Identifying Methamphetamine Users through EEG Analysis: Harnessing hctsa and Machine Learning Approaches

doi:10.21203/rs.3.rs-3052453/v1

Download PDF

Research Article

Identifying Methamphetamine Users through EEG Analysis: Harnessing hctsa and Machine Learning Approaches

https://doi.org/10.21203/rs.3.rs-3052453/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The objective of this study was to evaluate the potential of accurately distinguishing methamphetamine users from a cohort of healthy individuals by analyzing electroencephalography (EEG) signals and utilizing machine learning techniques. Ten participants with methamphetamine dependence and nine healthy individuals were subjected to a 19-channel EEG recording. A highly comparative time series analysis (hctsa) method was employed for feature extraction from the EEG signals. Subsequently, three machine learning techniques, namely logistic regression (LR), support vector machine (SVM), and random forest (RF), were implemented to process the data. A nine-fold cross-validation approach was utilized to prevent overfitting during the training process. Using the hctsa method, 6,070 features were extracted while discarding 1,682 erroneous or valueless data points. Forty informative features were selected for machine learning implementation. Although single features did not achieve 100% accuracy, combinations of two features resulted in two distinct states predicting values with 100% accuracy when employing the SVM approach. With three-feature combinations, SVM, LR, and RF techniques reached 100% accuracy in 134, 89, and 100 states respectively. The inclusion of four-feature combinations further increased these numbers, with SVM, LR, and RF achieving 100% accuracy in 2933, 3109, and 589 states respectively. Notably, only LR achieved 100% accuracy when using all 40 features. This study demonstrated that SVM, LR, and RF classifiers combined with feature extraction through the hctsa method exhibit an exceptional capacity to accurately identify methamphetamine users among healthy individuals using a single EEG channel with a classification accuracy of up to 100%.

Electroencephalography. Methamphetamine. Machine learning. hctsa

Methamphetamine (MA) is a widely abused substance that acts as a brain stimulant, emulating the physiological effects of monoamines like dopamine. The activation of the brain's reward system by MA, coupled with its affordability, primarily contributes to the intense cravings for MA. This can lead to dependence and addiction on a global scale (Khajehpour et al., 2019). Methamphetamine-dependent individuals (MDI) often suffer from various psychological issues, such as cognitive impairments and neurobehavioral disturbances, due to brain structural damage. This damage may include cortical loss, disruptions in normal cortical functions, and alterations in dopaminergic pathways connecting the limbic system, cingulate gray matter, putamen, caudate, and striatum nuclei (Shafiee-Kandjani et al., 2020). As a result of these pathological changes, chronic MA abuse can affect both neurotransmitter systems and neuro-electrical activity within the brain (Chen et al., 2020). Substance dependence disorders are characterized by their chronic relapsing nature and long-term health consequences. They pose significant challenges to society by contributing to a wide range of issues such as illness, criminal activity, accidents, domestic violence, and homelessness among others (Shahmohammadi et al., 2016). Developing a reliable method to distinguish MDI from normal controls (NC) based on noninvasive techniques could serve as an invaluable tool.

Time series, generated by collecting sequential measurements of a specific variable over time, is a prevalent method across various scientific disciplines and industries. The types of time series analyzed range widely (Fulcher et al., 2013). The highly comparative time series analysis (hctsa) software provides a systematic, algorithm-based platform that enables the calculation of numerous structural attributes from a single time series. These characteristics include fundamental distribution statistics, linear correlation configurations, stationarity, entropy, and information-theoretic metrics. The techniques used are derived from the nonlinear time-series analysis domain in physics as well as linear and nonlinear model fitting among others (Fulcher and Jones, 2017; Fulcher, 2013). hctsa has recently garnered considerable attention in electroencephalography (EEG) signal processing for applications such as sleep staging (Decat et al., 2022), cognitive impairment detection (Khatun et al., 2019), and visuomotor tracking state detection (Kuatsjah et al., 2019).

Machine learning (ML) originates from the convergence of statistics -which aims to identify relationships within data -and computer science- which emphasizes effective computational algorithm development (Sidey-Gibbons and Sidey-Gibbons, 2019). The application of ML for automated analysis for medical purposes is gaining significant interest, particularly in clinical diagnostics that rely on EEG data (Gemein et al., 2020). These techniques has been widely used in various fields of medicine such as seizure detection (Jiang et al., 2022; Thakare et al., 2023) and emotion recognition (Fang et al., 2021; Yalamanchili et al., 2022) using EEG data, diagnosis of cardiac disease using electrocardiograph (Sharma et al., 2023; Zhang et al., 2023), and also image classification (Menaka and Vaidyanathan, 2022; Samraj et al., 2023). In this research, we examined three ML methods with the top 40 features extracted from hctsa to assess the differences between healthy individuals and those with MA dependence.

2.1 Sampling

We used EEG recordings of MA users in abstinent periods and healthy individuals (Shafiee-Kandjani, 2020). The selected MA users had an average duration of 5 years of MA use and 7 to 9 days of abstinent period. All males aged between 20 and 45 years. Participants were recruited from the Razi Mental Hospital in Tabriz, Iran, with patients selected from the inpatient population and healthy individuals selected from the outpatient population. Inclusion criteria for the study required patients to have a history of prolonged MA use, lasting at least one year, as well as a positive MA screening test within the past month. Exclusion criteria for participants consisted of a history of seizures, a positive history of neurological disorders, cognitive impairment, long-term or high-dose intake of sodium valproate, chlorpromazine, lamotrigine, benzodiazepines, or topiramate. The maximum permissible doses of these medications were 0.5 mg alprazolam, 25 mg chlorpromazine, 250 mg sodium valproate, 25 mg lamotrigine, 1 mg lorazepam, 25 mg topiramate and 0.25 mg clonazepam. Furthermore, potential participants were excluded if they had a history of any mental illness or had tested positive for morphine within the last 90 days. After preprocessing the row data, we extracted the recordings of nine healthy and ten MDI.

2.2 EEG Recording

In this study, the EEG signals were acquired using a 19-channel silver-chloride (Ag/Cl) electrode array, in accordance with the conventional 10–20 system. The array was connected to an EEG amplifier (Mitsar, Russia) for signal enhancement. During the EEG recording process, a linked-ear reference was employed to maintain electrode impedance below 5 kΩ consistently. The acquired signals were sampled at a frequency of 500 Hz to ensure accurate data representation.

Both control and experimental groups participated in a 10-minute EEG recording session under resting conditions with eyes closed. This approach aimed to capture cerebral activity effectively and reliably. For participants with a history of MA use, the EEG recording session was conducted during their first week of admission, before administering increased dosages of psychiatric medications. This timing ensured that potential confounding factors were minimized and the recorded data accurately reflected their cerebral activity.

2.3 Preprocessing EEG signal

EEG signals are typically susceptible to contamination from various sources. In this study, we employed the EEGLAB toolbox v3.8 to preprocess the EEG data before extracting features. The elimination of artifacts was performed automatically using Artifact Subspace Reconstruction bad burst rejection, allowing for a maximum 0.5-second window standard deviation of 5 and a maximum of 5% out-of-bound channels.

High-frequency components within the EEG signal may be contaminated with muscle artifacts or eye movement interference. To address this issue, we applied a low-pass Finite Impulse Response (FIR) filter with a cutoff frequency of 40 Hz for artifact removal. Furthermore, low-frequency components of the EEG signal (below 0.5 Hz) could be influenced by head or body movements. To minimize these artifacts, we implemented a high-pass filter with a cutoff frequency of 1 Hz. Additionally, we utilized a Notch filter to eliminate the 50 Hz frequency component linked to power line noise since filters are not entirely ideal and may result in leakage of adjacent frequency components.

2.4 Feature extraction and selection

In this study, feature extraction was performed utilizing the hctsa v1.08 software, which encompasses a comprehensive set of 1080 master operations capable of extracting a total of 7752 distinct features. These features originate from various scientific disciplines, including neuroscience, seismology, physics, economics, and artificially generated simulated data (Fulcher and Jones, 2017; Fulcher, 2013). Notably, the extraction process can be executed using a single EEG channel within the hctsa time series analysis framework. Our previous research demonstrated a significant difference in coherency between the Cz and C3 channels (Shafiee-Kandjani, 2020), prompting us to select the Cz channel for subsequent analysis in feature extraction using hctsa. Furthermore, this investigation employed a dataset consisting of 50,000 EEG samples for analysis within the hctsa software.

2.5 Classification methods

Data normalization can substantially enhance the accuracy of specific ML models, and certain models may exhibit suboptimal performance in the absence of normalized data. In this study, the top 40 features extracted from the hctsa software were subjected to normalization prior to the classification analysis. The Min-Max Scaler (MMS) was selected as the normalization technique for our research, as it effectively rescales the data to a range of [0, 1]. This process can be mathematically represented by the following equation:

$$X\_normalized = (X - X\_min) / (X\_max - X\_min)$$

To evaluate the proficiency of feature extraction through hctsa in classifying data, we employed three ML methodologies: Support Vector Machine (SVM), Logistic Regression (LR), and Random Forest (RF). The SVM is a versatile supervised classification algorithm that accommodates various kernel types, such as linear, polynomial, and Radial Basis Function (RBF) kernels (Alvar et al., 2017). SVM has proven effective in high-dimensional spaces, especially when the number of dimensions exceeds the number of subjects. In our study, we utilized a Linear Support Vector Classifier (SVC), fine-tuned with a regularization parameter (C) set to 1, a tolerance threshold for stopping criteria of 1e-4, and a linear kernel applied. This configuration highlights SVM's capability in handling complex high-dimensional data landscapes and facilitating robust and precise categorization between healthy individuals and MA-dependence patients.

LR provides insights into the relationship between predictor variables and binary outcomes by fitting a logistic or sigmoid function to the given data. This function maps predictor values onto a continuum between 0 and 1, representing the likelihood of event occurrence. As a classification algorithm, LR is widely used due to its straightforward implementation, ease of interpretation, and ability to accommodate both categorical and continuous predictor variables (Hosmer Jr et al., 2013). In our study, we employed the Limited-memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) method as the optimization algorithm for LR. The model was augmented with an L2 regularization penalty, where the regularization parameter, C, was set to 1.0. The algorithm was set to cease iterations upon reaching a tolerance threshold of 1e-4, and the maximum number of iterations was capped at 100.

The RF classifier is an ensemble learning method that combines multiple decision tree (DT) classifiers. Each classifier is generated using a distinct random vector independently sampled from the input vector (Breiman, 2001). Every individual tree contributes a unit vote towards the classification of an input vector, favoring the class with the highest frequency of votes. In our RF application, we configured the ensemble to consist of 100 individual decision trees. The maximum depth parameter was left undefined, allowing nodes to expand without limitation until each leaf contained fewer samples than the specified minimum split threshold. Additionally, the minimum number of samples necessary for a leaf node was set to one.

2.6 Validation of classification models

The design of a classifier requires a comprehensive and impartial evaluation of its performance across a spectrum of features and classifier designs, with suitable coefficient values that correspond to a large cohort of subjects. To ensure a fair and robust evaluation, the K-fold cross-validation technique was employed, whereby the dataset was partitioned into 9 equal segments, each containing an equal number of study participants. In each iteration, 8 segments were used as the training subset, while the remaining segment was used as the test subset. This iterative process allows for an accurate assessment of the classifier's performance and convergence in scenarios where data points are limited or when features are used for both training and testing the model.

3.1 hctsa features

The results of analyzing data with hctsa indicated that 6070 features, out of the 7752, produced real values for all of the subjects, and the remaining certain features produced special values (fatal errors or non-real values). We employed the TS_Normalize function in hctsa to eliminate the features that produced special values in over 21% of the time series for each dataset. On average, 1682 features were eliminated from each subject. The range of values generated by the remaining hctsa features varied across features. Although, The TS_Normalize function includes a normalization method to transform the extracted data into a range between 0 and 1. To better visualize similar features, we used the TS_Cluster function implemented inside the hctsa to reorder rows and columns of extracted results. The results of clustered features are shown in Fig. 1; This is a data matrix consisting of rows of different time series (EEG data for each individual) and columns of extracted features from hctsa that were normalized and reordered.

Not all the features extracted from hctsa are useful for classification purposes. The most useful features can be extracted by the TS_TopFeatures function in hctsa. The top 40 features used in our study was shown in Fig. 2. The scatter distribution of each feature was shown in Fig. 2, (a) with blue color for the NC group and with red color for the MDI group. Additionally, We illustrated the median and inter-quarter range for each normalized 40 top feature for NC and MDI groups in Fig. 2, (b).

3.2 Classifying the data

Classifying the data with SVM, LR, and RF indicated that using hctsa for feature extraction resulted in high distinguishing power between the groups of study. we used Python code to evaluate the accuracy of using different combinations of the top 40 features in SVM, LR, and RF classifiers. The results of our inquiry show that there was not any state in which a single feature can lead to an accuracy of 100 percent with SVM, LR, and RF classifiers. However, there were two states with combinations of two features in which using the SVM classifier can lead to an accuracy of 100 percent. Furthermore, using three combinations of features led to the 134, 89, and 100 different states in which the accuracy is 100 percent in SVM, LR, and RF classifiers respectively. Using more combinations of features results in a higher number of states in which the accuracy is 100 percent. We reported the number of states in which the accuracy of using various combinations for each classifier is 100 percent in Table 1. The computational cost for evaluating the accuracy for all combinations is very high, so we computed the results for the combinations of selecting 1, 2, 3, 4, and all 40 features. Also, selecting all top features for classification resulted in 100 accuracy with just the LR model. It may be worth noticing that the total number of combinations of selecting 1, 2, 3, 4, and 40 features are 40, 780, 9880, 91390, and 1 respectively.

Table 1

The number of ways that reach the accuracy of 100 with various combinations of 40 top features. SVM, Support Vector Machine; LR, Logistic Regression; RF, Random Forest
		Number of utilized top features
		1	2	3	4	40
Classifying method	SVM	0	2	134	2,933	0
	LR	0	0	89	3,109	1
	RF	0	0	100	589	0

To the best of our knowledge, this study represents the first attempt to apply the hctsa method for classifying resting-state EEG data into healthy and MDI groups. The objective is to evaluate the efficacy of hctsa and fundamental ML classification approaches in distinguishing between healthy individuals and those with MA dependence. Three distinct ML methods were assessed, specifically SVM, random forests RF, and LR. The results demonstrate that multiple feature combinations achieved 100% accuracy across all three methods.

Overfitting is a common challenge in ML, occurring when a model becomes overly adapted to the peculiarities of the training dataset, thus failing to recognize a generalized predictive pattern (Dietterich, 1995). To address overfitting, this study employed stratified k-fold cross-validation for accuracy assessment, ensuring that the testing dataset remained independent from the training dataset.

Numerous studies have focused primarily on individuals with addiction and implemented knowledge-based systems using EEG signals to aid human decision-making in forecasting substance dependence and generating insights for developing treatment strategies targeting substance abuse. Table 2 presents recent studies that aimed to identify MA abuse by incorporating EEG methodology into their research framework.

Table 2

Comparing the result of using different methods of classifying between healthy and MA addicted individuals. EPNN, Enhanced Probabilistic Neural Network; SVM, Support Vector Machine; RF, Random Forest; LR, Logistic Regression; DT, Decision Tree; MLP, Multilayer Perceptron; RBFN, Radial Basis Function Networks; AB, Ada Boost; GB, Gradian Boost
Authors	feature extraction method	Classification method	Accuracy	reference
Ahmadlou et al. (2013)	visibility graph similarity in the gamma band	EPNN	83%	(Ahmadlou et al., 2013)
Shahmohammadi et al. (2016)	area of positive sections below the time window	subconscious craving-based algorithm	80%	(Shahmohammadi, 2016)
Khajehpour et al. (2019)	clustering coefficient, Node strength, Pairwise Weighted phase lag index	SVM	93%	(Khajehpour, 2019)
Ding et al. (2020)	power spectrum	RF	89%	(Ding et al., 2020)
		SVM	90%
		LR	90%
Chen et al. (2022)	spectral densities before and after virtual reality	SVM	36%	(Li et al., 2022)
		LR	64%
		DT	84%
		RF	68%
		MLP	72%
		RBFN	72%
		AB	68%
		GB	76%

Ahmadlou et al. pioneered the application of resting-state electroencephalography (rEEG) in distinguishing MA abusers from non-drug-using controls. Their research involved assessing individuals with a history of MA abuse, revealing significant disruptions in functional connectivity within the gamma band. These findings provided valuable insights into the neurophysiological consequences of MA abuse (Ahmadlou, 2013). Shahmohammadi et al. conducted a study employing event-related potentials (ERP) to differentiate between MA abusers and control subjects. The researchers utilized the area under the curve (AUC) of windowed ERPs, elicited by a visual paradigm comprised of both drug-related and neutral imagery. This novel approach proved to be effective, as it facilitated the differentiation between MDI and NC with an accuracy rate of 80% (Shahmohammadi, 2016).

In a more recent study conducted by Ding et al., the authors developed RF, SVM, and LR models based on power spectrum density. These models demonstrated superior accuracy, achieving rates of 89%, 90%, and 90% respectively (Ding, 2020). A further developed SVM model, augmented by an array of features such as the clustering coefficient, node strength, and the Pairwise Weighted Phase Lag Index (WPLI), demonstrated a superior predictive accuracy, quantifiably reaching 93% (Khajehpour, 2019). Chen et al. employed virtual reality as a method of differentiation, utilizing eight distinct machine-learning techniques. Nevertheless, it was the DT algorithm that produced the most accurate results, boasting a commendable accuracy rate of 84% (Li, 2022).

ML methodologies are also employed in the medical field to differentiate between healthy individuals and other kinds of addiction such as alcoholism. Prior research utilizing SVM and LR classifiers has demonstrated a classification accuracy ranging from 85 to 95 percent (Bae et al., 2017; Mumtaz et al., 2018; Mumtaz et al., 2017). Farsi et al. achieved a substantial accuracy of 93 percent utilizing a Long Short-Term Memory (LSTM) computational model (Farsi et al., 2020). The highest accuracy in distinguishing between healthy individuals and those with alcoholism is attributed to two recent studies; Salankar et al. (Salankar et al., 2022) employed four distinct classification methodologies, utilizing three features derived from Second Order Difference Plots. Remarkably, they achieved an accuracy rate of up to 97% when implementing RF and SVM classifiers. However, the accuracy escalated to an impressive 99% when they utilized the Multilayer Perceptron (MLP) classifier. Li et al. (Li and Wu, 2022) employed advanced deep learning techniques, achieving varying levels of accuracy with four different models: Convolutional Neural Network (CNN), LSTM, Bidirectional Long Short-Term Memory (Bi-LSTM), and a combined CNN + Bi-LSTM model. The respective accuracies for these models were reported as 96%, 87%, 91%, and 97%. Moreover, the authors successfully refined their model, resulting in a significant increase in performance with an impressive accuracy rate of 99%.

The present study acknowledges several limitations that warrant consideration. Firstly, the recruitment of solely male substance users significantly restricts the generalizability of the research findings. Consequently, future research should be multi-centric and incorporate both genders to adequately address this limitation. The number of participants in our study is also limited which must be considered in future studies. So, Further investigation using different samples is necessary to evaluate the generalizability of these trained models.

The second limitation pertains to the study's inability to account for instances of polydrug consumption. It is plausible that individuals engaged in polydrug use may exhibit unique EEG patterns, differentiating them from those exclusively dependent on MA. Therefore, an inclusive approach that considers polydrug users is essential for a more comprehensive understanding of the relationship between EEG patterns and substance abuse in future studies. Additionally, subsequent studies should aim to conduct a comparative analysis of EEG patterns between MDI and those dependent on other substances, such as cocaine, to determine whether the identified EEG patterns are specific to MA use.

Thirdly, this study primarily aimed to explore the feasibility of utilizing ML methodologies to differentiate between individuals who consume MA and healthy individuals based on EEG data. To enhance the applicability of these findings as markers for detecting MA cravings, it is necessary to measure craving levels and other variables such as treatment adherence. Further research is therefore warranted. Additionally, it is possible that after a year-long period of sobriety and treatment, the MA-using group could potentially resemble the healthy control group. To obtain a more nuanced understanding of this possibility, future research should involve diverse sample groups, including untreated MA users, for instance.

In conclusion, this study represents a pioneering attempt to apply the hctsa method combined with ML classification approaches to distinguish between healthy individuals and those with MA dependence using resting-state EEG data. The results demonstrate promising efficacy, with multiple feature combinations achieving 100% accuracy across three ML methods. The study addressed the challenge of overfitting through stratified k-fold cross-validation, ensuring independent testing datasets. Previous studies in the field have also utilized EEG methodologies to identify MA abuse, providing valuable insights into the neurophysiological consequences of substance dependence. However, there are limitations to consider, including the need for multi-centric studies with gender diversity, the inclusion of polydrug users, and comparative analyses with other substance dependencies. Future research should explore the potential of EEG patterns as markers for MA cravings, considering variables like treatment adherence and craving levels, and involve diverse sample groups to further enhance understanding.

Conflict of interest The authors state no financial or non-financial conflict of interest.

Funding The EEG data utilized in this study was collected with the generous assistance of the Research Center of Psychiatry and Behavioral Sciences, located in Tabriz, Iran.

Ethical approval The study received approval from a Regional Ethics Committee associated with Tabriz University of Medical Sciences with the assigned code TBZMED.REC.1394.26. The patients were treated properly and no intervention was forced upon them.

Acknowledgements We would like to express our gratitude to OpenAI for their GPT-4 model, which provided invaluable assistance in the writing and debugging of the codes used in this project, as well as in the editing process of this manuscript.

Author contributions Reza Meynaghizadeh-Zargar contributed to analyzing the data and writing the manuscript. Sareh Kazmi and Saeed Sadigh-Eteghad contributed to writing and editing the manuscript. Ali Reza Shafiee-Kandjani and Abdollah Barati contributed to recording the EEG data.

Ahmadlou, M., Ahmadi, K., Rezazade, M., & Azad-Marzabadi, E. (2013). Global organization of functional brain connectivity in methamphetamine abusers. Clinical Neurophysiology, 124(6), 1122–1131. https://doi.org/10.1016/j.clinph.2012.12.003.
Alvar, A. A., Deevband, M. R., & Ashtiyani, M. (2017). Neutron spectrum unfolding using radial basis function neural networks. Applied Radiation and Isotopes, 129, 35–41. https://doi.org/10.1016/j.apradiso.2017.07.048.
Bae, Y., Yoo, B. W., Lee, J. C., & Kim, H. C. (2017). Automated network analysis to measure brain effective connectivity estimated from EEG data of patients with alcoholism. Physiological Measurement, 38(5), 759. https://doi.org/10.1088/1361-6579/aa6b4c.
Breiman, L. (2001). Random forests. Machine learning, 45, 5–32. https://doi.org/10.1023/A:1010933404324.
Chen, T., Su, H., Zhong, N., Tan, H., Li, X., Meng, Y., Duan, C., Zhang, C., Bao, J., & Xu, D. (2020). Disrupted brain network dynamics and cognitive functions in methamphetamine use disorder: insights from EEG microstates. Bmc Psychiatry, 20(1), 1–11. https://doi.org/10.1186/s12888-020-02743-5.
Decat, N., Walter, J., Koh, Z. H., Sribanditmongkol, P., Fulcher, B. D., Windt, J. M., Andrillon, T., & Tsuchiya, N. (2022). Beyond traditional sleep scoring: Massive feature extraction and data-driven clustering of sleep time series. Sleep Medicine, 98, 39–52. https://doi.org/10.1016/j.sleep.2022.06.013.
Dietterich, T. (1995). Overfitting and undercomputing in machine learning. ACM computing surveys (CSUR), 27(3), 326–327. https://doi.org/10.1145/212094.212114.
Ding, X., Li, Y., Li, D., Li, L., & Liu, X. (2020). Using machine-learning approach to distinguish patients with methamphetamine dependence from healthy subjects in a virtual reality environment. Brain and Behavior, 10(11), e01814. https://doi.org/10.1002/brb3.1814/v2/review1.
Fang, Y., Rong, R., & Huang, J. (2021). Hierarchical fusion of visual and physiological signals for emotion recognition. Multidimensional Systems and Signal Processing, 32(4), 1103–1121. https://doi.org/10.1007/s11045-021-00774-z.
Farsi, L., Siuly, S., Kabir, E., & Wang, H. (2020). Classification of alcoholic EEG signals using a deep learning method. IEEE Sensors Journal, 21(3), 3552–3560. https://doi.org/10.1109/jsen.2020.3026830.
Fulcher, B. D., & Jones, N. S. (2017). hctsa: A computational framework for automated time-series phenotyping using massive feature extraction. Cell systems, 5(5), 527–531. https://doi.org/10.1016/j.cels.2017.10.001.
Fulcher, B. D., Little, M. A., & Jones, N. S. (2013). Highly comparative time-series analysis: the empirical structure of time series and their methods. Journal of the Royal Society Interface, 10(83), 20130048.
Gemein, L. A., Schirrmeister, R. T., Chrabąszcz, P., Wilson, D., Boedecker, J., Schulze-Bonhage, A., Hutter, F., & Ball, T. (2020). Machine-learning-based diagnostics of EEG pathology. Neuroimage, 220, 117021. https://doi.org/10.1016/j.neuroimage.2020.117021.
Hosmer, D. W. Jr., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression. John Wiley & Sons.
Jiang, T., Zhu, J., Hu, D., Gao, W., Gao, F., & Cao, J. (2022). Early seizure detection in childhood focal epilepsy with electroencephalogram feature fusion on deep autoencoder learning and channel correlations. Multidimensional Systems and Signal Processing, 33(4), 1273–1293. https://doi.org/10.1007/s11045-022-00839-7.
Khajehpour, H., Mohagheghian, F., Ekhtiari, H., Makkiabadi, B., Jafari, A. H., Eqlimi, E., & Harirchian, M. H. (2019). Computer-aided classifying and characterizing of methamphetamine use disorder using resting-state EEG. Cognitive Neurodynamics, 13, 519–530. https://doi.org/10.1007/s11571-019-09550-z.
Khatun, S., Morshed, B. I., & Bidelman, G. M. (2019). A single-channel EEG-based approach to detect mild cognitive impairment via speech-evoked brain responses. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 27(5), 1063–1070. https://doi.org/10.1109/tnsre.2019.2911970.
Kuatsjah, E., Zhang, X., Khoshnam, M., & Menon, C. (2019). Two-channel in-ear EEG system for detection of visuomotor tracking state: A preliminary study. Medical Engineering and Physics, 68, 25–34. https://doi.org/10.1016/j.medengphy.2019.03.016.
Li, H., & Wu, L. (2022). EEG Classification of Normal and Alcoholic by Deep Learning. Brain Sciences, 12(6), 778. https://doi.org/10.3390/brainsci12060778.
Li, Y., Luo, J., & Zhang, J. (2022). Classification of Alzheimer’s disease in MRI images using knowledge distillation framework: an investigation. International Journal of Computer Assisted Radiology and Surgery, 17(7), 1235–1243. https://doi.org/10.1007/s11548-022-02661-9.
Menaka, D., & Vaidyanathan, S. G. (2022). Chromenet: a CNN architecture with comparison of optimizers for classification of human chromosome images. Multidimensional Systems and Signal Processing, 33(3), 747–768. https://doi.org/10.1007/s11045-022-00819-x.
Mumtaz, W., Kamel, N., Ali, S. S. A., & Malik, A. S. (2018). An EEG-based functional connectivity measure for automatic detection of alcohol use disorder. Artificial Intelligence in Medicine, 84, 79–89. https://doi.org/10.1016/j.artmed.2017.11.002.
Mumtaz, W., Vuong, P. L., Xia, L., Malik, A. S., & Rashid, R. B. A. (2017). An EEG-based machine learning method to screen alcohol use disorder. Cognitive Neurodynamics, 11, 161–171. https://doi.org/10.1007/s11571-016-9416-y.
Salankar, N., Qaisar, S. M., Pławiak, P., Tadeusiewicz, R., & Hammad, M. (2022). EEG based alcoholism detection by oscillatory modes decomposition second order difference plots and machine learning. Biocybernetics and Biomedical Engineering, 42(1), 173–186. https://doi.org/10.1016/j.bbe.2021.12.009.
Samraj, D., Ramasamy, K., & Krishnasamy, B. (2023). Enhancement and diagnosis of breast cancer in mammography images using histogram equalization and genetic algorithm. Multidimensional Systems and Signal Processing. https://doi.org/10.1007/s11045-023-00880-0.
Shafiee-Kandjani, A. R., Jahan, A., Moghadam-Salimi, M., Fakhari, A., Nazari, M. A., & Sadeghpour, S. (2020). Resting-State Electroencephalographic Coherence in Recently Abstinent Methamphetamine Users. International Journal of High Risk Behaviors and Addiction, 9(4), https://doi.org/10.5812/ijhrba.103606.
Shahmohammadi, F., Golesorkhi, M., Kashani, M. M. R., Sangi, M., Yoonessi, A., & Yoonessi, A. (2016). Neural correlates of craving in methamphetamine abuse. Basic and clinical neuroscience, 7(3), 221. https://doi.org/10.15412/J.BCN.03070307.
Sharma, L. D., Rahul, J., Aggarwal, A., & Bohat, V. K. (2023). An improved cardiac arrhythmia classification using stationary wavelet transform decomposed short duration QRS segment and Bi-LSTM network. Multidimensional Systems and Signal Processing, 34(2), 503–520. https://doi.org/10.1007/s11045-023-00875-x.
Sidey-Gibbons, J. A., & Sidey-Gibbons, C. J. (2019). Machine learning in medicine: a practical introduction. BMC Medical Research Methodology, 19, 1–18. https://doi.org/10.1186/s12874-019-0681-4.
Thakare, A., Anter, A. M., & Abraham, A. (2023). Seizure disorders recognition model from EEG signals using new probabilistic particle swarm optimizer and sequential differential evolution. Multidimensional Systems and Signal Processing, 34(2), 397–421. https://doi.org/10.1007/s11045-023-00870-2.
Yalamanchili, B., Samayamantula, S. K., & Anne, K. R. (2022). Neural network-based blended ensemble learning for speech emotion recognition. Multidimensional Systems and Signal Processing, 33(4), 1323–1348. https://doi.org/10.1007/s11045-022-00845-9.
Zhang, X., Liang, T., Su, C., Qin, S., Li, J., Zeng, D., Cai, Y., Huang, T., & Wu, J. (2023). Deep learn-based computer-assisted transthoracic echocardiography: approach to the diagnosis of cardiac amyloidosis. The International Journal of Cardiovascular Imaging, 39(5), 955–965. https://doi.org/10.1007/s10554-023-02806-0.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Identifying Methamphetamine Users through EEG Analysis: Harnessing hctsa and Machine Learning Approaches

Status:

Version 1

Abstract

Figures

1 Introduction

2 Materials and Methods

2.1 Sampling

2.2 EEG Recording

2.3 Preprocessing EEG signal

2.4 Feature extraction and selection

2.5 Classification methods

2.6 Validation of classification models

3 Results

3.1 hctsa features

3.2 Classifying the data

4 Discussion

Declarations

References

Additional Declarations

Status:

Version 1