Ensemble classifiers fed by functional connectivity during cognitive processing differentiate Parkinson’s disease even being under medication

Brain–computer interface technologies, as a type of human-computer interaction, provide a control ability on machines and intelligent systems via human brain functions without needing physical contact. Moreover, it has a considerable contribution to the detection of cognitive state changes, which gives a clue for neurodegenerative diseases, including Parkinson’s disease (PD), in recent years. Although various studies implemented different machine learning models with several EEG features to detect PD and receive remarkable performances, there is a lack of knowledge on how brain connectivity during a cognitive task contributes to the differentiation of PD, even being under medication. To fill this gap, this study used three ensemble classifiers, which were fed by functional connectivity through cognitive response coherence (CRC) with varying selected features in different frequency bands upon application of the 3-Stimulation auditory oddball paradigm to differentiate PD medication ON and OFF and healthy controls (HC). The results revealed that the most remarkable performances were exhibited in slow frequency bands (delta and theta) in comparison to high frequency and wide range bands, especially in terms of target sounds. Moreover, in the delta band, target CRC distinguishes all groups from each other with accuracy rates of 80% for HC vs PD-OFF, 80% for HC vs PD-ON, and 81% for PD-ON vs PD-OFF. In the theta band, again target sounds were the most distinctive stimuli to classify HCxPD-OFF (80% accuracy), HCxPD-ON (80.5% accuracy) with quite good performances, and PD-ONxPD-OFF (76% accuracy) with acceptable performance. Besides, this study achieved a state-of-the-art performance with an accuracy of 87.5% in classifying PD-ONxPD-OFF via CRC of standard sounds in the delta band. Overall, the findings revealed that brain connectivity contributes to identifying PD and HC as well as the medication state of PD, especially in the slow frequency bands.


Introduction
In the last decades, intelligent systems have become an indispensable part of our lives and humans interact with machines, more simply computers, for various purposes.Consequently, human-computer interaction (HCI), which is a field of study focused on designing and implementing enhanced user interfaces between the machine and the end user, started to play a crucial role in daily life.One of the main components of HCI is users (i.e.humans), so, not only physical contact but also mental communication could provide a way for interacting with computing systems (Chalmers 2003).As a part of HCI, brain-computer interface (BCI) technology based on neuroimaging techniques provides a control ability on machines via human brain functions without needing physical contact (Alimardani and Hiraki 2020).In recent years, BCI is also used for the detection of cognitive state changes and even for treatment, especially in neuropsychiatric diseases (Tromp 2016, Tülay et al 2019).
The functionality of a human brain is encoded with connectivity between different brain regions as well as cortical neural activities in those regions.In addition, they are believed to underpin multiple cognitive functions including attention, perception, and memory (Petsche and Etlinger 1998) which are fundamental abilities to lead a healthy life.Therefore, they are mostly preferred in HCI and/or BCI systems as features.Although the studies in the literature usually focus on the activities of specific regions, brain connectivity has also a prominent contribution to understanding the brain's working mechanism (Miljevic et al 2022, Chiarion et al 2023).Functional connectivity (FC) which is the one type of brain connectivity is defined as statistical dependencies among spatially distant brain locations (Friston 2011).Among various FC measures, coherence is one of the most used measurements and provides information about the variability of time differences between two brain signals in the frequency domain (Thatcher 2012, Bastos andSchoffelen 2016; see section 2.2.2).Moreover, according to Fries (2005), neural communication is mechanistically subserved by neural coherence.
Electroencephalography (EEG) is the most preferred neuroimaging technique to measure FC in recent years due to having a high temporal resolution (Cao et al 2021) and significantly improved source localization (He et al 2019).Although FC is mostly investigated during the resting state (van Diessen et al 2015), according to Bas ¸ar et al (2010), there are remarkable differences between 'EEG coherence' in spontaneous (resting state) activity and 'Cognitive Response Coherence (CRC)' in the cognitive processing of incoming stimuli.CRC plays an important role in the human brain and even is accepted as a potential biomarker in various neuropsychiatric diseases (Bas ¸ar et al 2016).In a series of studies, the alterations in CRCs were shown in Alzheimer's disease (AD) (Güntekin et al 2008, Bas ¸ar et al 2010), bipolar patients (Özerdem et al 2010, 2011), and schizophrenia (Bucci et al 2007, Prieto Alcántara et al 2023) upon application of different cognitive tasks.Besides these diseases, several studies have contributed also to the investigation of biomarkers of both cognitively normal and impaired Parkinson's disease (PD) by means of CRC elicited by several cognitive (Cassidy and Brown 2001), and emotional (Yuvaraj et al 2015) tasks.
PD is the second most prevalent neurodegenerative disease worldwide after AD in the elderly population (Bloem et al 2021).PD is primarily characterized by motor symptoms (e.g.hypokinesia, resting tremors, mask face), and various clinical tests are used for diagnosing PD (Rizzo et al 2016).According to Virameteekul et al (2023), significant enhancement in clinical diagnostic accuracy for PD has been achieved in clinical practice over the last decade, even at the early stages of the disease.The study reported that the International Parkinson and Movement Disorder Society clinically established early PD criteria which are highly specific allowing the identification of de novo PD individuals with a high level of accuracy (98.4%).However, potential limitations in documenting clinical information due to variations in assessments by different clinicians could cause inaccurate results (Virameteekul et al 2023).Besides motor symptoms, clinical manifestations of PD also involve non-motor symptoms (e.g.cognitive and sensory deficits including hearing loss, olfactory dysfunction, and sleep problems) (Jafari et al 2020).For example, PD patients with subjective cognitive decline have decreased cognitive capacity without detectable impairment on neuropsychological tests (Huang et al 2023).Consequently, to overcome subjective assessment, especially in the prodromal phase of the PD, using cognitive symptoms with current methodologies of machine learning (ML) and/or deep learning (DL) as an objective tool plays a considerable role in increasing the sensitivity of clinical diagnosis (Alzubaidi et al 2021, Loh et al 2021), even if the symptoms are controlled with medication (Aljalal et al 2022).
Across years of research, various EEG characteristics fed ML/DL models to predict PD with different cognitive stages (Maitín et al 2020).A large majority of these studies used several resting state features for their classification models (Yuvaraj et al 2018, Chaturvedi et al 2017, Oh et al 2020, Anjum et al 2020, Aljalal et al 2022, Chang et al 2022, Yang and Huang 2022, Qiu et al 2022).Moreover, the sleep data of PD individuals were also used to detect different kinds of clinical states of PD (Zhang et al 2022).On the other hand, a few studies investigated the contribution of cognitive task-related features, even though there is vast literature showing cognitive changes in PD (Fang et al 2020, Güntekin et al 2020, Hünerli-Gündüz et al 2023).Event-related potentials (Hassin-Baer et al 2022; also see review Wang et al 2020), Fourier coefficients (Cavanagh et al 2018), theta activity (Singh et al 2018), common spatial patterns (CSPs) (Smrdel 2022), Sparse Dynamical Features (Meghnoudj et al 2023), Event-related spectral perturbation (Lin et al 2023), and inter-trial phase coherence (Tülay et al 2023) are some of the classifier features that were used in this concept.Although most of the previous studies have achieved good performances to differentiate healthy control (HC) and PD via different features during a cognitive task, these features represent just one aspect of the functionality of the brain, which is brain activities in local regions.However, brain functionality is more than just that aspect.Since FC is a part of the brain's working mechanism, it could also contribute to the reliable diagnosis of PD.There is only one study that uses FC features, partial directed coherence (de Oliveira et al 2020), but these features were obtained in response to photic stimulation that does not elicit cognitive functions (e.g.perception, memory, attention, decision-making) of the brain.
To the best of my knowledge, there is a big gap in distinguishing individuals with PD, independent from the medication state via CRC, even via all types of FC measurements, upon application of a cognitive task by using ML approaches.The current study aimed to examine mainly two aspects.First, to differentiate the individuals with PD on both ON and OFF medication states from HC via CRC features.Second, to detect the best CRC feature set among different frequency bands to classify PD.To achieve this goal, multiple ensemble ML algorithms were used (see section 2.2.4).It is hypothesized that both PD patients with and without medication would be discriminated with the delta and theta CRCs in response to target stimuli of the auditory cognitive task.

Dataset
A public dataset (Cavanagh 2021) was adopted (available at https://openneuro.org/datasets/ds003490/ versions/1.1.0),which includes EEG signals of 25 individuals clinically diagnosed with Parkinson's disease (PD group) and 25 matched healthy elderly controls (HC group).After disregarding the four and five participants respectively from PD and HC groups due to persistent artifacts or noise (muscle artifacts and spikes) that caused a low number of epochs (less than 20) in their recordings, the remaining data were included in the study.
The mean age of the PD group with 21 participants (9 female, 12 male) was 70.24 (Standard deviation: 8.5) and the mean age of the HC group with 20 participants (8 female, 12 male) was 68.8 (Standard deviation: 17.5) years.The EEG data of the PD group were recorded twice separated by a week when they were either ON or OFF medication (after a 15-hour overnight withdrawal from their individual prescriptions of dopaminergic medication), whereas only once for the HC group.Hereafter the groups for PD are referred to as PD-ON or PD-OFF, respectively.The inclusion criteria for the participants were as follows; Mini-Mental State Exam (MMSE) score ⩾26.Detailed information about the rest of the neuropsychological and questionnaire assessments could be reached in the study of Cavanagh et al (2018).All participants provided written informed consent.
During EEG recording, the participants underwent a task called the 3-stimulus auditory oddball paradigm to evoke cognitive processes.In the paradigm, there were three stimuli named standard (440 Hz sinusoidal tones, and 80 dB), target (660 Hz sinusoidal tones, and 80 dB), and novel distractors that were naturalistic sounds (Bradley and Lang 1999), varying with each presentation (65 dB with an inter-quartile range of +/− 6.5 dB).The distribution of these three stimuli among the 200 stimuli was as follows: 140 standard, 30 target, and 30 novel sounds.Each stimulus was presented for 200 ms and a random inter-trial interval (ITI) was selected from a uniform distribution of 0.5-1 s for the novel sounds, and 950-1450 ms for both the standard and target sounds.The participants mentally counted the target sounds and ignored standards and novels (Cavanagh et al 2018).
The EEG data of participants were recorded using the 64-channel Brain Vision system with a sampling rate of 500 Hz.During the recording, a CPz served as a reference electrode, and an AFz electrode served as ground.

EEG signal analysis
The offline data analysis was completed mainly in four steps including preprocessing, feature extraction, feature selection, and classification.The preprocessing and feature extraction steps were fulfilled using custom-written scripts via the FieldTrip Toolbox (version no: 20220104) (Oostenveld et al 2011) running under Matlab R2018b (MathWorks, Natick, MA, U.S.A.) whereas the feature selection and classification steps were fulfilled using Python programming language (Python version 3.9) with various modules (see sections 2.2.3 and 2.2.4).

Preprocessing
The current study applied a preprocessing pipeline described by Tülay (2023) which differed from the reference study by Cavanagh et al (2018) to the data.The pipeline has broken down into six sub-steps starting from importing the data carrying on with artifact and eye movement removal, detrending, dividing the data into separate conditions, removing residual bad trials, and equalizing the number of epochs among conditions.
In the importing data step, the data around the stimulus onset (−2000 to 2000 ms) were extracted from the raw data for all types of conditions and were imported to the Matlab platform.Moreover, the number of channels was also reduced to 30 (F7, F5, F3, F1, Fz, F2, F4, F6, F8, T7, C5, C3, C1, Cz, C2, C4, C6, C8, P7, P5, P3, P1, Pz, P2, P4, P6, P8, O1, Oz, O2) from 64 channels.During the importing process, the data were re-referenced to an average reference of selected channels, and each epoch was baseline corrected according to the mean amplitude of −200 to 0 ms pre-stimulus time window.Finally, the line noise was removed from the data via the discrete Fourier transform (DFT) filter.
After importing the data of all participants, the following steps were performed to clean the data from multiple types of artifacts.Firstly, fast muscular artifacts and unexpected jumps were eliminated by visual inspection to enhance the efficiency of the further step that applies independent component analysis (ICA).Then, the fast ICA1 was applied for removing eye movements, heartbeat effects, and spiky patterns as a later process.For the last type of artifact, slow low-frequency drifts, detrending was applied to remove per trial.As a result, the cleaned data was reconstructed.
The clean data that included all the types of stimuli (target, standard, novel) were divided into separate sub-data based on the conditions per participant, and the last bad trials were detected manually and removed from the data.As a last step of preprocessing, the number of epochs among conditions was equalized by removing randomly selected trials.In the end, if a participant had less than 20 trials for a condition, the data were excluded from the dataset.

Feature extraction: coherence analysis
Coherence is a linear measure that determines whether neuronal oscillations at similar frequencies in different brain regions engage in oscillatory coupling with a preferred phase difference (Bastos and Schoffelen 2016).To compute the coherence magnitudes in the frequency domain, the magnitude of the summed cross-spectral density between two brain signals is normalized by their respective power.The mathematical background of the coherence measure could be found in the study of Bastos and Schoffelen (2016).The value of the coherence ranges from 0 to 1. Synchronized signals produce high coherence magnitude that is close to 1, but if the synchronization decreases, the coherence magnitude approaches 0.
FieldTrip Toolbox contains the ft_connectivityanalysis function that computes coherence for each frequency bin.Before running that function, Fourier coefficients were computed between 0.1 and 50 Hz with 0.5 Hz steps by using ft_freqanalysis with 'mtmfft' method and conventional single taper (Hanning) to obtain the cross-spectral density matrix that is essential for ft_connectivityanalysis2 .Fourier coefficients were obtained for each individual trial (1 s after stimulus onset), and each selected channel (F3, Fz, F4, C3, Cz, C4, P3, Pz, P4, O1, Oz, O2, where F, C, P, and O represents Frontal, Central, Parietal, and Occipital brain regions, respectively).

Feature selection
In the current study, there were twelve electrode pairs as features, and the sequential feature selection algorithm (SFS) (Pudil et al 1994) was used to find the most contributing features among them.Therefore, in the application of SFS, there would be a minimum of 1 and a maximum of 11 features.SFS was performed using sklearn.featureselection module in Python programming language (Python version 3.7.13).

Feature classification
During the classification process of PD (ON and OFF) and HC by means of CRC features, multiple settings were implemented by using different classification models, frequency bands, stimulations, and electrode pairs to investigate which models with which features contribute most to classification performance (figure 2).
The three models of ensemble methods (RandomForest, Gradient Tree Boosting, and AdaBoost) were developed using Python programming language (Python version 3.9) in Google Colaboratory (Google Colab in short) environment to assess whether the CRC features could be decoded to predict the three groups, PD-ON, PD-OFF, and HC.Various Scikit-learn (Sklearn) 1.2.2 libraries were used to implement the models.
The Random Forest algorithm is an averaging algorithm based on randomized decision trees.The number of trees in the forest (n_estimators) was 100.The sub-sample size to train each base estimator and best split were controlled with the default setting of max_samples and max_features parameters, respectively, to decrease the variance of the forest estimator.Unlike the reference study of Breiman (2001), the scikit-learn implementation that combines classifiers by averaging their probabilistic prediction was adopted in the study.
Another family of ensemble methods is boosting methods which allow you to reduce bias and combine multiple weak learners (decision stumps) into a single strong and efficient learner.For the current study, two major boosting algorithms, AdaBoost (Adaptive Boosting) and gradient boosting, have been used.During the study, AdaBoost-SAMME.R (Zhu et al 2009) was implemented for the AdaBoost classifier.In addition, the gradient boosting classifier with 100 decision stumps as weak learners (n_estimators) was applied.For both boosting classifiers, the maximum number of estimators at which boosting is terminated was 100 and the learning rate was 0.1.The other hyper-parameters were kept as default.Unlike AdaBoost, which uses an exponential loss function, Gradient Boosting uses an arbitrary differentiable loss function.
A stratified 10-Folds cross-validation technique was used for model evaluation and the results are averaged over all 10 folds.Moreover, the performances of the classifiers were evaluated using multiple metrics, including accuracy, recall (Sensitivity), Specificity, and Receiver operating characteristic (ROC) curve.
There were three groups (HC, PD-ON, PD-OFF) in the dataset.Each classification model was utilized to differentiate pairs of groups (HC x PD-ON, HC x PD-OFF, and PD-ON x PD-OFF) by means of CRCs in different frequency bands (Please see section 2.2.2).
Figure 2 summarizes the methods of the study and provides an overview of the classification analysis steps.

Results
In this study, classification performances were evaluated for three stimuli types and six different frequency bands separately to differentiate group pairs among HC, PD-ON, and PD-OFF via three types of ensemble ML algorithms.According to the results, unsurprisingly, the performances especially in the delta and theta bands are better than high frequencies and the wide-range band except for a few cases that will be explained in further paragraphs.Moreover, similarly, target is the predominant distinguishing condition in delta and theta bands again with a few exceptions (Please see tables 1 and 2).Therefore, the results of CRC for target stimuli in delta and theta bands are going to be shared in this section.However, all the results obtained for high frequencies and overall (wide range) frequencies were presented in supplementary materials 1-4.
After calculating CRC values elicited by different stimuli in the wide-range frequency band for each participant and channel pairs, the grand averages were obtained over all participants.Figure 3 depicts the target CRC values in delta and theta bands for each channel pair used as features of the classifiers and for all groups.According to the figure, the biggest visible difference among the three groups has been seen for fronto-central connectivity.However, the results of classification revealed that the other channel pairs also contribute to the differentiation of all groups (tables 1 and 2).
The performance metrics (Accuracy, Recall, and Specificity) of three ensemble methods that differentiated three group pairs (HC x PD-OFF, HC x PD-ON, PD-ON x PD-OFF) were listed with selected features in tables 1 and 2, respectively for delta and theta bands.Table 1 shows that the CRC in response to target stimuli was quite successful to differentiate all group pairs in the delta band.Random Forest classifier achieved the highest accuracy (80%) to differentiate HC and PD-OFF where the classifier was slightly better at predicting HC (85%) than it was at predicting PD-OFF (76,19%).To classify HC and PD-ON, Adaboost gave the best performance with 80% accuracy by predicting PD-ON (85,71%) better than HC (75%).However, when PD individuals with different medication states wanted to be distinguished, although the target condition was still good enough with the 81% accuracy obtained by Adaboost, the highest accuracy (87,5%) was obtained by the standard condition which fed the Gradient Boosting classifier.Considering all the results, it could be easily said that CRC in response to target stimuli in the delta band contributed to the classification of each group with remarkable performances.
Similarly, the CRC in response to target stimuli in the theta band also increased the performances of classifiers to differentiate PD from HC regardless of whether on medication.However, it was not successful in classifying ON and OFF states as much as the other group pairs.Table 2 contains the classification metrics of all ensemble models for CRC in the theta band.As seen in the table, in the case of target stimuli applied, PD-OFF (80% accuracy) and PD-ON (80,5% accuracy) were distinguished from HC respectively by Adaboost and Gradient Boosting classifiers.Besides, to classify PD-ON and PD-OFF, the maximum accuracy rate of 76% was obtained again with the target CRC that fed the Gradient Boosting classifier.The results revealed that theta CRC in response to target stimuli has a greater contribution than the novel and standard stimuli to detect PD at different medication states.The figures also represent the mean area under the curve (AUC) values and standard deviations.Figure 4(a) depicts the ROC curves and AUCs in each fold for the differentiation of HC and PD-OFF by Random Forest, which gave the highest accuracy rate (80%) for this group pair.Accordingly, the mean AUC value was 0.77, which could be considered acceptable.In the classification process of HC and PD-ON by Adaboost, the mean AUC value was 0.80, which could be considered quite excellent (figure 4(b)).Lastly, in the detection of medication status (PD-ON vs PD-OFF), the AdaBoost classifier gave the best mean AUC value (0.88) as seen in figure 4(c).Similarly, figure 5 depicts the ROC curves and mean AUCs in the theta band resulting from the classification of HC vs PD-OFF (0.82), HC vs PD-ON (0.79), and PD-ON vs PD-OFF (0.80) respectively by Adaboost, Gradient Boosting, and Gradient Boosting classifiers.

Conclusion
The current study performed the classification analyses by using CRCs with varying selected features in different frequency bands upon application of the 3-Stimulation auditory oddball paradigm to understand which features have a higher contribution to identifying PD and HC as well as the medication state of PD.In accordance with this purpose, multiple ensemble classifiers including Random Forest, AdaBoost, and Gradient Boosting were employed with the forward feature selection method to reveal which classifier exhibits good distinction between PD with different medication states and HC.To my knowledge, this is the first study that attempts to use FC through CRC in response to cognitive stimuli as features for classifying PD and HC.
The results of the classification performances revealed that PD could be able to be classified successfully with the contribution of CRC features.The most important observations of classification analyses performed with multiple ensemble classifiers, frequency bands, and stimuli types could be summarized as follows; (1) The most remarkable performances were exhibited in slow frequency bands (delta and theta) in comparison to high frequency and wide range bands, especially in terms of target sounds.(2) Although the highest performance (87.5% accuracy) was obtained in the detection of medication states (PD-ON vs PD-OFF) via standard stimuli responses in the delta band, target CRC distinguishes all groups from each other with accuracy rates of 80% for HC vs. PD-OFF group pair, 80% for HC vs. PD-ON group pair, and 81% for PD-ON vs. PD-OFF group pair.On the other hand, the findings revealed that delta CRC features in the application of novel sounds were not powerful enough to distinguish PD from HC and detect medication states.The classification performances for all group pairs were lower than the 70% accuracy rate.
(3) In the theta frequency band, again target sounds were the most distinctive stimuli to classify HC and PD-OFF (80% accuracy), HC and PD-ON (80.5% accuracy) that could be considered as quite good performances, and PD-ON vs PD-OFF (76% accuracy) that could be considered as acceptable performance based on the claim of the study by Shawen et al (2020).Novel and standard sounds have almost equal contributions that were close to but lower than target sounds to classify each group.
These findings revealed that changes in cognitive performance may be linked to alterations in FC, implying that FC is impacted early in PD (Droby et al 2022).Also, the first finding has not been surprising and totally in line with the expectations since the delta and theta are the dominant oscillatory responses of EEG in the case that a cognitive task is applied (please see the reviews; Harmony 2013, Güntekin and Bas ¸ar 2016, Karakas ¸2020).In addition, these slow oscillatory responses were altered in various neurodegenerative diseases, including PD (Güntekin et al 2022).For example, Solís-Vivanco et al (2018) reported a diminished power and phase alignment of slow oscillations (delta-theta) after the onset of deviant stimuli in PD patients (ON and OFF combined) in comparison to HC.Also, they compared ON and OFF states and showed higher total delta-theta power in medicated patients compared to non-medicated patients without interactions with tone type.Another study by Güntekin et al (2018) showed a gradual decrease in delta response according to the cognitive impairment in patients with PD.The same reduction was seen during the visual stimuli applied to PD-ON in the study of Emek-Savas et al (2017).They reported significantly lower delta event-related oscillations (EROs) in PD-ON patients than HC.In light of these literature backgrounds, it could be guessed that including all frequencies in a wide range would decrease the sensitivity of detecting PD (please see supplementary material 4).Moreover, Güntekin and Bas ¸ar (2010) indicated that target responses have higher coherence values during the auditory oddball paradigm.Therefore, it is an expected finding that the target sounds were the dominant stimuli to differentiate PD in delta and theta bands.Additionally, the second and third findings expectedly revealed that the delta and theta CRC in response to target sounds rather than novel sounds were discriminative features, even though both sounds were rare.As mentioned in the study of Cole et al (2023), slow oscillatory responses could not be specific to novelty, however, could be associated with cognitive-control processes.The second finding also presents a state-of-the-art performance that was achieved with an accuracy of 87.5% in classifying PD-ON x PD-OFF via CRC of standard sounds in the delta band.This finding may reveal that the auditory sensory gating mechanism (Jones et al 2016) which refers to the brain's ability to filter out irrelevant stimuli is affected by medication, so using CRC of standard sounds has a contribution to differentiate medication states of PD.
Various studies employed different modalities to detect PD (Mei et al 2021, Ngo et al 2022).However, cognitive dysfunction is another commonly observed biomarker in individuals with PD, including those who have recently been diagnosed clinically (Fang et al 2020).It is known that almost 39% of individuals newly diagnosed with PD exhibit subjective cognitive impairment (Ay and Gürvit 2022).Additionally, cognitive impairment is identified as a risk factor throughout the PD with dementia (PDD) continuum (Hoogland et al 2017), with 15%-20% of individuals with PD progressing to dementia (Yu and Wu 2022).Therefore, early diagnosis has a crucial role in preventing the progress of impairment.
In the literature, although various studies showed abnormalities of cognitive processes in PD (Güntekin et al 2022) and the association between EROs in specific frequency ranges and specific cognitive domains (Bas ¸ar et al 2001), ERP is the commonly used neural marker of attention and cognition in the studies that adopted ML approaches (Wang et al 2020).A few studies evaluated cognitive oscillatory features, mostly obtained from brain activities in individual locations, in classifying PD patients (table 3).However, according to a comprehensive review, PD is associated with disruptions in the normal functioning of neural networks  (Gao and Wu 2016).Therefore, in light of the literature given above, the current study serves as a pioneering investigation, offering a valuable indicator for the early diagnosis of PD by examining brain connectivity.
One of the studies given in table 3 was reported by Cavanagh et al (2018) which used the SVM classifier fed by Fourier coefficients during auditory stimuli.They achieved 82% of accuracy with novel sound stimuli to dissociate PD-OFF from HC and asserted that the maximally discriminating frequency band was the delta band.They also classified PD-ON and HC and presented lower accuracy around 75% for the same condition.The rest of the conditions in the cognitive task produced lower performance than the novelty condition to discriminate all group pairs.By contrast, the current study obtained the most discriminating features in the target condition, which is associated with cognitive-control processes, and paved the way for higher responses in comparison to standard stimuli, in particular for the delta band (Güntekin and Bas ¸ar 2016).The findings of this study also showed that the novel sound stimuli had an acceptable level of accuracies in the higher frequency bands but there was no consistent evidence of differentiating PD from HC (see supplementary material 1-4).Furthermore, it is notable that not only two group pairs but also PD-ON vs. PD-OFF, which provide detection of medication states, were classified during this study.The findings revealed that delta CRC features in response to target condition differentiate all group pairs successfully with around 80% accuracy.Coelho et al (2023) conducted another study that used the same cognitive task.They run multiple ML models fed by Hjorth features that proposed three measures called activity, mobility, and complexity to classify all group pairs suchlike this study.Although they achieved prominent performances for HC x PD-OFF (89.56% with SVM), and HC x PD-ON (87.78% with k-nearest neighbors (KNN)) by means of accuracy metric, the performance could not reach the same levels for PD-ON x PD-OFF (68.67% with KNN) classification.In addition to using different features of EEG and classifiers, the major distinction between the aforementioned study and the current study is evaluating conditions.Coelho et al (2023) created their feature vector with a combination of all types of stimuli whereas this study evaluated their contributions separately since in an oddball task, the impact as well as the importance of target stimuli in comparison to standard stimuli is not the same (Güntekin and Bas ¸ar 2016).
Two more studies investigated the contribution of cognitive features during 3-stimulus auditory oddball to distinct PD from HC. Smrdel (2022) extracted the features calculated using CSPs and Laplacian operator and archived promising performance, accuracy of 90% with the Laplacian operator and 85% with CSP, for differentiating HC and PD with several conventional ML algorithms.However, the detection ability of the medication state with the given models was not investigated.Moreover, only standard stimuli of the oddball task were employed for the classification process although the brain activity is altered by target stimuli more.The other study is Meghnoudj et al (2023).They extracted Sparse Dynamical Features over different numbers of channels and tested multiple feature vectors by simple classification algorithms.The maximum accuracy rate was 94% with three channels.However, unlike this study, they separate only PD-OFF and HC.
In the literature, several features of brain activity during varying cognitive tasks were employed for the classification of HC vs. PD-ON.Singh et al (2018) trained a linear SVM model with mid-frontal activity features during Simon reaction-time task that involves cognitive control for conflict and error processing and achieved lower than 70% accuracy with cognitive features.They also evaluated resting state data and with all the combination of features, they reached the accuracy level of around 70%, which was lower than this study achieved.Moreover, in addition to studies that distinguish cognitively normal PD patients and without any comorbidity, Tülay et al (2023) classified PD with dementia (PDD)-ON and HC with a promising accuracy of 94% by using inter-trial phase coherence (ITPC), which is a measure of the appearance and degree of consistency over trials, in the delta band at fronto-central channels during visual target stimuli.They also demonstrated the contribution of ITPC in the delta and theta bands during auditory target stimuli with an accuracy of 85%.To differentiate PD with impulse control disorders (ICDs) comorbidity, Lin et al (2023) used event-related spectral perturbation (ERSP) features obtained during the visual Go/Nogo paradigm and ran SVM and support vector regression (SVR) models.They reported accuracies of 63.3% and 53.3% for the classification of PD with ICD vs PD and HC vs PD, respectively.Comparing the aforementioned study and this study is not very reliable because it is not clear in the aforementioned study whether PD is medicated or not.
The only study that is close to this study in terms of the type of features is de Oliveira et al (2020).They provide a distinction of PD with outstanding results, above 99% of accuracy by using partial directed coherence features via ML techniques.However, the used features were obtained during a photic stimulation instead of a cognitive task.Therefore, as mentioned earlier, the current study is the pioneering study in the classification of PD by means of CRCs during auditory stimuli and has shown the valuable contributions of connectivity between distinct brain regions in the delta and theta bands to detect the disease.
As seen in table 3, few studies used cognitive task-related features of EEG to distinguish PD from HC.Among the studies, three of them attempted to classify HC and PD-ON, and only Coelho et al (2023) exceeded the performance of this study.However, they combined all stimuli types, which prevented measuring the contribution of pure cognitive features.Moreover, the aforementioned study is again the only study that classifies PD-ON and PD-OFF with a low accuracy of 68.67% compared to this study which achieved 87.5% accuracy.On the other hand, the studies in the literature on this concept mostly differentiate PD-OFF.Four studies achieved better performances than this study in classifying HC vs PD-OFF, although two of them (Smrdel 2022, Coelho et al 2023) did not evaluate the cognitive effects.Even so, this study provides an alternative feature for identifying PD with closer accuracy rates.
Various cognitive domains, including difficulties in memory, attention, and executive functions, are affected in the spectrum of cognitive impairment associated with PD.According to the literature, dysfunction in the frontostriatal dopaminergic circuitry leads to executive dysfunction (Ay and Gürvit 2022) that is characterized by deficits in internal control of attention, planning, inhibitory control, and on a range of decision-making and can be present from the early stages of PD (Dirnberger and Jahanshahi 2013).Also, patients with this dysfunction do not always progress to PDD.Moreover, treatment with dopaminergic medication has a positive impact on executive deficits in PD.Therefore, detecting PD in the early stages, even in the ON state, has crucial importance in many aspects.First of all, it provides an improvement in quality of life.Moreover, detecting ON state could provide the follow-up of the disease and pave the way for developing new treatment approaches.Clinicians can assess how well the medication is managing the motor symptoms and make adjustments to dosage, timing of medication administration, or more personalized treatment plans as needed, thus minimizing the impact of motor fluctuations and dyskinesias, ultimately improving the overall management of PD.
Although the findings have revealed promising classification performances to detect different medication states of PD and PD in general in this study, several biases/challenges/limitations could be mentioned.One limitation arising is that PD patients were applied to the cognitive task two times, on medication ON and OFF states.This process could cause learning the paradigm.However, according to Singh et al (2018) and references therein, the effect of exposure would not be significant.Additionally, the brain FC features during the cognitive task help to detect Parkinsonian states.However, for more robust detection systems via neural mechanisms underlying specific cognitive functions in PD, more comprehensive studies with different types of stimuli and tasks should be done in the future.Besides, the limited amount of data poses a limitation to running deep learning algorithms.It should be also acknowledged that testing different feature selection methods, classifiers, and hyperparameters would be more informative to find the best classification model.

Figure 1 .
Figure 1.An example to obtain the mean coherence value for the delta band.Red points represent the steps, which are 0.5 Hz.

Figure 2 .
Figure 2. The pipeline of the classification analysis.

Figure 3 .
Figure 3. Target CRC values in delta and theta bands.The blue line represents healthy controls (HCs), the red line represents Parkinson's Disease with medication (PD-ON) and the green line represents Parkinson's Disease without medication (PD-OFF).

F3O1HC:Figure 4 .
Figure 4. ROC curves created from 10-fold cross-validation and the mean AUCs in the delta band for all the classification processes of group pairs.(a) HC vs PD-OFF (b) HC vs -PD-ON (c) PD-ON vs PD-OFF.

Figure 5 .
Figure 5. ROC curves created from 10-fold cross-validation and the mean AUCs in the theta band for all the classification processes of group pairs.(a) HC vs PD-OFF (b) HC vs -PD-ON (c) PD-ON vs PD-OFF.

Table 1 .
Classification metrics of ensemble models with the optimum number of features of CRC in the delta band.

Table 2 .
Classification metrics of ensemble models with the optimum number of features of CRC in the theta band.

Table 3 .
The summary of the studies that focus on cognitive oscillatory features of EEG for identification of Parkinson's disease with machine learning techniques.