Addiction can be defined as the loss of control over drug use or the compulsive seeking and use of drugs despite adverse consequences. It is a neuropsychiatric disorder caused by substance abuse that is strongly influenced by a person’s genetic structure and the psychological and social context in which drug use occurs. The disease cycle of addiction mainly results from dopaminergic dysfunction, particularly in dopamine (DA) secreted from the mesencephalic ventral tegmentum area (VTA) to the nucleus accumbens (NAcc), prefrontal cortex (PFC), and amygdala. Substances of abuse affect the same neural circuitry as primary biological rewards such as food, water, and sex, leading to substance dependence (1, 2). The understanding of addiction has evolved, from being viewed as a moral condition in DSM-II to being considered more based on psychobiological constructs in DSM-III and beyond. DSM-IV added cognitive factors to the contributing factors of addiction, while DSM-V focused on the psychological changes caused by substance abuse that lead to cognitive impairment (3). Consumed substances have different effects on neural processes, and substance use can lead to increased dopamine consumption in the collecting systems, which becomes an important element in goal-directed behavior and can ultimately result in substance dependence (2).

Nonlinear dynamics methods are widely used to analyze neurophysiological data due to the brain’s inherent complexity, which spans multiple spatial and temporal scales. Higuchi Fractal Dimension (HFD) specifically quantifies the self-similarity and complexity of neural oscillations, offering insights into the balance between stochasticity and determinism in brain dynamics. In the context of addiction, altered neural complexity may reflect disruptions in cognitive control, reward processing, and executive function. Various complexity measures have been developed, including those based on random fractal theory, information theory, and chaos theory. Gao et al. distinguish between chaos and random phenomena and found that the variations of complexity measures with time are either similar or reciprocal in their study on the relations among different complexity measures for EEG (4, 5).

Emotion recognition and addiction detection using electroencephalogram (EEG) signals have garnered increasing interest due to their potential applications in mental health monitoring, affective computing, and neurorehabilitation (6–8). Among various techniques applied to analyze EEG signals, fuzzy logic has emerged as a powerful mathematical framework to manage uncertainty and imprecision. Its inherent ability to model nonlinear and complex systems makes it highly suitable for EEG analysis, which often involves noisy, nonstationary, and high-dimensional data (9–12).

Fuzzy logic has been employed in diverse EEG-related domains, such as brain-computer interfaces (BCIs), neurological disorder diagnosis, and cognitive state assessment. For instance, studies have demonstrated the efficacy of fuzzy logic in classifying visual perception-related EEG signals and diagnosing epilepsy through fuzzy expert systems (13–16). The flexibility of fuzzy set theory has also enhanced feature processing by enabling partial membership-based selection, which improves the robustness of classification tasks. Tools like partition generator functions have been leveraged to filter and transform EEG feature vectors into sparse representations, particularly valuable in multi-instance data scenarios. The Partition generator function can be used in feature processing methods that are based on fuzzy set theory. These methods use the idea of partial set membership to identify relevant features that have high membership degrees for a given class or target variable. By using fuzzy set theory in feature selection, these methods can provide more flexible and accurate results than traditional binary-based methods. This filter supports multi-instance data and can be applied to a given dataset and features for any partition generator to obtain these filtered vectors for all instances. As a result of this, filtered instances are composed of the relevant values and the class attribute (if set in the input data) and rendered as sparse instances (14, 15).

Recent advancements in BCI technologies have further facilitated the integration of fuzzy logic with machine learning. Notably, Dhara et al. (2023) introduced a fuzzy ensemble-based deep learning model that attained over 97% accuracy on the DEAP dataset and 95% on AMIGOS, showcasing the synergy between fuzzy logic and deep learning in refining model predictions (11). In parallel, deep learning has gained significant momentum in EEG-based emotion recognition. Hassouneh et al. (2020) developed a real-time system combining facial expressions and EEG signals using machine learning and neural networks, highlighting the feasibility of accurate multimodal emotion classification (17). A comprehensive review by Chutia and Baruah (2024) emphasized the impact of convolutional and recurrent neural networks in capturing spatiotemporal dynamics of EEG data (6). Similarly, Khare et al. (2024) systematically reviewed a decade of emotion recognition studies, underlining the increasing shift toward hybrid models that integrate EEG with other physiological signals for improved contextualization (7). This trend is echoed in the review by Computers in Biology and Medicine (2023), which notes the growing importance of transfer learning, attention mechanisms, and multimodal fusion techniques in advancing EEG-based emotion detection (8).

Beyond emotion recognition, fuzzy logic and EEG complexity metrics have also proven useful in addiction research. For example, Marvi et al. (2023) applied recurrence quantification analysis and entropy indices to distinguish multidrug users from healthy individuals with 90% accuracy using support vector machines (18). Zhou et al. (2024) further extended these findings by employing wavelet-transformed P300 components and BiLSTM networks to detect methamphetamine abuse with 83.85% accuracy (19). Additionally, Crane et al. (2021) revealed altered neural reward processing in cannabis users via EEG event-related potentials, suggesting that EEG-based measures may serve as biomarkers for addiction-related dysfunctions (20). Finally, Hosseini et al. examined the effects of computer gaming on brain function, offering insights using quantitative EEG (QEEG) complexity analysis (21).

Collectively, these findings confirm that fuzzy logic approaches—whether standalone or hybridized with deep learning—and EEG-based signal complexity analysis substantially enhance both the accuracy and interpretability of brain activity monitoring in emotional and addiction-related research domains.

As mentioned previously, despite the widespread use of EEG in clinical neuroscience, there remains a critical need for reliable, noninvasive tools to assess substance addiction and monitor treatment progression. We introduce a novel EEG dataset comprising recordings from four distinct groups of male participants: active opium addicts, individuals undergoing early treatment (≤3 days), individuals in extended treatment (>2 weeks), and healthy controls. EEG signals were recorded from frontal and parietal regions, preprocessed to remove artifacts, and analyzed using the Higuchi Fractal Dimension to quantify EEG complexity. Additionally, fuzzy logic-enhanced machine learning model was applied to classify subjects based on their addiction stage. Addicted individuals exhibited reduced EEG complexity in regions associated with attention, memory, and executive function. These differences were partially reversed in long-term treated subjects. Fuzzy approach to HFD features of EEG complexity led to high classification accuracy across groups. This study advances the field by addressing limitations in previous works, refining EEG-based complexity analysis with fuzzy logic, and moving toward practical applications in addiction diagnosis and treatment monitoring.

During the preprocessing steps, some records were detected as being corrupted or disturbed; therefore, the final numbers of analyzed participants of different groups were displayed in Table 1 .

This study aims to propose a machine learning (ML) model based on fuzzy logic for analyzing EEG data in order to enhance the understanding of the biological mechanisms underlying opium addiction and enable personalized diagnostics based on individual neural activity patterns. We analyzed EEG data from four groups based on their addiction status and treatment abstinence period by using of extracted complexity features that serve as indicators of neural dynamics in the brain. Classification experiments using different methods and attribute sets showed that MultilayerPerceptron function in combination with WEKA partition membership preprocessing resulted in outstanding results. All participants in this study were male and between 18–40 years of age. While this demographic restriction limits the generalizability of our findings to young adult male opioid users, it also provides an advantage in terms of methodological rigor. By studying a relatively homogeneous group, we were able to partially offset the limitations of a small sample size and reduce the influence of confounding factors such as sex, age, and ethnicity. This approach increases the internal validity of our findings, although future studies with larger and more diverse populations will be needed to confirm their broader applicability.

Table 2 and Supplementary Table S1 present the results of psychological assessments that were conducted to evaluate the mental health and addiction status of the participants.

The DASS questionnaire revealed that despite addiction treatment, the depression factor remained unchanged and did not differ between the control and addiction groups, but it decreased in the healthy group. In terms of anxiety, there was no significant difference between the recently quit group and the control group, but both differed significantly from the healthy group. The stress factor showed no significant difference between the recently quit group and the control group, but both significantly differed from the healthy group. Overall, the DASS questionnaire indicated that distress levels were significantly different between the groups, and even after quitting, the difference persisted.

Regarding the GHQ questionnaire, individuals who had recently quit exhibited higher levels of physical symptoms, which resolved immediately after quitting. The control group, who had quit more recently, showed similarities to the healthy group. The sleep factor also significantly differed between the recently quit group and both the control and healthy groups. Social interaction did not significantly differ between the three groups and did not change after quitting. The depression factor in the GHQ questionnaire aligned with the findings from the DASS questionnaire. In general, the total factors of the GHQ questionnaire demonstrated that addicted individuals had lower general health compared to healthy individuals. Treatment improved general health but still posed challenges compared to the healthy group.

Regarding the OCDUS questionnaire, cravings, and mental preoccupation did not show significance or differences between the recently quit group and the control group. However, the impact of drug use on work and life significantly differed between the recently quit group and the control group. Significant differences were also found in motivation, excitement, and violation of control between the recently quit group and the control group, indicating that emotional experiences were more consciously regulated in different situations. There was no significant difference in resistance to drug use between the two groups. Overall, the OCDUS questionnaire showed a significant difference in factors between the two groups, indicating a decrease in the urge to consume in the past week for both compared groups. However, intention and desire to consume were not significant, and there was high variability in dispersion and standard deviation.

Finally, the DDQ questionnaire did not yield any significant differences between the two groups in all factors.

The fuzzy-rough feature selection method involves using fuzzy logic to calculate the relevance of each feature and then using rough set theory to select the most relevant features. The method was shown to outperform other feature selection methods in terms of accuracy and efficiency (36–38). However, here we used WEKA partition membership as a filter based on fuzzy logic. It is clear from Table 3 that Classification experiments using different methods and attribute sets were conducted, revealing the varying performance of the MultilayerPerceptron, LibSVM, and MLPClassifier methods across different categories. In consistence with our previous finding (21), the performance of MultilayerPerceptron function in combination with WEKA partition membership preprocessing shows outstanding results ( Table 3 ).

The pattern of brain function in people who are using can be different from those who are quitting and healthy people, also the function of the brain can change under the influence of the type of substance consumed (39) For example, by comparing the EEG of alcoholics with the normal control group, an increase in beta power and a decrease in alpha power, as well as a decrease in the ratio of delta to theta, were confirmed (40). Marijuana consumption also leads to instability of the alpha frequency band (41, 42). According to the EEG recorded from people who have recently quit heroin a decrease in alpha, an increase in beta and an increase in Delta and Theta have been encountered in the central areas (43, 44). As a result, the people in this research were divided into four groups, the Addicted group included people who had just visited and the treatment process had not yet started for them, in other words, they were addicted, the First group included people who had less than three days of treatment. The second group included people who had been treated for more than two weeks, and finally the Normal group, who were healthy people who had no history of drug use. Using fuzzy logic techniques, the extracted features were processed to develop an ML model. Thus, we employed HFD algorithms to convert EEG data into scalar values, resulting in a set of features for each participant. These features were then filtered and combined using the Partition Membership method.

To evaluate the ability of different Higuchi variables(features) in discriminating different groups the output of four different feature selection methods is presented in Supplementary Table S2 . Additionally, the study used a statistical analysis with the same aim ( Supplementary Table S3 ). According to these results, we find out that different channels are responsible for the segregation of different groups.

The comparison between the Addict group and the other groups identified significant differences in Higuchi variables across channels 2, 4, 6, 8, 9, 10, 12, 16, and 17. These channels likely correspond to brain regions associated with addictive behaviors and related cognitive processes. Addiction involves complex interactions between several brain regions, including the prefrontal cortex, limbic system, and reward pathways. The observed differences in Higuchi variables within these channels may reflect altered neurophysiological activity linked to addictive tendencies.

The comparison of all groups against each other revealed significant differences in Higuchi variables across channels 2, 4, 5, 8, 9, 10, 16, 17, and 18. These channels potentially correspond to brain regions involved in various neurocognitive functions shared among the groups, such as attention, memory, and executive control. The observed differences in Higuchi variables within these channels suggest variations in the complexity of neural activity, reflecting the diversity of cognitive processes exhibited across the groups.

Comparisons of the First group with other groups revealed distinct sets of EEG channels exhibiting significant differences in Higuchi Fractal Dimension (HFD) variables, reflecting variations in neural complexity. For instance, when comparing the First group with the Addict group, channels 1, 4, 6, 9, 10, 15, 16, 17, and 19 showed significant differences, which may correspond to brain regions implicated in substance use disorders and associated cognitive impairments, such as executive control, reward processing, and attention regulation (45). Similarly, comparisons between the First group and the Other groups (channels 4, 5, 8, 9, 10, 15, 16, 17, and 19), and between the First and Second groups (channels 2, 3, 4, 5, 6, 13, 16, 17, and 19) revealed significant HFD differences, suggesting variations in neurophysiological activity that may underlie the unique cognitive and behavioral profiles of the First group (31).

Additional group comparisons also demonstrated channel-specific differences in HFD values: Normal vs. Addict (channels 2, 6, 9, 11, 12, 13, 17, 18), Normal vs. First (channels 3, 4, 5, 9, 11, 15, 17, 19), Normal vs. Others (channels 4, 8, 9, 10, 11, 16, 17, 18), Normal vs. Second (channels 2, 3, 4, 5, 8, 9, 11, 14, 18), Second vs. Addict (channels 2, 3, 4, 5, 8, 9, 10, 12, 16, 17), and Second vs. Others (channels 2, 3, 5, 8, 9, 10, 14, 16, 17, 18). These results indicate group-specific variations in neural complexity, potentially reflecting differential engagement of brain regions involved in addiction-related cognitive processes, including decision-making, inhibitory control, and reward sensitivity (46). Overall, these findings highlight the utility of HFD-based EEG analysis for capturing subtle neurophysiological differences across addiction stages and treatment groups.

The very high classification performance reported in Table 3 (e.g., Normal vs. Others, F1 = 0.98; Normal vs. Addict, F1 = 0.91) should be interpreted with caution. Although we employed strict 10-fold cross-validation, evaluation on a held-out independent test set, and dimensionality reduction via the Partition Membership filter, the possibility of overfitting cannot be entirely excluded. One potential risk is that the model may have captured non-neural artifacts (e.g., head movements, muscle activity, or subtle recording differences) that coincidentally distinguished the groups. At the same time, the application of fuzzy logic to EEG feature processing constitutes a key methodological innovation of this study. Unlike traditional feature selection methods (e.g., correlation-based or chi-square) that operate on binary inclusion–exclusion rules, fuzzy logic allows for partial membership of features to classes, offering a more flexible and realistic way to address the inherent variability and uncertainty of EEG data. This approach not only enhances classification accuracy by retaining weak but informative signals that would otherwise be discarded under strict thresholding, but also improves interpretability: membership degrees provide an intuitive measure of the relative importance of each feature for group separation. Furthermore, the Partition Membership filter generates sparse feature representations that effectively reduce dimensionality while preserving subtle discriminative patterns, thereby improving robustness to noise. Taken together, these theoretical advantages ( Supplementary Table S1 of Supplementary File 1 ) highlight fuzzy logic as a promising framework for uncovering latent neural patterns that may be overlooked by conventional methods, while also underscoring the need for careful validation to ensure that observed performance reflects genuine brain signal differences rather than overfitting artifacts. It is important to consider that the functional interpretations of these EEG channels and their relationships with neurocognitive and neurophysiological activities may require further investigation. Incorporating additional techniques, such as neuroimaging and comprehensive cognitive assessments, would contribute to a more comprehensive understanding of the brain regions involved and the specific neurocognitive and neurophysiological activities underlying the observed differences in Higuchi variables across these EEG channels. One of the main goals underlying this study is to focus on and interpret the issue from various dimensions, as it is a social problem that can expand to a large-scale social level. In this regard, as is evident in the study, an interdisciplinary study has been carried out by making use of both social sciences and engineering. As a result of the study, it would be a good idea and useful if the artificial intelligence-based model in this study could be connected to online tools that can evaluate the effectiveness of addiction treatment institutions and government spending. In this way, both public and private institutions can be managed effectively, and more accurate and rapid steps can be taken for patients. Therefore, policy development and implementation principles can be presented on the public side. The findings of the study provide important information for better understanding the pathophysiology of opioid addiction and personalizing individual diagnoses. This information can be used in the field of public administration to develop programs to combat substance addiction and to take and implement preventive measures. In particular, new substance addiction policies can be developed to increase the effectiveness of treatment programs and to take individual differences into account in addiction treatment. In addition, this study can make significant contributions for improving health services in public administration, increasing social welfare and security, and activating educational programs.

The second activity in the public administration dimension is the efficient management of health services. With the right plans for this, both the management of treatment centers and resource allocation and planning will become easier. As the third activity, education and awareness raising can be emphasized. In this regard, public awareness programs can be organized about the causes, effects, and treatment methods of addiction. The content of such training programs can be updated and made more effective in light of the findings of the study.

The last activity in public administration is social welfare and security. This may be among the issues that politicians need to deal with directly and put forward solutions. Of course, it may vary from country to country. In this context, the relationship between addiction and crime needs to be evaluated. Considering the impact of substance abuse on crime rates, effective management of addiction treatment can increase social safety. Public administration can develop effective policies for the rehabilitation and reintegration of individuals struggling with addiction. Social services and support programs can be developed to ensure the integration of individuals receiving addiction treatment into society. In this context, the results of the study can be used to increase the effectiveness of social service models used in the fight against addiction. Also, dependent people will not be excluded. Therefore, with these four basic activities in public administration, important steps will be taken towards the United Nations Sustainable Development Goals.

This study presents a machine learning framework integrating fuzzy logic to analyze EEG data, offering an objective tool for assessing addiction severity and treatment progress. By extracting Higuchi Fractal Dimension features, refining them with the Partition Membership method, and applying classifiers, we identified Multilayer Perceptron as the most effective model for distinguishing addiction stages. Traditional addiction assessments rely heavily on self-reported data, which can be incomplete, biased, or unreliable. The challenge of obtaining accurate treatment histories underscores the value of our approach: EEG-based complexity measures combined with fuzzy logic provide an objective and data-driven complement to traditional demographic and clinical assessments, helping to reduce dependence on self-explanation by addicted individuals. The findings highlight persistent brain function disruptions even after treatment initiation, emphasizing the need for objective, brain-based biomarkers. Overall, this work demonstrates the potential of AI-driven neurophysiological analysis to strengthen addiction research. Future efforts should prioritize broader validation and explore real-time monitoring applications, paving the way toward more personalized and effective treatment strategies.