Fractal-Based Functional Connectivity Analysis of Electroencephalography Signals for Alcohol Use Disorder Characterization

Alcohol consumption is widespread globally, and excessive intake can result in significant neurological, cognitive, and behavioral impairments [1, 2]. Prolonged alcohol dependence, clinically categorized as Alcohol Use Disorder (AUD), has been associated with structural and functional alterations in the brain that adversely affect cognitive processing, motor coordination, memory, and emotional regulation [3-7]. According to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) [8], AUD diagnosis is primarily based on symptom clusters and behavioral assessments, which remain inherently subjective and vulnerable to reporting bias.

AUD has been consistently associated with disruptions in large-scale brain network organization, including altered functional connectivity (FC), impaired inter-regional communication, and reduced efficiency of information transfer [9-11]. In addition to these connectivity-level abnormalities, prior studies have reported that AUD is characterized by alterations in the intrinsic dynamics of electroencephalograph (EEG) signals, including reduced signal complexity and modified nonlinear behavior [12, 13].

From a signal processing perspective, EEG signals exhibit scale-free and self-similar characteristics, reflecting long-range temporal dependencies and broadband background activity that are not fully captured by conventional oscillatory or frequency-domain analyses. Fractal decomposition provides a principled framework for separating oscillatory and non-oscillatory components, thereby enabling the characterization of both rhythmic activity and scale-invariant dynamics within the same signal representation. In the context of AUD, where both FC and signal complexity are affected, such decomposition facilitates a more comprehensive analysis of neural interactions by isolating components that may differentially reflect pathological alterations. Consequently, the use of fractal decomposition in this study is motivated by its ability to capture underlying scale-free properties of EEG signals that are directly relevant to AUD-related neural dysfunction.

Recent EEG-based studies have demonstrated that alterations in rhythmic activity and inter-regional coupling can serve as informative markers for neurological and neuropsychiatric conditions. In particular, FC analysis provides a principled framework for characterizing coordinated activity between spatially distributed brain regions, thereby enabling network-level interpretation of neural signals [14, 15]. Alterations in FC patterns have been consistently reported across a range of neurological and neuropsychiatric conditions, including depression, schizophrenia, epilepsy, stroke, and substance-related disorders, highlighting the diagnostic potential of connectivity-based signal representations [16-20].

In particular, FC analysis provides a framework for characterizing coordinated activity between spatially distributed brain regions, enabling network-level interpretation of neural signals. Variations in FC patterns have been reported in conditions such as depression, schizophrenia, stroke, and substance-related disorders, highlighting the diagnostic potential of connectivity-based signal representations [15, 18, 21-24].

1.1 Related work on Alcohol Use Disorder classification

Recent studies have increasingly highlighted the potential of FC derived from EEG signals as a promising feature for classifying various neuropsychiatric disorders, including AUD. For example, one study involving EEG data from 24 patients with Major Depressive Disorder (MDD) and 29 healthy controls applied intrinsic time-scale decomposition (ITD) and a graph-based adaptive least absolute shrinkage and selection operator (GA-LASSO) model to extract time–frequency characteristics of FC for classification [25]. Several classifiers were evaluated, including Support Vector Machine (SVM), k-Nearest Neighbour (kNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN), with SVM achieving the highest accuracy of 97.43%.

Similarly, another study used FC features as input to a SelectKBest algorithm combined with a CNN to classify autism spectrum disorder (ASD) [26]. Resting-state EEG data from 189 subjects (97 with ASD and 92 healthy controls) were analyzed, achieving an accuracy of 81.08%. These findings emphasize the importance of FC in capturing brain activity and its effectiveness as a biological feature for neuropsychiatric disorder classification.

In 2022, a study employed a multilayer discrete wavelet transform (DWT) to decompose, denoise, and extract EEG features for AUD classification from 122 subjects [27]. As a time–frequency analysis model, DWT enables the identification of key signal characteristics across both domains. A hybrid deep learning model combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks (CNN–Bi-LSTM) was then applied, achieving an accuracy of 99.32%. In earlier work, AUD classification using EEG signals was also explored with effective connectivity (EC) as a biomarker rather than FC [28]. In particular, the study [28] used Partial Directed Coherence (PDC) as the EC feature extraction method, paired with a 3D-CNN classifier on data from 31 AUD patients and 31 healthy controls, yielding an accuracy of 87.85 ± 4.64%. By employing frequency-domain analysis, the study highlighted the strength of causal interactions between brain regions within the default mode network (DMN) as a distinguishing feature between AUD patients and healthy controls (HC).

Recent studies have also explored fractal analysis of EEG signals as an effective analytical feature for machine learning and deep learning algorithms. In one such study [13], discriminative features were extracted from EEG signals of 10 healthy subjects and 10 AUD patients using the Davies–Bouldin (DB) criterion, which assesses the separability of feature clusters based on their distribution. The results of the DB analysis were subsequently fed into two classifiers—SVM and kNN—achieving accuracies of 98.77% and 98.5%, respectively, thereby demonstrating the effectiveness of fractal features as reliable indicators for AUD classification using EEG signals.

Another study applied Hurst exponent estimation and adaptive fractal analysis (AFA) of EEG signals to classify epileptic and non-epileptic patients [29]. Hurst exponent-based time-domain analysis was employed to examine the signal’s scaling properties, while Adaptive decomposition analysis (ADA) was utilized for its robustness to noise. The study [29] used an extensive dataset of 11,500 samples with 178 features each. The fractal analysis output was then fed into a Support Vector Regression (SVR) model and a multilayer Gradient Boosting Decision Tree (GBDT), achieving a peak accuracy of 98.23%.

Wavelet-based fractal analysis of fMRI data has also emerged as a promising approach for diagnostic modeling [30]. Another study used fMRI signals to capture the fractal behavior of FC features for Alzheimer’s disease classification [31]. This comparative study included two datasets: the first with 21 normal controls and 35 AD patients, and the second with 30 normal controls and 30 AD patients. Wavelet-based fractal analysis was conducted to decompose the fMRI signals into fractal and non-fractal components for feature extraction. In addition, Pearson correlation was employed as a baseline method for comparison. The resulting features were then used to train an SVM classifier for the classification task. A maximum accuracy of 90.3% was achieved using the non-fractal-based method, highlighting that non-fractal components are effective for representing spontaneous FC in characteristics neurological disorders.

Hadiyoso et al. [32] examined EEG signal complexity for early Alzheimer’s Dementia detection using entropy and fractal features from resting-state EEG. Spectral entropy and fractal dimensions (Higuchi, Katz, and Sevcik) were extracted from 19-channel EEG recordings of 34 subjects (normal elderly and MCI). Using Linear Discriminant Analysis, the study found reduced EEG complexity in MCI subjects, particularly in frontal and occipital regions. The combined entropy–fractal feature set achieved 82.4% accuracy, 87.5% sensitivity, and 77.8% specificity, demonstrating the effectiveness of complexity-based features in capturing subtle neural changes while reducing sensitivity to low-frequency noise.

This work reinforces the relevance of complexity-driven EEG signal analysis as a complementary methodology for neurodegenerative disorder assessment and motivates further exploration of fractal-informed signal features at higher representational levels, such as FC analysis.

Beyond conventional time–frequency analysis, EEG signals are known to exhibit scale-free and self-similar characteristics, often described using fractal models [30, 33]. Fractal analysis enables the separation of broadband, non-oscillatory background activity from narrowband oscillatory components, thereby offering a more comprehensive representation of neural signal structure. While fractal measures have previously been employed as standalone features for classification, their integration into connectivity-level signal analysis remains relatively underexplored.

Motivated by this gap, the present study proposes a fractal-based FC framework for EEG analysis in AUD. By applying fractal decomposition prior to connectivity estimation, the proposed approach isolates oscillatory and scale-free components and examines their respective contributions to network-level interactions within the DMN. This signal-centric methodology enables the extraction of biologically meaningful connectivity features that reflect long-range temporal dependencies and coordination patterns often obscured in conventional analyses.

Machine learning classifiers are employed in this study solely as quantitative evaluation tools to assess the discriminative capability of the extracted signal features. The main contribution of this work lies in the systematic application and evaluation of fractal decomposition within a FC framework for EEG analysis in AUD. Specifically, the study examines how separating oscillatory and scale-free components influences connectivity-based feature representations, and compares their respective contributions through an ablation-style analysis. By providing a structured comparison of original, fractal, and non-fractal connectivity features within the same analytical pipeline, this work offers additional insight into the role of scale-invariant dynamics in EEG-based connectivity analysis, rather than introducing a fundamentally new methodological approach.

1.2 Default mode network channels

The brain network associated with cognitive functioning is known as the DMN. According to Menon [34], the DMN is closely associated with cognitive functions, particularly retrospective processes such as episodic memory, social cognition, and semantic memory, which are sustained over time—making it a suitable framework for detecting neural irregularities during simulated tasks. Resting-state EEG signals, especially those originating from the DMN, exhibit intrinsic dynamics that not only reflect brain adaptability but also serve as predictors of cognitive performance [35, 36]. These patterns have been widely used to investigate the pathophysiology of psychological disorders such as post-traumatic stress disorder (PTSD), schizophrenia (SZ), and MDD [37, 38].

The DMN primarily comprises key brain regions such as the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), and the precuneus, which maintains FC with the inferior parietal cortex (IPC) [34, 39], as illustrated in Figure 1. These regions collectively underpin the brain’s baseline neural activity and are crucial for internally directed cognitive processes such as self-referential thinking and memory consolidation [39]. Rather than assessing isolation, the correlation of these regions was studied to provide insight into brain activity, known as FC [37]. To capture this connectivity, EEG channels located nearest to canonical DMN regions were identified and mapped to corresponding Brodmann areas [40]. Consequently, six EEG electrodes (Pz, Fz, F3, F4, P3, and P4) were exclusively selected, as they provide the most reliable representation of the DMN’s functional architecture and are therefore most appropriate for feature extraction and detailed analysis [28], as summarized in Table 1. This electrode selection is further supported by prior EEG and neuroimaging studies that have established spatial correspondences between scalp-recorded signals and underlying DMN regions. Frontal electrodes such as Fz, F3, and F4 are commonly associated with activity in the medial prefrontal cortex (mPFC), a core hub of the DMN involved in self-referential and executive processes [7, 34]. Similarly, parietal electrodes including Pz, P3, and P4 have been linked to posterior DMN regions such as the precuneus and PCC, which play a central role in internally directed cognition and memory integration [36, 39]. EEG-fMRI studies have further demonstrated that FC patterns observed at these electrode sites reflect large-scale network interactions consistent with DMN organization [37]. Moreover, alterations in frontal–parietal connectivity involving these regions have been consistently reported in neuropsychiatric conditions, including AUD, reinforcing their relevance for capturing network-level dysfunction [28, 41].

Figure 1. Lateral and medial view of the default mode network (DMN) of the left hemisphere (adapted and modified from [37])

Table 1. Electroencephalograph (EEG) electrode locations corresponding to default mode network (DMN) regions and associated Brodmann areas used for feature extraction. L and R denote the left and right hemispheres, respectively

The proposed method is evaluated across three dimensions: (i) feature-based analysis, which assesses the discriminative power of individual components—original, fractal, and non-fractal—using their respective feature vectors; (ii) sample-based testing, which measures overall classification performance using randomized training and testing splits; and (iii) subject-based testing, which examines subject dependency by reserving specific individuals exclusively for testing. All classifications are performed using two selected classifiers; Optimizable SVM and Ensemble Learning.

3.1 Linear and log-log scale power spectra analysis of original, fractal, and non-fractal connectivity

Power spectral density (PSD) analysis reveals how signal power is distributed across frequencies, capturing both rhythmic oscillations and background activity. The linear scale enables direct comparison of absolute power, making dominant rhythms and amplitudes more visible, while the log–log scale highlights scaling relationships and separates fractal background activity from oscillatory peaks. Using both views provides a fuller understanding of spectral patterns and the effectiveness of fractal decomposition in distinguishing neural components between AUD and HC subjects. Figure 8 shows the power spectra of a 2-second EEG segment for an AUD subject (top) and an HC subject (bottom) after fractal decomposition, in linear scale and log–log scale.

Figure 8. Power spectra in linear (left) and log-log (right) scale for fractal decomposition of a 2-second EEG segment from Alcohol Use Disorder (AUD) (top) and healthy controls (HC) (bottom). Each subplot compares the original, fractal, and non-fractal components, highlighting spectral differences between the AUD and HC

For the linear scale in Figure 8, the AUD subject exhibits a much higher low-frequency amplitude (16 × 10⁶) with a large peak near 0 Hz, while the HC subject shows a lower amplitude (5 × 10⁵) but a broader distribution of power between 0 and 0.4 Hz. Smaller peaks appear near 0.4 Hz in AUD and 0.1–0.2 Hz in HC, suggesting specific oscillatory components beyond the primary slow-wave peak. These patterns indicate reduced variety in rhythmic activity and stronger fractal background power in AUD, whereas HC displays more varied rhythms, which may be consistent with healthier and more active neural dynamics [42]. In AUD, the original component closely follows the fractal curve across 0–0.4 Hz, masking much of the non-fractal contribution. In contrast, the original component of the HC subject shows a clearer mix of both fractal and non-fractal patterns, with more distinct oscillatory peaks.

For the log–log scale in Figure 8, the AUD subject’s spectrum shows irregular deviations from the expected smooth slope, including a marked drop in power near the frequency corresponding to −1 on the log scale. Such irregularities may reflect unstable oscillatory activity or increased noise influence. In contrast, the HC subject displays a more stable slope with well-defined oscillatory peaks, indicating more consistent rhythmic activity across frequencies. Moreover, the relationship between components shifts across frequencies. At low frequencies (around −6 to −2 on the log scale), the original component closely follows the fractal curve, indicating that fractal, non-oscillatory background activity dominates this range. However, from approximately −1 onwards, the fractal and non-fractal components align more closely with each other than with the original. This convergence reflects the reduced influence of the fractal background at higher frequencies, where the decomposition outputs share similar scaling behaviour. This characteristic is consistent with that observed in the linear plot.

3.2 P-value analysis of original, fractal, and non-fractal connectivity

To evaluate the discriminative capability of effective connectivity features between the Healthy Control (HC) and AUD groups, a statistical significance analysis was conducted using one-way ANOVA in MATLAB. A total of 45 connectivity features were extracted from pairwise interactions among six EEG channels within the DMN, comprising 15 original, 15 fractal, and 15 non-fractal features. The statistically significant connections with zero or near-zero p-values are summarized in Table 2. Following Benjamini–Hochberg FDR correction (q = 0.05), all previously identified significant connections (including those with near-zero p-values of 10⁻⁷ and 10⁻⁶) remained significant after FDR correction, indicating that the findings are not due to multiple-comparison effects.

For both the original and fractal representations, the same eight connectivity links exhibited p-values equal to zero, indicating highly significant group-wise differences. These links primarily involve frontal–parietal and inter-frontal interactions, suggesting consistent alterations in long-range and frontal connectivity in the AUD group. Such disruptions are consistent with impaired frontoparietal executive networks reported in AUD, which underlie deficits in cognitive control and decision-making [28, 46]. In the non-fractal representation, six of these eight connections also retained zero p-values, while the remaining two links (Pz↔F3 and Pz↔F4) showed slightly higher but still highly significant p-values (10⁻⁷ and 10⁻⁶, respectively). This indicates a minor reduction in statistical sensitivity for these specific connections when isolating the non-fractal component. Alterations involving frontal and parietal DMN nodes further align with evidence of DMN dysregulation in AUD, which has been linked to impairments in attention, memory, and executive function [47, 48].

These findings are in agreement with the power spectral analysis presented in Section 3.1. The original and non-fractal components preserved clearer oscillatory characteristics, which are known to be closely associated with neural communication and information transfer. In contrast, the fractal component predominantly captured broadband background activity, resulting in comparatively weaker discriminative capability despite the presence of statistically significant links. Moreover, the pronounced low-frequency dominance observed in the AUD group spectra further supports the observed p-value patterns, reinforcing the role of altered frontal–parietal connectivity in AUD-related neural dysfunction. Collectively, these results indicate that disrupted frontoparietal and DMN connectivity reflects impaired executive control and network-level dysregulation, key neurobiological mechanisms underlying AUD [28, 49].

Table 2. Comparison of effective connectivity measures between Healthy Control (HC) and Alcohol Use Disorder (AUD) Groups having zero p-value

3.3 Comparative analysis of original, fractal, and non-fractal connectivity in Alcohol Use Disorder vs. healthy controls classification

The feature sets for the non-fractal, fractal, and original FC were divided into feature vectors of equal dimension (7717 × 15) and evaluated over 100 iterations. The results are presented in Table 3 and Table 4. In terms of classification accuracy, the original and non-fractal features exhibited comparable performance across both models, with the original feature set achieving the highest overall accuracy of 78.32% for the optimizable SVM and 80.23% for the Ensemble model. These values represent an advantage of 0.8% and 0.7%, respectively, over the non-fractal features, further supporting the role of fractal decomposition by showing how separating components influences discriminative power, suggesting that the original feature set retains more discriminative information relevant to distinguishing AUD from HC subjects. Furthermore, the original feature set yielded superior specificity (87.05%) and precision (82.02%) with the Ensemble model, indicating stronger discriminative capability, particularly in minimizing false positives.

However, this gain in accuracy and specificity came at the expense of sensitivity, where both models showed reduced performance—70.78% for the SVM and 74.48% for the Ensemble—indicating a higher likelihood of false negatives in AUD detection. In contrast, the non-fractal feature set offered a more balanced trade-off, improving sensitivity (75.65% for the Ensemble) while maintaining competitive accuracy (79.53%) and stability across other metrics. Fractal features alone yielded the lowest sensitivity (68.54% for SVM and 73.10% for the Ensemble), although they maintained reasonable specificity and precision, which suggests that fractal components contribute complementary information rather than acting as a standalone discriminative feature set.

Table 3. Mean classification accuracy (in %) with standard deviation for original, fractal, and nonfractal functional connectivity (FC) features, evaluated using Optimizable Support Vector Machine (SVM) and Ensemble classifiers

Table 4. Sensitivity, specificity, accuracy, and precision (in %)—for original, fractal, and non-fractal functional connectivity (FC) features using Optimizable Support Vector Machine (SVM) and Ensemble classifiers

Ultimately, while the original feature set produces higher accuracy, specificity, and precision values, it does so at the expense of sensitivity, leading to greater performance variability. Conversely, the non-fractal feature set offers a better balance across all performance metrics, maintaining lower variability while achieving comparable accuracy to the original feature classifications, thereby reinforcing the value of the ablation framework in revealing the distinct and complementary roles of fractal and non-fractal components.

From a clinical perspective, the observed trade-off between sensitivity and specificity has important implications for the practical utility of the proposed approach. Sensitivity reflects the ability to correctly identify AUD subjects, which is critical for early detection and timely intervention, as missed cases (false negatives) may delay treatment and worsen clinical outcomes. In contrast, specificity represents the ability to correctly identify healthy individuals, thereby reducing false positives and minimizing unnecessary psychological burden or clinical follow-up. In this study, the original feature set achieved higher specificity and precision but at the cost of reduced sensitivity, indicating a tendency toward conservative classification that favors minimizing false positives. Conversely, the non-fractal feature set provided improved sensitivity with slightly lower specificity, suggesting a more balanced detection capability. In real-world clinical screening scenarios, a higher sensitivity is often preferred to ensure that individuals with AUD are not overlooked, even if it leads to a moderate increase in false positives. However, for confirmatory diagnostic settings, higher specificity may be more desirable to avoid misclassification of healthy individuals. Therefore, the choice of feature representation and classification strategy should be guided by the intended clinical application, with the non-fractal features offering a more balanced trade-off and the original features favoring specificity-driven decision-making.

3.4 Sample-based evaluation of Alcohol Use Disorder vs. healthy controls classification using fractal-based functional connectivity

Table 5 presents the sample-based evaluation results for AUD vs. HC classification using fractal-based FC features. Overall, the Ensemble model demonstrates superior performance across most evaluation metrics, achieving the highest average accuracy (85.86%) compared to the Optimizable SVM (83.94%). This advantage extends to precision (87.88% vs. 85.98%) and sensitivity (81.26% vs. 78.96%), indicating the Ensemble’s stronger ability to capture true positives. The higher AUC score of 0.904, relative to 0.8345 for the SVM, further underscores its superior discriminative capacity and adaptability to complex, high-dimensional EEG features.

The Ensemble’s advantage can be attributed to its integration of multiple classifiers, which provides a more comprehensive interpretation of the fractal-based connectivity patterns. In contrast, the Optimizable SVM, while effective, relies on a single decision boundary, which may limit its ability to generalize across the variability inherent in EEG-derived connectivity features.

Table 5. Sensitivity, specificity, accuracy, precision (in %) and AUC of fractal-based functional connectivity (FC) classification using Optimizable Support Vector Machine (SVM) and Ensemble classifiers

Notably, the SVM shows a relative strength in specificity, achieving 88.45%, only slightly lower than the Ensemble’s 89.98%. This reflects the SVM’s conservative bias in classifying HC subjects, thereby reducing the likelihood of false positives. Such a property can be valuable in clinical settings, where minimizing the misdiagnosis of healthy individuals is important. Ultimately, while both models exhibit distinct strengths, the Ensemble model performs slightly better than the Optimizable SVM, indicating enhanced adaptability to the complexity and high dimensionality of EEG data and providing a higher level of confidence in AUD patient diagnosis.

3.5 Subject-based evaluation of Alcohol Use Disorder vs. healthy controls classification using fractal-based functional connectivity

Resting-state FC features derived from EEG signals effectively capture subject-specific neural patterns associated with AUD and HC, supporting their utility in subject-level classification and evaluation [28, 50]. The dataset was divided and tested under this paradigm, with classification accuracy ranging from 30.3% to 100% for the Optimizable SVM model and from 28.30% to 100% for the Ensemble model, as shown in Table 6. Taking 70% as the baseline accuracy, the results indicate instances of misclassification and inconsistency across different subjects. The Optimizable SVM misclassified three HC subjects (20% HC misclassification rate) and one AUD subject (6.7% AUD misclassification rate), whereas the Ensemble model misclassified only two HC subjects (13.3% HC misclassification rate) and correctly classified all AUD cases (0% AUD misclassification rate).

Despite slightly higher variance in a few HC predictions, the Ensemble model consistently outperforms or matches the SVM, with its perfect AUD classification further highlighting its robustness. These findings suggest that although both models are influenced by inter-subject variability, the Ensemble model offers superior generalization and greater diagnostic reliability, particularly in AUD detection.

Analysis of the extended confusion matrix in Figure 9 reveals that although both models demonstrate stronger performance in classifying AUD subjects than HC subjects, the Ensemble model consistently outperforms the Optimizable SVM across key metrics. Specifically, it achieves higher AUD classification accuracy (93.89% vs. 91.96%) and better HC classification precision (83.15% vs. 81.12%).

This consistent advantage suggests that the Ensemble model employs a more flexible decision boundary, better capturing interclass variability, particularly within the more heterogeneous HC group. It also demonstrates higher specificity (94.53% vs. 92.85%), reducing false positives—an essential consideration in clinical settings to avoid mislabelling healthy individuals.

Further, the Ensemble model demonstrates higher precision (81.46%) compared to the SVM (79.10%), along with lower false positive (5.47% vs. 7.15%) and false negative rates (6.11% vs. 8.04%). Notably, the reduction in false negatives is clinically significant, as misclassifying AUD subjects as healthy could delay timely diagnosis and treatment. Although the improvements of approximately 2% may appear modest, they contribute to a higher overall accuracy (87.89% vs. 85.87%), highlighting the Ensemble model’s superior reliability and robustness for AUD diagnosis.

Table 6. Classification performance of 15 healthy controls (test subjects 1 to 15) and 15 Alcohol Use Disorder (AUD) (test subjects 16 to 30)

Table 7. Comparison of proposed technique with related works

Figure 9. Extended confusion matrices for AUD classification using fractal-based functional connectivity (FC) with Optimizable SVM (left) and Ensemble (right). HC and AUD are encoded as 0 and 1, respectively; the vertical axis denotes the true class and the horizontal axis denotes the predicted class

Generally, both models demonstrate broadly comparable response patterns, indicating a modest degree of subject dependency in their classification behavior. A more detailed analysis of variability across subjects reveals that differences in classification performance may be attributed to several factors, including signal quality and inter-subject heterogeneity. Variations in EEG signal quality, such as residual artifacts, noise contamination, and differences in recording conditions, can influence the stability of extracted connectivity features. In addition, inter-subject heterogeneity in neural dynamics, including differences in AUD severity, duration of alcohol exposure, and individual brain network organization, may contribute to inconsistencies in classification outcomes. These factors are particularly relevant in EEG-based studies, where both physiological variability and acquisition-related noise can significantly impact model generalization [14, 36]. The observed variability highlights the importance of developing more robust feature representations and incorporating subject-independent validation strategies to improve the reliability of EEG-based AUD classification.

The Ensemble model consistently delivers more favorable outcomes across precision, specificity, and class-wise accuracy. These characteristics position it as a more reliable tool for AUD classification, meriting further exploration and validation in larger, more diverse datasets to assess its potential for deployment in real-world clinical environments.

3.6 Performance comparison with related work

Table 7 presents a comparison of the proposed approach with representative studies in EEG- and neuroimaging-based classification. It should be noted that direct numerical comparison across studies is inherently limited due to differences in datasets, experimental protocols, feature extraction methods, and evaluation settings. The studies included in this table vary in terms of data modality (EEG vs. fMRI), task conditions (resting-state vs. task-based), sample size, and validation strategies, which may influence reported performance metrics.

Accordingly, the purpose of this comparison is not to establish a strict benchmark, but rather to provide contextual insight into how the proposed method relates to existing approaches in the literature. In particular, the table highlights the diversity of methodologies employed, including time–frequency analysis, deep learning architectures, and connectivity-based techniques.

Within this broader context, the proposed fractal-informed FC approach demonstrates competitive performance while maintaining a relatively simple and interpretable signal-processing pipeline. Unlike more complex deep learning models, the present method focuses on feature representation by separating oscillatory and scale-free components prior to connectivity estimation, enabling a more structured analysis of EEG signal characteristics.

Nevertheless, a more rigorous evaluation would require implementing standard FC pipelines and alternative feature extraction methods on the same dataset for direct comparison. Such a controlled benchmarking study is beyond the scope of the current work but represents an important direction for future research to further validate the effectiveness of the proposed approach.