Sun, Nov 24, 2024
| فارسی
Volume 25, Issue 2 (summer 2023)
Advances in Cognitive Sciences 2023, 25(2): 119-132 |
Back to browse issues page
Research code: 1401-1-15-22171
Ethics code: IR.IUMS.FMD.REC.1401.451
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Farokhzadi A, Motamedi H, Shahbazi A, Ghadirivasfi M, Nazari M A. Investigating the specific spectral features of obsessive-compulsive disorder in quantitative electroencephalography. Advances in Cognitive Sciences 2023; 25 (2) :119-132
URL: http://icssjournal.ir/article-1-1552-en.html
URL: http://icssjournal.ir/article-1-1552-en.html
Amirreza Farokhzadi1 
, Hadi Motamedi2 , Ali Shahbazi * 3, Mohammad Ghadirivasfi4 , Mohammad Ali Nazari4
, Hadi Motamedi2 , Ali Shahbazi * 3, Mohammad Ghadirivasfi4 , Mohammad Ali Nazari4
1- Student Research Committee, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran/ Department of Neuroscience, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran
2- Student Research Committee, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran/Department of Neuroscience, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran
3- Department of Neuroscience, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran
4- Research Center for Addiction and Risky Behaviors (ReCARB), Medical Faculty, Iran University of Medical Sciences (IUMS), Tehran, Iran
2- Student Research Committee, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran/Department of Neuroscience, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran
3- Department of Neuroscience, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences (IUMS), Tehran, Iran
4- Research Center for Addiction and Risky Behaviors (ReCARB), Medical Faculty, Iran University of Medical Sciences (IUMS), Tehran, Iran
Abstract: (823 Views)
Introduction
Obsessive-compulsive disorder (OCD) is a debilitating mental illness that often manifests in childhood and adolescence. It is among the top ten causes of “life with disability” and has a significant impact on society. The global lifetime prevalence of OCD is 2-3%, with a 12-month prevalence of up to 1%. In Iran, the prevalence of OCD in children and adolescents is estimated at about 10%. Moreover, The COVID pandemic has led to an increase in obsessive symptoms (intrusive thoughts or images, compulsive behaviors, or both) and psychiatric referrals. Despite advances in understanding the neurophysiological processes of OCD, diagnosis remains primarily clinical. Quantitative electroencephalography (QEEG) has become a popular brain mapping technique for examining psychiatric diseases due to its ease of use and accessibility. On the other hand, machine learning methods have been applied to electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) data to improve diagnosis. These techniques have the potential to advance diagnostic methods, treatment, and measurement of response to treatment. In particular, using QEEG spectral features as input for machine learning algorithms such as TREE can provide valuable information about neural activities involved in OCD. This study evaluated algorithm performance using widely recognized spectral features associated with OCD based on past research as input data.
Methods
This cross-sectional analytical study examined a population of 42 individuals, comprising a combination of patients and healthy individuals with at least 40% patient dispersion. Patients were diagnosed based on the Yale-Brown questionnaire and confirmed by a psychiatrist in one visit for evaluation of neurocognitive disorders and entry criteria. Besides, healthy individuals were confirmed in terms of psychiatric issues and entered the study as the healthy group in one visit. The patients were referred to the Brain and Cognition Clinic, and their written consent was obtained upon entering the study.
EEG signals were obtained from patients and healthy individuals in resting state with eyes closed. This method was performed using a 21-channel cap. The cap’s sensors were placed according to the 10-20 electrode placement standard. These signals were sampled at a frequency of 625 Hz and bandpass filtered between 0.5 Hz and 70 Hz.
In the first stage, common noise and artifacts related to facial movements present in the signals were removed using preprocessing methods. For this purpose, the MATLAB Field Trip toolbox was used. Furthermore, the ICA (Independent Component Analysis) algorithm was used for artifact removal using MATLAB EEG Lab toolbox by manual inspection of the components in time and frequency domains. At this stage, an EEG recording was divided into 1-minute segments to produce additional samples and enhance the dataset’s diversity that is needed for the machine learning training.
In this study, based on findings from Prichep et al.’s study (11) and other main references (23), six features, including relative power in the theta band for bipolar central and temporal derivations, relative power in the theta band for monopolar frontal derivations, relative power in the alpha band for bipolar central and temporal derivations and the relative power in the beta band for frontal, and central and occipital derivations were calculated. Their medians were compared between patients and healthy individuals using the Mann-Whitney U test.
A nonlinear machine learning algorithm was used as a classifier, and the model hyperparameters were optimized using a Bayesian optimizer. The model cost was also modified to compensate for data imbalance using relative cost sensitivity. Finally, 10-fold cross-validation was used to evaluate the classification performance of the proposed method. Evaluation metrics used in this study were accuracy (AC), sensitivity (SE), specificity (SP), and false discovery rate (FDR).
Results
This study included a total of 42 individuals, of which 27 were labeled as patients (cases) and 14 were labeled as healthy (controls). Three patients were excluded from the demographic analysis due to needing more consent, demographic information, and outlier removal. Among the 24 obsessive-compulsive patients in the patient group, 15 were female, and nine were male. The youngest individual was a 15-year-old woman, and the oldest was a 56-year-old woman. The average age of the control group was 30 years.
In order to compare the performance of the machine learning algorithm with other common statistical methods for group comparison, feature distributions were examined and compared between the two groups. The normality of the distribution of the extracted features was assessed using the Kolmogorov-Smirnov test. The results indicated that all of the features under examination exhibited a non-normal distribution, with p-values for all six features being less than 0.001.
After removing outliers for comparison between two independent groups with non-normal distribution, the Mann-Whitney (Wilcoxon) non-parametric test was used. The obtained results showed that only the fourth feature was significantly different between the healthy and patient groups (P=0.002), but none of the first (P=0.692), second (P=0.643), third (P=0.406), fifth (P=0.062), sixth (P=0.057) extracted features had significant difference between the two.
A total of 42 EEG signal recordings obtained for a duration of 4 minutes each were divided into 1-minute sections with similar labels, and eventually, the number of data reached 168 cases. Indicatively, 1-min slicing led to better results for the proposed method compared with other slicing times.
The proposed method achieved a sensitivity of 83.3% and a specificity of 80.0%. An accuracy of 82.1% was obtained with cross-validation of 10 folds using the algorithm. The false discovery rate (FDR) was found to be 11.8%.
Conclusion
Quantitative electroencephalography (QEEG) is a non-invasive imaging technique that has been widely used to diagnose and evaluate treatment responses in patients with OCD using spectral findings. This study aimed to evaluate the performance of the machine learning algorithm in differentiating patients with OCD from healthy individuals based on these spectral parameters. The performance of this algorithm was compared with common statistical methods (Wilcoxon and Mann-Whitney). The results showed that the nonlinear machine learning algorithm achieved an accuracy of 82.1%, sensitivity of 83.3%, specificity of 80.0%, and false discovery rate of 11.8% in differentiating between the healthy and patient groups.
Some suggestions for performance improvement of the machine learning algorithm in this study are presented. One possible improvement could be achieved by artifact removal from the prefrontal cortex using the ICA method, removing features such as eye blinking or muscle activity. Another improvement could be achieved by taking into account the generalizability factors between studies, such as differences in study populations, cultural, religious, and linguistic differences, and even differences in data recording methods. Lastly, keynotes such as the small sample size, the 10-fold cross-validation performed, and the hyperparameters of the selected machine learning algorithm are possible areas for improvement in this study. In addition, they could bring about a plausible performance of spectral features and their use in differentiating between patients with OCD and healthy individuals in the guidelines.
Ethical Considerations
Compliance with ethical guidelines
The studies involving human participants were reviewed and approved by the Ethics Committee of Iran University of Medical Sciences (IR.IUMS.FMD.REC.1401.451). Ethical considerations for participants in this study included obtaining written consent and respecting the principle of confidentiality. The principle of confidentiality was respected by coding and removing names from questionnaires to protect participants’ privacy. Additionally, all participants were provided with sufficient information about the research procedures to enable them to make an informed decision about their participation.
Authors’ contributions
The study was designed and prepared by all five authors; data collection, data analysis, and preparation of the text of the article were conducted by the first two authors; the third author was responsible for the supervision and final review of the work as the corresponding author; the fourth and fifth authors were responsible for technical and scientific assistance during the data acquisition and analysis respectively.
Funding
This research was conducted with the funding support of the Cognitive Sciences and Technologies Council (tracking code 10976) and Iran University of Medical Sciences (research code: 22171)
Acknowledgments
The authors want to thank the Brain and Cognition Clinic for providing the types of equipment and Doctor Mohammadreza Shalbafan for case evaluation and referral.
Conflicts of interest
The authors declared no potential conflict of interest.
Obsessive-compulsive disorder (OCD) is a debilitating mental illness that often manifests in childhood and adolescence. It is among the top ten causes of “life with disability” and has a significant impact on society. The global lifetime prevalence of OCD is 2-3%, with a 12-month prevalence of up to 1%. In Iran, the prevalence of OCD in children and adolescents is estimated at about 10%. Moreover, The COVID pandemic has led to an increase in obsessive symptoms (intrusive thoughts or images, compulsive behaviors, or both) and psychiatric referrals. Despite advances in understanding the neurophysiological processes of OCD, diagnosis remains primarily clinical. Quantitative electroencephalography (QEEG) has become a popular brain mapping technique for examining psychiatric diseases due to its ease of use and accessibility. On the other hand, machine learning methods have been applied to electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) data to improve diagnosis. These techniques have the potential to advance diagnostic methods, treatment, and measurement of response to treatment. In particular, using QEEG spectral features as input for machine learning algorithms such as TREE can provide valuable information about neural activities involved in OCD. This study evaluated algorithm performance using widely recognized spectral features associated with OCD based on past research as input data.
Methods
This cross-sectional analytical study examined a population of 42 individuals, comprising a combination of patients and healthy individuals with at least 40% patient dispersion. Patients were diagnosed based on the Yale-Brown questionnaire and confirmed by a psychiatrist in one visit for evaluation of neurocognitive disorders and entry criteria. Besides, healthy individuals were confirmed in terms of psychiatric issues and entered the study as the healthy group in one visit. The patients were referred to the Brain and Cognition Clinic, and their written consent was obtained upon entering the study.
EEG signals were obtained from patients and healthy individuals in resting state with eyes closed. This method was performed using a 21-channel cap. The cap’s sensors were placed according to the 10-20 electrode placement standard. These signals were sampled at a frequency of 625 Hz and bandpass filtered between 0.5 Hz and 70 Hz.
In the first stage, common noise and artifacts related to facial movements present in the signals were removed using preprocessing methods. For this purpose, the MATLAB Field Trip toolbox was used. Furthermore, the ICA (Independent Component Analysis) algorithm was used for artifact removal using MATLAB EEG Lab toolbox by manual inspection of the components in time and frequency domains. At this stage, an EEG recording was divided into 1-minute segments to produce additional samples and enhance the dataset’s diversity that is needed for the machine learning training.
In this study, based on findings from Prichep et al.’s study (11) and other main references (23), six features, including relative power in the theta band for bipolar central and temporal derivations, relative power in the theta band for monopolar frontal derivations, relative power in the alpha band for bipolar central and temporal derivations and the relative power in the beta band for frontal, and central and occipital derivations were calculated. Their medians were compared between patients and healthy individuals using the Mann-Whitney U test.
A nonlinear machine learning algorithm was used as a classifier, and the model hyperparameters were optimized using a Bayesian optimizer. The model cost was also modified to compensate for data imbalance using relative cost sensitivity. Finally, 10-fold cross-validation was used to evaluate the classification performance of the proposed method. Evaluation metrics used in this study were accuracy (AC), sensitivity (SE), specificity (SP), and false discovery rate (FDR).
Results
This study included a total of 42 individuals, of which 27 were labeled as patients (cases) and 14 were labeled as healthy (controls). Three patients were excluded from the demographic analysis due to needing more consent, demographic information, and outlier removal. Among the 24 obsessive-compulsive patients in the patient group, 15 were female, and nine were male. The youngest individual was a 15-year-old woman, and the oldest was a 56-year-old woman. The average age of the control group was 30 years.
In order to compare the performance of the machine learning algorithm with other common statistical methods for group comparison, feature distributions were examined and compared between the two groups. The normality of the distribution of the extracted features was assessed using the Kolmogorov-Smirnov test. The results indicated that all of the features under examination exhibited a non-normal distribution, with p-values for all six features being less than 0.001.
After removing outliers for comparison between two independent groups with non-normal distribution, the Mann-Whitney (Wilcoxon) non-parametric test was used. The obtained results showed that only the fourth feature was significantly different between the healthy and patient groups (P=0.002), but none of the first (P=0.692), second (P=0.643), third (P=0.406), fifth (P=0.062), sixth (P=0.057) extracted features had significant difference between the two.
A total of 42 EEG signal recordings obtained for a duration of 4 minutes each were divided into 1-minute sections with similar labels, and eventually, the number of data reached 168 cases. Indicatively, 1-min slicing led to better results for the proposed method compared with other slicing times.
The proposed method achieved a sensitivity of 83.3% and a specificity of 80.0%. An accuracy of 82.1% was obtained with cross-validation of 10 folds using the algorithm. The false discovery rate (FDR) was found to be 11.8%.
Conclusion
Quantitative electroencephalography (QEEG) is a non-invasive imaging technique that has been widely used to diagnose and evaluate treatment responses in patients with OCD using spectral findings. This study aimed to evaluate the performance of the machine learning algorithm in differentiating patients with OCD from healthy individuals based on these spectral parameters. The performance of this algorithm was compared with common statistical methods (Wilcoxon and Mann-Whitney). The results showed that the nonlinear machine learning algorithm achieved an accuracy of 82.1%, sensitivity of 83.3%, specificity of 80.0%, and false discovery rate of 11.8% in differentiating between the healthy and patient groups.
Some suggestions for performance improvement of the machine learning algorithm in this study are presented. One possible improvement could be achieved by artifact removal from the prefrontal cortex using the ICA method, removing features such as eye blinking or muscle activity. Another improvement could be achieved by taking into account the generalizability factors between studies, such as differences in study populations, cultural, religious, and linguistic differences, and even differences in data recording methods. Lastly, keynotes such as the small sample size, the 10-fold cross-validation performed, and the hyperparameters of the selected machine learning algorithm are possible areas for improvement in this study. In addition, they could bring about a plausible performance of spectral features and their use in differentiating between patients with OCD and healthy individuals in the guidelines.
Ethical Considerations
Compliance with ethical guidelines
The studies involving human participants were reviewed and approved by the Ethics Committee of Iran University of Medical Sciences (IR.IUMS.FMD.REC.1401.451). Ethical considerations for participants in this study included obtaining written consent and respecting the principle of confidentiality. The principle of confidentiality was respected by coding and removing names from questionnaires to protect participants’ privacy. Additionally, all participants were provided with sufficient information about the research procedures to enable them to make an informed decision about their participation.
Authors’ contributions
The study was designed and prepared by all five authors; data collection, data analysis, and preparation of the text of the article were conducted by the first two authors; the third author was responsible for the supervision and final review of the work as the corresponding author; the fourth and fifth authors were responsible for technical and scientific assistance during the data acquisition and analysis respectively.
Funding
This research was conducted with the funding support of the Cognitive Sciences and Technologies Council (tracking code 10976) and Iran University of Medical Sciences (research code: 22171)
Acknowledgments
The authors want to thank the Brain and Cognition Clinic for providing the types of equipment and Doctor Mohammadreza Shalbafan for case evaluation and referral.
Conflicts of interest
The authors declared no potential conflict of interest.
Type of Study: Research |
Received: 2023/05/23 | Accepted: 2023/08/15 | Published: 2023/09/20
Received: 2023/05/23 | Accepted: 2023/08/15 | Published: 2023/09/20
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
