.795	1.149	3.477	2.586	$-0.92$ [$-1.35$, $-0.63$]	<0.0001 ***
C3	alpha	2.157	2.178	3.922	3.542	$-0.63$ [$-1.07$, $-0.25$]	$0.0016$ **
C3	beta	2.238	1.835	2.573	2.055	$-0.17$ [$-0.52$, $0.19$]	$0.2397$ ns
C3	gamma	0.777	0.900	0.566	0.399	$0.28$ [$-0.01$, $0.53$]	$0.8543$ ns
C4	delta	1.378	0.858	2.173	1.184	$-0.80$ [$-1.24$, $-0.43$]	<0.0001 ***
C4	theta	1.687	1.141	3.433	2.532	$-0.97$ [$-1.35$, $-0.66$]	<0.0001 ***
C4	alpha	1.938	1.952	4.073	3.696	$-0.78$ [$-1.20$, $-0.42$]	$0.0003$ ***
C4	beta	2.367	2.761	2.554	1.881	$-0.07$ [$-0.52$, $0.23$]	$0.1053$ ns
C4	gamma	0.981	1.870	0.594	0.750	$0.25$ [$-0.06$, $0.45$]	$0.8574$ ns
P3	delta	3.813	1.927	6.728	4.220	$-0.97$ [$-1.39$, $-0.67$]	<0.0001 ***
P3	theta	4.853	3.090	9.288	6.734	$-0.92$ [$-1.33$, $-0.64$]	<0.0001 ***
P3	alpha	4.337	3.840	8.869	7.866	$-0.79$ [$-1.16$, $-0.46$]	$0.0001$ ***
P3	beta	3.722	2.769	4.992	4.314	$-0.37$ [$-0.72$, $0.01$]	$0.0310$ *
P3	gamma	0.867	0.574	0.983	1.220	$-0.13$ [$-0.47$, $0.38$]	$0.8637$ ns
P4	delta	3.731	1.841	7.326	5.670	$-0.95$ [$-1.52$, $-0.72$]	<0.0001 ***
P4	theta	4.684	3.015	9.382	7.237	$-0.93$ [$-1.34$, $-0.65$]	<0.0001 ***
P4	alpha	3.970	3.715	9.051	8.486	$-0.85$ [$-1.22$, $-0.52$]	<0.0001 ***
P4	beta	3.346	2.453	4.801	4.284	$-0.44$ [$-0.78$, $-0.07$]	$0.0090$ **
P4	gamma	0.815	0.528	0.840	0.680	$-0.04$ [$-0.43$, $0.38$]	$0.9155$ ns
Pz	delta	3.068	2.461	4.623	2.170	$-0.65$ [$-1.40$, $-0.25$]	<0.0001 ***
Pz	theta	3.403	2.438	5.750	3.693	$-0.79$ [$-1.25$, $-0.42$]	<0.0001 ***
Pz	alpha	2.899	2.970	6.147	6.404	$-0.71$ [$-1.10$, $-0.40$]	<0.0001 ***
Pz	beta	2.519	1.992	3.222	2.814	$-0.30$ [$-0.65$, $0.07$]	$0.0399$ *
Pz	gamma	0.547	0.414	0.520	0.313	$0.07$ [$-0.31$, $0.38$]	$0.8605$ ns
Cz	delta	24.343	11.009	36.748	32.306	$-0.57$ [$-0.85$, $-0.33$]	$0.0006$ ***
Cz	theta	17.872	7.289	27.721	14.174	$-0.94$ [$-1.34$, $-0.59$]	<0.0001 ***
Cz	alpha	9.383	7.380	15.572	11.617	$-0.67$ [$-1.11$, $-0.31$]	$0.0012$ **
Cz	beta	10.435	5.688	10.659	3.948	$-0.04$ [$-0.45$, $0.27$]	$0.2397$ ns
Cz	gamma	4.803	3.755	4.473	2.689	$0.10$ [$-0.27$, $0.40$]	$0.9155$ ns

Figure 2. Comparison of mean band power (${\mathsf{\mu }\mathrm{V}}^2$, log scale) between the AUD ($n=77$, red) and control ($n=45$, blue) groups at channel Fz. Error bars represent standard error of the mean. Significance markers are from the Mann–Whitney U test (uncorrected p-values): * $p<0.05$, ** $p<0.01$, *** $p<0.001$, and ns = not significant. FDR correction was applied to the primary eight-channel analysis reported in Table 1; channel Fz is shown here for illustrative purposes only and its values are not included in Table 1. Band power was computed as the trial-level Welch periodogram averaged across trials at the subject level.

Figure 2. Comparison of mean band power (${\mathsf{\mu }\mathrm{V}}^2$, log scale) between the AUD ($n=77$, red) and control ($n=45$, blue) groups at channel Fz. Error bars represent standard error of the mean. Significance markers are from the Mann–Whitney U test (uncorrected p-values): * $p<0.05$, ** $p<0.01$, *** $p<0.001$, and ns = not significant. FDR correction was applied to the primary eight-channel analysis reported in Table 1; channel Fz is shown here for illustrative purposes only and its values are not included in Table 1. Band power was computed as the trial-level Welch periodogram averaged across trials at the subject level.

Figure 3. Heatmap of mean band power (${\mathsf{\mu }\mathrm{V}}^2$) across all eight analysis channels and five frequency bands, determined separately for AUD and control groups. Both panels use an identical colour scale to enable direct visual comparison between groups.

Figure 3. Heatmap of mean band power (${\mathsf{\mu }\mathrm{V}}^2$) across all eight analysis channels and five frequency bands, determined separately for AUD and control groups. Both panels use an identical colour scale to enable direct visual comparison between groups.

3.3. Topographic Distribution of Band Power

Topographic scalp maps of mean band power for the AUD group, control group, and group difference are presented in Figure 4 for all five frequency bands.

In the delta band, the difference map revealed a broad pattern of AUD-related power reduction, with the largest deficits concentrated in the parietal and occipital regions and the smallest difference in the central midline, consistent with the comparatively small effect at Cz ($g=-0.57$). Theta topography showed the highest power over the frontal midline and central regions, with the difference map showing a widespread frontocentral reduction in AUD. Alpha topography showed the expected posterior dominance, with the largest reductions related to AUD in the parietal and central regions. Beta and gamma difference maps showed no systematic spatial organisation, consistent with their weaker or absent inferential effects.

3.4. P300 Analysis

Trial-averaged event-related potential waveforms for the S2 match (target) trials and the S2 match minus S1 contrast for the AUD and control groups across all eight analysis channels are presented in Figure 5 and Figure 6. P300 amplitude was extracted from the 250–400 ms window of S2 match trials only.

In the control group, a clear positive deflection was observed at approximately 300 ms post-stimulus at parietal channels P3, P4, and Pz in S2 match (target) trials, consistent with the canonical P300 component (Figure 5 and Figure 6). This deflection was substantially attenuated in the AUD group at all parietal channels (Pz: $g=-0.75$, $p=0.0002$; P3: $g=-0.59$, $p=0.005$; P4: $g=-0.63$, $p=0.008$). At frontal channels F3 and F4, the negativity present in the control group was reduced in the AUD group (F3: $g=0.75$, $p<0.0001$; F4: $g=0.48$, $p=0.006$). Central channels C3, C4, and Cz did not show significant group differences in the S2 match condition ($p>0.05$).

The S2 match minus S1 contrast did not reach significance at parietal channels (Pz: $g=0.12$, $p=0.90$), indicating that both S1 and S2 match stimuli elicited attenuated responses in the AUD group. This pattern is consistent with a general reduction in stimulus-evoked neural activity rather than a deficit specific to stimulus matching per se.

P300 peak amplitudes extracted from the 250–400 ms window are presented in

Table 2. P300 measures from S2 match (target) trials across eight analysis channels. Peak amplitude: Maximum value in the 250–400 ms window. Mean-window amplitude: Mean value across the 250–400 ms window. Peak latency: Time of peak amplitude within the 250–400 ms window. Amplitude values reported as mean ± SD. Hedges’ g and significance (Mann–Whitney U test, uncorrected) are reported for all three measures. Note: ERPs were computed without baseline correction; amplitude differences may reflect pre-stimulus offset or slow drift in addition to evoked response differences. * $p<0.05$, ** $p<0.01$, *** $p<0.001$, ns = not significant.

Ch.	AUD Peak	Ctrl Peak	g	Sig	AUD Mean	Ctrl Mean	g	Sig	AUD Lat.	Ctrl Lat.	Sig
	(μV)	(μV)			(μV)	(μV)			(ms)	(ms)
F3	$0.52$	$-1.39$	$0.79$	***	$-1.60$	$-4.20$	$0.98$	***	$283.5$	$281.1$	ns
F4	$0.21$	$-1.02$	$0.48$	**	$-1.82$	$-3.79$	$0.71$	***	$297.8$	$292.7$	ns
C3	$0.35$	$0.08$	$0.17$	ns	$-1.03$	$-1.73$	$0.43$	*	$305.7$	$316.1$	ns
C4	$0.16$	$0.28$	$-0.07$	ns	$-1.19$	$-1.21$	$0.01$	ns	$325.9$	$322.7$	ns
P3	$1.04$	$3.07$	$-0.63$	**	$-1.36$	$-0.19$	$-0.38$	*	$313.3$	$317.4$	ns
P4	$0.89$	$3.10$	$-0.67$	**	$-1.44$	$0.25$	$-0.54$	*	$311.6$	$319.2$	ns
Pz	$1.22$	$3.43$	$-0.79$	***	$-0.64$	$1.07$	$-0.64$	**	$336.2$	$336.5$	ns
Cz	$-2.41$	$-3.78$	$0.23$	ns	$-8.50$	$-10.85$	$0.37$	ns	$266.6$	$278.7$	ns

and Figure 7.

The largest parietal reductions in peak amplitude were observed at P4 ($3.10$ to $0.89\mathsf{\mu }\mathrm{V}$), P3 ($3.07$ to $1.04\mathsf{\mu }\mathrm{V}$), and Pz ($3.43$ to $1.22\mathsf{\mu }\mathrm{V}$). At frontal channels F3 and F4, the negativity present in the control group was reduced in the AUD group, indicating a global rather than spatially isolated reduction in stimulus-evoked response.

3.5. Bayesian Classification

The Gaussian Naive Bayes classifier trained on five band power features extracted from channel Fz as a conservative single-channel baseline achieved a mean cross-validated ROC AUC of $0.591$ ($\pm 0.128$) across five stratified folds, using a balanced sample ($n=40$ per group) and a uniform prior $P\left(\mathrm{AUD}\right)=0.50$. When extended to all eight analysis channels (40 features total), repeated stratified cross-validation ($10\times 5$ folds, $n=122$) yielded ROC AUC = $0.713\pm 0.088$ (fold score range: $0.526$ to $0.844$) for GNB and ROC AUC = $0.709\pm 0.113$ (fold score range: $0.521$ to $0.887$) for logistic regression, indicating comparable performance between the two approaches. A permutation test (1000 permutations) confirmed that the observed GNB AUC of $0.713$ significantly exceeded chance level ($p=0.001$; permuted mean AUC = $0.503\pm 0.065$), establishing that the classification result was not an artefact of chance alignment between features and labels. Class-conditional likelihood were illustrated with Figure 8. The ROC curve with uncertainty across cross-validation folds is presented in Figure 9.

The delta band showed the greatest separation between AUD and control likelihood distributions, consistent with its status as the band with the largest effect sizes in the inferential analysis (Table 1). Theta and alpha bands showed moderate separation; beta and gamma bands showed near-complete distributional overlap. This correspondence is expected, since the classifier likelihoods are fitted to the same features that were tested, and it is reported here as a consistency check rather than as independent evidence.

As a complementary summary, Figure 10 and Figure 11 present an integrated topographic overview for four of the five frequency bands, with channel-level statistical annotations restricted to the eight primary analysis channels. Each panel displays the scalp distribution of band power for the AUD group, control group, and their difference, with significance markers (*, **, ***) overlaid at F3, F4, C3, C4, P3, P4, Pz, and Cz to indicate the result of the corresponding Mann–Whitney U test reported in Table 1.

This consolidated visualisation allows direct comparison of the spatial and spectral pattern of group differences across the bands shown, reinforcing the observation that the magnitude and significance of AUD-related power reduction follow a consistent gradient: largest and most significant in delta and theta, intermediate in alpha, and minimal or absent in beta and gamma.

4. Summary and Discussion

Our findings demonstrate a broad-spectrum reduction in EEG band power in individuals with AUD relative to healthy controls, spanning delta, theta, alpha, and to a lesser extent beta frequencies, across frontal, central, and parietal scalp regions. This pattern is consistent with the established literature on task-evoked EEG abnormalities in AUD [27,32,40] and is consistent with and complements prior work by providing effect size quantification across a standardised eight-channel montage. These findings are consistent with, and provide effect size quantification complementary to, prior work on EEG abnormalities in AUD. We acknowledge that generalisability beyond the Begleiter dataset cannot be established from a single-database analysis, and we explicitly frame this as a limitation requiring multi-cohort replication.

The most striking finding is the magnitude of the parietal delta deficit, with Hedges’ g reaching $-0.97$ at P3 and $-0.95$ at P4, which is a large effect by conventional psychological research standards and among the largest effect sizes reported in the AUD EEG literature for task-evoked spectral features. Delta oscillations are thought to reflect thalamocortical gating mechanisms that regulate the flow of sensory information to cortical association areas [41]. The pronounced parietal delta reduction observed here is consistent with disrupted sensory gating and attentional filtering in AUD, though the cross-sectional design precludes causal inference.

The frequency-dependent gradient of effect sizes, which is strongest in delta and theta, weaker in alpha and beta, and absent in gamma, is consistent with the hypothesis that AUD preferentially disrupts slow cortical oscillations generated by thalamocortical and corticocortical feedback loops. The spatial specificity of the effects, which are strongest at parietal sites and weakest at the central midline, further supports a network-specific rather than globally diffuse interpretation, consistent with the known density of dopaminergic mesocortical projections to posterior association cortex.

The P300 findings replicate and extend the foundational observation of P300 reduction in AUD first reported by Begleiter and colleagues [25]. The magnitude of attenuation observed here, 74–82 at parietal channels, is substantial, representing a near-complete absence of the canonical P300 positivity in the AUD group at these sites. A theoretical interpretation consistent with the prior literature is that P300 attenuation may reflect impaired dopaminergic facilitation of parietal attention and stimulus-evaluation networks, though the present data cannot establish this mechanistic link directly. This interpretation is consistent with PET imaging studies demonstrating reduced D2 receptor availability in AUD [12].

It is important to note that this study cannot establish whether the observed P300 attenuation precedes AUD onset or reflects a consequence of chronic alcohol exposure. The cross-sectional design does not permit this distinction, though convergence with prior family-risk studies [25,42] suggests the present sample may reflect a combination of pre-existing vulnerability and exposure-related neuroadaptation.

The Gaussian Naive Bayes classifier achieved a cross-validated ROC AUC of $0.591$ using only five band power features from a single frontal channel. This performance is modest but should be interpreted as a lower-bound estimate, given that the model employed only one channel rather than the full eight-channel montage, and the Naive Bayes independence assumption did not exploit known correlations between frequency bands. Comparison with the existing AUD EEG classification literature, which reports ROC AUC values typically in the range of $0.65$ to $0.85$ using more complex feature sets [19,43], suggests that incorporation of P300-derived features and parietal channel information would likely improve classification performance beyond the present preliminary estimate.

A deliberate methodological choice in this study was the selection of computationally lightweight classical signal processing and machine learning techniques over deep learning architectures that dominate much of the recent EEG classification literature. This decision merits explicit discussion in light of growing concern within the machine learning community regarding the environmental and accessibility costs of computationally intensive approaches, a concern formalised under the framework of Green AI [44]. The complete pipeline, including data loading for all 122 subjects, trial-level PSD computation, band power extraction, statistical analysis, repeated stratified cross-validation ($10\times 5$ folds), permutation tests (1000 permutations), and calibration analysis, completed in 258 s ($4.3$ min) on a standard consumer laptop without GPU acceleration.

Convolutional neural networks and recurrent architectures applied to raw or minimally preprocessed EEG signals can achieve strong classification performance, but typically require graphics processing unit (GPU) acceleration, large training datasets to avoid overfitting, and substantial energy expenditure for both training and hyperparameter optimisation. These requirements present a meaningful barrier to entry for researchers in resource-constrained settings, including many clinical research groups and institutions in low- and middle-income countries where the burden of AUD-related healthcare is often highest. They also carry a non-trivial carbon footprint when training is repeated across architecture searches, cross-validation folds, and ablation studies, a cost increasingly scrutinised in the broader AI research literature [45].

The complete analytical pipeline presented in our study, encompassing data parsing, Welch power spectral density estimation, band power extraction, statistical testing, Gaussian Naive Bayes classification with cross-validation, and topographic visualisation, was developed and executed entirely on a standard consumer-grade laptop computer without GPU acceleration, specialised hardware, or cloud computing resources.

The total computation time for the full pipeline, including processing of all 122 subjects, was on the order of minutes rather than hours, reflecting the inherent computational efficiency of classical spectral methods and probabilistic classifiers relative to deep learning alternatives. The Gaussian Naive Bayes classifier in particular requires only the estimation of per-feature means and variances within each class, an operation with negligible computational and memory overhead even as sample size and feature dimensionality scale.

This computational accessibility is relevant to future research in this domain. The present results demonstrate that exploratory group-associated EEG differences can be detected using methods executable on consumer-grade hardware, providing a lightweight baseline against which more computationally intensive approaches can be compared in future work.

This is not to suggest that more complex models are inappropriate for future work in this domain; rather, it is to argue that the choice of model complexity should be justified by demonstrated necessity rather than defaulted to the most computationally intensive available option. The present study establishes that classical, lightweight methods are sufficient to detect statistically robust and clinically interpretable group differences in this dataset, providing a computationally efficient and environmentally responsible baseline against which the marginal benefit of more complex architectures can be evaluated in future research.

The convergence of band power reduction and P300 attenuation, both pointing toward parietal cortical dysfunction with a plausible dopaminergic mechanism, strengthens the case for EEG frequency-domain features as candidate biomarkers for AUD risk. The non-invasive, low-cost, and computationally lightweight nature of the approach presented here suggests that these group-associated EEG features may warrant further investigation as candidate markers. However, the present cross-sectional comparison cannot support conclusions regarding early risk identification, screening utility, or vulnerability markers. Longitudinal and family-risk designs would be required to establish whether these EEG features precede disorder onset or reflect its consequences.

5. Conclusions

This work presents a preliminary investigation of EEG frequency-domain biomarkers for Alcohol Use Disorder, combining power spectral density analysis, P300 event-related potential characterisation, and Bayesian classification within a unified, computationally efficient analytical framework applied to the publicly available Begleiter EEG dataset ($n=122$).

Four principal findings emerge from this analysis:

• The AUD is associated with broad-spectrum EEG power reduction spanning delta, theta, alpha, and beta bands, with the largest effects observed in parietal delta power (Hedges’ $g=-0.97$ at P3 and $g=-0.95$ at P4) and central theta power ($g=-0.97$ at C4).
• P300 amplitude from S2 match (target) trials is significantly attenuated in the AUD group at parietal electrodes (P3, P4, PZ; Hedges’ g range $-0.59$ to $-0.75$), consistent with hypothesised dopaminergic dysregulation of attentional processing networks.
• A Gaussian Naive Bayes classifier using all eight analysis channels achieves ROC AUC = $0.713\pm 0.088$ [$95\%$ CI: $0.526$, $0.833$] under repeated stratified cross-validation ($10\times 5$ folds), confirmed above chance by permutation test ($p=0.001$), suggesting that multichannel spectral EEG features carry group-discriminative information in this exploratory analysis.
• This entire analytical pipeline was executed on consumer-grade hardware without specialised computational resources, demonstrating that exploratory group-associated EEG differences can be detected using lightweight, computationally accessible methods without GPU acceleration or cloud computing.

5.1. Study Limitations

Several limitations of the present study should be considered when interpreting these findings:

• The AUD and control groups were unbalanced ($n=77$ vs. $n=45$), introducing asymmetric statistical power in the descriptive comparisons, though this was addressed through balanced subsampling for the classification analysis.
• The dataset comprises a single recording session per subject, precluding assessment of the temporal stability of the observed EEG signatures.
• The two-dimensional spatial resolution of scalp-recorded EEG limits anatomical specificity, and the topographic patterns reported here do not constitute source-localised estimates of underlying neural generators.
• The Gaussian Naive Bayes classifier’s assumption of conditional feature independence does not capture known inter-band correlations in EEG spectral data, likely constraining classification performance relative to multivariate alternatives.
• The causal direction of the observed associations cannot be determined from this cross-sectional design.
• The analysis is based on a single publicly available dataset (Begleiter, 1995 [25]), and generalisability of the reported effect sizes and classification performance to other AUD cohorts, recording protocols, or clinical settings cannot be assumed without multi-cohort replication.
• The Gaussian Naive Bayes classifier does not automatically produce calibrated probabilities. The uncalibrated model yields a Brier score of $0.272$, worse than the uninformative baseline of $0.25$. Post hoc isotonic calibration improves this to $0.199$, but formal prospective calibration validation and clinically appropriate prior adjustment would be required before the model output could be interpreted as a risk probability in a clinical context.
• The use of a single-segment Welch periodogram, necessitated by the short epoch length of 256 samples, limits spectral smoothing relative to multi-segment approaches. Future work with longer recordings should employ multitaper or multi-segment Welch estimation.
• Band power extracted from task-evoked epochs reflects a mixture of stimulus-locked and non-stimulus-locked neural activity and cannot be directly equated with resting-state power estimates. Future work should employ dedicated resting-state recordings to obtain spectral features independent of task-evoked activity.
• The dopaminergic interpretive framework invoked in the Discussion is theoretical and drawn from the prior literature. The present study does not measure dopamine levels, D2 receptor availability, or genetic markers, and the observed EEG group differences cannot be attributed to a dopaminergic mechanism on the basis of the present data alone.
• Several limitations concerning the P300 analysis must be acknowledged. Trial-averaged ERPs were computed without baseline correction, since the trial records were parsed and averaged in their full recorded form. Peak amplitude measured on a non-baseline-corrected average is sensitive to pre-stimulus offset and slow drift. A systematic group difference in pre-stimulus offset would therefore produce an apparent amplitude difference in the absence of any difference in the evoked response itself. The reported parietal attenuation cannot be separated from this confound on the basis of the present analysis. The P300 results should accordingly be treated as provisional and require confirmation on baseline-corrected epochs.
• A second limitation follows from the amplitude measure. Peak amplitude is an upward-biased estimator whose bias decreases with the number of averaged trials, so unequal trial counts between groups would contribute to the observed difference. Mean amplitude over the measurement window is the more robust alternative and is identified as a required step in any replication.

5.2. Future Research Directions

Several directions for future work emerge from the limitations identified above, while preserving the computational accessibility emphasised in this study:

• Replication in larger, balanced cohorts would improve the precision of effect size estimates and enable more robust classification performance assessment, and could be achieved without departing from lightweight computational methods given the linear scaling properties of the spectral and Bayesian approaches employed here.
• Longitudinal designs incorporating family-risk samples would permit direct evaluation of whether the EEG signatures identified here precede AUD onset and predict future risk.
• Expansion of the feature set to include P300 latency, inter-channel coherence measures, and time–frequency representations would likely improve classification performance while remaining within the computational reach of standard consumer hardware, in contrast to raw-signal deep learning approaches.
• Multivariate classification approaches that explicitly model inter-feature correlations, including regularised logistic regression and linear discriminant analysis, both of which remain computationally lightweight relative to deep learning alternatives, should be compared against the Naive Bayes baseline established here to quantify the performance cost of the independence assumption while maintaining Green AI principles.
• Where deep learning architectures are explored in future extensions of this work, their adoption should be accompanied by explicit reporting of computational cost and energy consumption and justified by a demonstrated performance gain that outweighs the associated environmental and accessibility costs.
• Finally, integration of EEG-derived features with complementary measures within a unified, lightweight Bayesian framework represents a direction for future exploratory research, contingent on longitudinal validation in independent cohorts.