(PDF) Classifying oscillatory brain activity associated with Indian Rasas using network metrics

Pandey et al. Brain Informatics (2022) 9:15 https://doi.org/10.1186/s40708-022-00163-7 Brain Informatics RESEARCH Open Access Classifying oscillatory brain activity associated with Indian Rasas using network metrics Pankaj Pandey1*, Richa Tripathi2 and Krishna Prasad Miyapuram1,3  Abstract  Neural signatures for the western classification of emotions have been widely discussed in the literature. The ancient Indian treatise on performing arts known as Natyashastra categorizes emotions into nine classes, known as Rasas. Rasa—as opposed to a pure emotion—is defined as a superposition of certain transitory, dominant, and tempera- mental emotional states. Although Rasas have been widely discussed in the text, dedicated brain imaging studies have not been conducted in their research. Our study examines the neural oscillations, recorded through electro- encephalography (EEG) imaging, that are elicited while experiencing emotional states corresponding to Rasas. We identify differences among them using network-based functional connectivity metrics in five different frequency bands. Further, Random Forest models are trained on the extracted network features, and we present our findings based on classifier predictions. We observe slow (delta) and fast brain waves (beta and gamma) exhibited the maxi- mum discriminating features between Rasas, whereas alpha and theta bands showed fewer distinguishable pairs. Out of nine Rasas, Sringaram (love), Bibhatsam (odious), and Bhayanakam (terror) were distinguishable from other Rasas the most across frequency bands. On the scale of most network metrics, Raudram (rage) and Sringaram are on the extremes, which also resulted in their good classification accuracy of 95%. This is reminiscent of the circumplex model where anger and contentment/happiness are on extremes on the pleasant scale. Interestingly, our results are consist- ent with the previous studies which highlight the significant role of higher frequency oscillations in the classification of emotions, in contrast to the alpha band that has shows non-significant differences across emotions. This research contributes to one of the first attempts to investigate the neural correlates of Rasas. Therefore, the results of this study can potentially guide the explorations into the entrainment of brain oscillations between performers and viewers, which can further lead to better performances and viewer experience. Keywords:  EEG, Emotion, Classification, Natyashastra, Rasas, Movie clips, Random Forest, wPLI, Graph theory 1 Introduction grasp on the audience’s attention and generating certain Our emotions affect our daily lives in many ways and kinds of emotions are driven by the structure of audio– they contribute to cognitive processes such as percep- video placement in a film. A neurocinematics study tion, attention, and decision-making. Films engage view- explores different brain processes and mental states while ers through experiences by capturing their attention and watching movies. In line with this, neuroaesthetic is the stimulating perception, cognition, and emotion. The field that involves the study of esthetic processing in the brain while watching a structured video pertaining to a set of emotions. Esthetic components of audio–video *Correspondence:

stimuli evoke various emotions in our daily lives. 1 Computer Science and Engineering, Indian Institute of Technology The previous studies in neuroaesthetics are mostly Gandhinagar, 382355 Gandhinagar, India Full list of author information is available at the end of the article based on the western classification of emotions. Several © The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. Pandey et al. Brain Informatics (2022) 9:15 Page 2 of 20 such categorizations of emotions have been discussed and patterns, which can be decoded to connect previ- in the literature. Ekman discusses the six basic emo- ous knowledge to new ones to get more depth of brain tions including anger, disgust, fear, joy, sadness, and processes. Slow-to-fast brain rhythms have been dis- surprise [1], besides another six categorizations compris- cussed widely spanning numerous domains of cognitive ing desire, love, sorrow, wonder, happiness, and interest neuroscience. Delta rhythm is the slowest and strong- [2, 3]. Tomkins et  al. [4] in their approach to emotion, est brainwave and is usually associated with the deepest describe nine basic emotions including anger, contempt, form of dreamless sleep. Theta waves are observed in disgust, distress, fear, interest, joy, shame, and surprise. deep meditation and relaxation. Alpha band is associ- The cognitive structure of emotions has also been dis- ated with relaxation or calmness, and alertness. Beta cussed in further 22 forms. In this study, we present our band is marked by the state of wakefulness/conscious- work on the Indian categorization of emotions into nine ness, observed when performing any cognitive tasks classes as described in ‘Natyashastra’: a treatise on per- (e.g., problem-solving, decision-making, etc.). Gamma forming arts. These nine dimensions of emotions corre- frequencies are the fastest brainwaves, correlated with spond to the nine Rasa s (esthetic impact of an artwork). long-range neuronal communication, and facilitating We study these Rasa s as evoked via watching audio-vis- the neural mechanisms underlying attention [10–13]. ual entertainment (movie clips) through electroencepha- Previous studies on emotions highlight the discussion lographic recordings. A Rasa describes a state of mind on frequency bands. Gamma band is ultra-fast brain to indicate emotion. This research is pivotal in under- waves identified to play an important role in human standing the theoretical work done by researchers in the emotions [9, 14]. Gamma band is shown to find the dif- domain of neuroaesthetics, especially Indian esthetics ferences between happy and sad emotions [15]. Beta of performing arts. Several works on Rasa s have been bands are indicated for identifying three emotions: produced, including dance, drama, and paintings [5–7]. positive, neutral, and negative [16, 17]. A recent study However, there is a need to understand the underlying finds the significance of beta and gamma bands in the cognitive processes while observing various Rasa forms. discrimination of low/high valence, low/high arousal This article investigates the role of different brain oscil- [18]. A study of event-related oscillations involving lations while watching nine Rasa s in the form of audio- event-related synchronization/desynchronization has visual clips. discussed the role of the slow (delta) waves in emo- This study recorded electroencephalography (EEG) tional processing in the passive viewing of emotionally responses while participants were watching movie clips evocative pictures [19]. depicting nine Rasa s. EEG has been a principal tool for Previous studies have suggested that functional con- brain research because it reflects electrophysiological nectivity in different frequency bands preserves sig- activity that is representative of brain function and EEG nificant network topology, which may be employed to recording can be conducted at a relatively low cost with classify emotions. In recent years, complex network the- high temporal and useful spatial resolution [8]. EEG ory has gained popularity [20], researchers have shown signals produce high-resolution images of neural oscil- that EEG can be used to build brain networks and the lations, which opens several ways to study the human resulting networks show a number of important topo- brain, from treating mental disorders to understand- logical traits [21]. A functional connection in the brain ing emotions. For example, to study the brain processes is defined typically as the temporal correlation between involved during happy or sad emotions, participants can remote neurophysiological events [22]. Brain activi- view emotional images while recording their brain activ- ties require interactions among multiple brain regions. ity [9]. This opens the research avenue to explore dif- Emotional processing requires the cooperation of many ferent emotional processes based on the stimuli. EEG brain regions, as it is a high-level cognitive function signals represent synchronized electrical pulses from [23]. The study of brain activity mechanisms often relies masses of neurons interacting with one other. Brain on brain networks, which depict relationships between rhythms are primarily divided into five frequency bands, brain regions and information exchange between them differentiated via their morphological and functional [24, 25]. Using functional connectivity, Zhang and col- aspects. These are majorly classified into five frequency leagues identify the interaction of the prefrontal area to bands: delta (1– 4 Hz), theta (4– 7 Hz), alpha (8–13 Hz), most other areas in emotional processing [26]. Gamma beta (13–30 Hz), and gamma (30–45 Hz). Figure  2 dis- waves form more dense connections during the nega- plays the five brain rhythms. tive and neutral valences than beta waves with specific Brain waves are the windows to understanding cogni- sites of right frontal and parietal–occipital regions [27]. tive functions and their underlying brain processes. The According to previous research, functional connectivity morphology of EEG signals encodes complex properties measurements based on EEG data effectively generate Pandey et al. Brain Informatics (2022) 9:15 Page 3 of 20 the representation that may depict neural signatures for This is analogous to neural entrainment [34]—where a different emotional states. rhythmic sensory stimulus synchronizes neuronal activ- Furthermore, functional connectivity has been stud- ity. In the case of performing arts, the performer gen- ied widely by various graph theoretical measures, which erates certain kinds of emotions that may induce the reveal crucial topological features of the brain network entrainment between the performer and viewer. There- [28]. Graph theoretical analysis of human brain net- fore, this research has potential implications for studying works has been utilized in a variety of imaging modali- the entrainment of brain oscillations between performer ties, including EEG/MEG, functional MRI, diffusion and viewer. Such synchrony of oscillations are the key MRI, and structural MRI [29]. The impact of emotional to generating better performances and a better viewer stimuli on large-scale functional brain networks can be experience. measured through the evaluation of parameters such as To the best of our knowledge, this is one of the first centrality and global efficiency [9, 30]. Other network attempts at the scientific study of Rasa s that involves properties such as modularity, node betweenness cen- modern experimental techniques and methodology, e.g., trality, clustering coefficient, and the existence of highly brain imaging through EEG, network construction based connected hub regions have been consistently discussed on weighted phase lag index, and machine learning for in the EEG studies [21, 29, 31]. Several network meas- classification of Rasa s. Such a study is novel and interest- ures are explored to identify the characteristics of emo- ing, especially in the domain of neuroaesthetics, because tional states. Alpha frequency has been found to have the Rasa s are defined as the esthetics associated with an art closest community structure across nine emotions [32]. form experienced by an audience, and are not pure emo- Another study discussed that the clustering coefficient tional states. Through our analyses, we not only find dif- is higher in the left anterior regions of the negative emo- ferences and commonalities in how the nine Rasa s are tions than positive groups [33]. exhibited as brain waves, but also discover results that Evidence from previous studies strongly suggests that complement our contemporary understanding of emo- functional connectivity of different frequency bands tions and brain waves. preserves significant network topology, which may be This article is organized in seven sections: (a) Introduc- employed to classify emotions. In line with these find- tion, (b) The Natyashastra and Rasa s, (c) Data descrip- ings, we extract network features from EEG responses for tion and preprocessing, (d) Methodology, (e) Results, classification between Rasa s. This research is motivated (f ) Discussion, (g) Limitations and future scope, and (h) by the hypothesis that each Rasa may exhibit characteris- Conclusion. tics that are indistinguishable or distinguishable from one another. The following three points state the two primary 2 The Natyashastra and Rasas objectives of this research and the expected outcome: The ‘Natyashastra’ (NS), the ancient Indian treatise on performing arts, which dates back to the second century 1. Which frequency band represents the maximum AD, provides a major basis for the Indian system of cat- indistinguishable and distinguishable pair of Rasas? egorizing emotional states [35]. 2. What pair of Rasa s are indistinguishable and distin- Attributed to Bharata Muni (Sage), the NS provides guishable? instructions on topics such as dramatic composition, 3. We anticipate that the results of our research will be structuring of a play, construction of the stage, acting in line with previous neuroimaging studies on emo- styles, kinds of body movements, costumes, goals of the tion, especially on the role of fast brain waves in clas- art director, etc. [36]. NS has not only influenced various sifying emotions. Some of the indistinguishable and literary traditions in India, such as dance, music, and act- distinguishable pairs reflect the relationship on the ing but propounded Rasa Theory. The prime highlight of pre-defined emotion model. the theory is that although entertainment is the definite desired effect of performance art, it is not the primary This work provides neural correlates of Rasa s in the form goal. As a method of performance by movie actors, Rasa , of brain networks and identifies brain waves that distin- has been an undeniable part of Indian cinema (Bolly- guish them the most and the least. Our research provides wood). In contrast to western method acting, where an insights into the brain processing of emotionally laden actor embodies the character they play, the focus of the movie clips that elicit a certain mood. We believe that our Rasa method is to convey the emotion. Hence, according analysis and results may provide opportunities for per- to Rasa theory, the performers must become the living formers to understand the brain frequencies generated embodiment of the Rasa they depict [37]. while doing an act among an audience; and the same goes A word non-existent in the English language, for other art forms like music, literature and paintings. Rasa expresses a combination of the ‘artist’ and the Pandey et al. Brain Informatics (2022) 9:15 Page 4 of 20 Fig. 1  The nine-dimensional classification of emotions as described in Natyashastra (Indian Rasa Theory). The figure on the left depicts facial expressions corresponding to nine different Rasa s. In the table we give closest English translation of these Rasa s, and the corresponding dominant emotional state (or Sthayi Bhava ) also with the meaning in English. (Image source: https://​www.​youtu​be.​com/​watch?v=​sSdMU​aF3-​18) Fig. 2  Brain rhythms depict the primary five waveforms. The figure shows various frequencies present in the EEG signal. Delta band (1–4 Hz) depicts lowest frequency waves, followed by theta band (4–7Hz), alpha band (8–13 Hz), beta band (13–30 Hz) and gamma band (30–45 Hz) ‘aesthetic’ [38]. Its origins refer to the concept of taste described as an ‘ecstasy’ caused by watching or listen- of cuisine and can mean the essence or flavor. Bharata ing to an art form such as a play or music. Addition- Muni described Rasa as ’extract’, to imply something ally, as opposed to being a single pure thing, Rasa is worthy of being tasted, and asserted that without Rasa a superposition of many sensory inputs that produce the purpose of art is unfulfilled [38]. In [39], Rasa is “a richly textured, emotionally resonant experience Pandey et al. Brain Informatics (2022) 9:15 Page 5 of 20 larger than the sum of its parts” [40]. These parts (or 3 Data description ingredients, described in analogy to a cuisine) of Rasa s The Institute Ethical Committee (IEC) of Indian Institute are the bhavas. These distinguishable bhavas (emo- of Technology, Gandhinagar, approved this study. Prior tional states), when combined creatively, add to give to conducting experiments, all of the participants pro- enjoyable esthetics of a mixture of emotions. Bharata vided informed consent. describes Rasa s as “moods” experienced by the audi- ence, and bhavas are “state of being” portrayed by 3.1 Subjects actors in performing arts. He describes Rasa s and The study involved 20 healthy (mean age: 26 years, 16 bhavas as “cause one another to originate”. Uppal males, 4 females), right-handed students from Indian (2018) [38] describes Rasa s as taste in food, or melody Institute of Technology Gandhinagar. All participants in music, or movement of the body in a dance, while were proficient in the Hindi language, which was also the the bhavas are more discretely conveyed through language of the video clips. All participants were briefed words, gestures, acting, expressions, etc. In light of this about the task and asked to maintain their attention while definition of the Rasa , and the traditional pertinence watching the film clips. Small groups of subjects indepen- of Rasa theory in Indian cinema, we design our study dently scored movie clips from each category of emotion. and look at it through the lens of modern theories of Only those clips were selected with the highest ranking cognition, perception, and computational esthetics. for evoking a particular response for all categories. In the Natyashastra, Rasas (pg. LXXXVI: [41]) are considered as superposition of certain dominant 3.2 Audio‑visual stimuli states (sthayi bhava), transitory states (vyabhicari Bollywood is popular Indian cinema based on the Hindi bhava), and temperamental states (sattvika bhava) of language. We selected nine Bollywood movie clips cover- emotions (pgs. 102, 105: [41]). Out of these only the ing four decades from the 1980s to recent, as shown in the sthayi bhava is transformed into Rasa [38]. There Table 1. These movie clips depicted each Rasa and selec- are eight Rasa s in classical Indian performing arts tion was based on the independent rating from a small which are: Sringaram (erotic), Hasyam (comic), Karu- group of participants. Each film segment had a different nayam (pathetic), Raudram (furious), Veeram (heroic), length because the clips contained narration that had to Bhayanakam (terrible), Bibhatsam (odious), and Adb- be shown for a certain time to evoke a specific Rasa . Film hutam (marvelous). A later addition to the Sanskrit clips ranged in length from 42 s to 2 min 37 s, as shown poetic tradition is a ninth sentiment called Santam in Table 1. (peace) (pg. 102: [41]). The facial expressions and the dominant state (bhava) corresponding to each of these 3.3 EEG data acquisition and preprocessing Rasa s are depicted in Fig.  1 We based our selection EEG recordings were collected while a participant was of movie clips on this classification system and chose asked to watch the selected nine film clips correspond- ones that correspond to each Rasa . In light of the fact ing to nine Rasa s. A high-density Geodesic system of that there are no defined movie clips for this classifi- 128 channels was used for this acquisition with a sam- cation system, the movies we selected represent one pling rate of 250 Hz. A white fixation cross on a blank set of selections. screen preceded each film clip for 10 s, and the order of Table 1  Movie clips used in EEG data collection Movie ID Rasa genre Film name Director Year Duration Start time End time 1 Adbhutam Mr. India Shekhar Kapur 1987 1m 48s 1h 1m 40s 1h 3m 28s 2 Bhayanakam Bhoot Ram Gopal Varma 2003 1m 34s 1h 2m 57s 1h 4m 31s 3 Bibhatsam Rakhta Charitra Ram Gopal Varma 2010 1m 12s 43m 55s 45m 7s 4 Hasyam 3 Idiots Rajkumar Hirani 2009 2m 33s 59m 55s 1h 2m 28s 5 Karunayam Kal Ho Naa Ho Nikhil Advani 2003 2m 37s 2h 47m 41s 2h 50m 18s 6 Raudram Ghajini A.R. Murugadoss 2008 2m 9s 2h 38m 43s 2h 40m 52s 7 Santam Zindagi Na Milegi Dobara Zoya Akhtar 2011 2m 22s 48m 22s 50m 44s 8 Sringaram Umrao Jaan Muzaffar Ali 1981 42s 43m 08s 43m 50s 9 Veeram Lagaan: Once Upon a Time in India Ashutosh 2001 2m 3s 2h 10m 57s 2h 13m Pandey et al. Brain Informatics (2022) 9:15 Page 6 of 20 the films was randomized for each participant. The com- 4.3 On the choice of network metrics as features plete experiment was designed and run in E-primeTM We chose 14 structural metrics calculated from the final and recordings were captured using Net-stationTM. The weighted and thresholded brain networks as features. preprocessing was performed using the Matlab EEGLAB These were: average degree, maximum degree, average package. High-frequency signals after 60 Hz were fil- edge weight, maximum edge weight, network density, tered to avoid noise effects. Raw EEG data mostly contain average clustering coefficient, local efficiency, global movement and eye blink artifacts that can be checked efficiency, number of communities, modularity, transi- carefully and removed to make data useful for analysis. tivity, mean degree centrality, mean node betweenness Therefore, we applied artifact subspace reconstruction to centrality, and mean edge betweenness centrality. As keep the clean continuous data [42]. Following this, we stated before, we hypothesize that the network meas- chunked the data respective to each Rasa across subjects ures obtained from the connection topology carry infor- and used it for further analysis. mation specific to different Rasa  s in different brain frequency bands. This assumption is based on two of 4 Methodology the research findings in neuroscience: one linking graph 4.1 Construction of brain networks theory with brain conditions/states, and the other high- We constructed the functional connectivity networks lighting the role of frequency bands in brain processes. using the EEG signals from each of the participants Several studies that use the graph theoretic framework and for each of the Rasa s. The nodes of these networks [48, 49] to study the complex system of the brain, have were the EEG electrodes and the edges representing the proven that different structural and functional aspects of strength of connections between the nodes were evalu- the brain are captured by EEG-based connectivity pat- ated using a measure called weighted Phase Lag Index terns of the brain network [30, 50–54]. These studies (wPLI). The wPLI that quantifies the phase synchrony have highlighted that such brain functional networks can between any two time-varying signals, is a standard func- be characterized in terms of complex network proper- tional connectivity measure used in the network neuro- ties, such as node betweenness, small-worldness, hubs, scientific community. The wPLI is defined as the extent of and modularity. Moreover, they demonstrated that these absolute phase lag or lead between two signals weighted structural connectivity metrics could also distinguish by the imaginary component of the cross-spectral power between different cognitive states and pathophysiologi- density between these signals. It is robust to the vol- cal states of brain [54]. Since our network connections ume conduction, presence of noise, and biases induced are governed by the phase relationship of EEG signals by sample size in the electrophysiological data [43–45]. between electrodes, they capture the functional dynamic Firstly, the EEG time-series signals from each of the elec- connectivity pertaining to the activation of brain path- trodes were segmented into 5-s-long windows or epochs ways of emotions. The brain frequencies as observed in with an overlapping window of length 2.5 s. Followed clinical EEG, on the other hand, have played an enor- by filtration in five frequency bands, namely, delta: 1–4 mous role in cognitive research [55]. Different frequency Hz, theta: 4–7 Hz, alpha: 8–13 Hz, beta: 13–30 Hz, and bands, their power content, and amplitude have been gamma: 30–45 Hz. The wPLI measure between a pair of found to be specific to various basic cognitive engage- signals calculates the average over the number of epochs. ment states such as wakefulness, sleep, and attention, For our computation of wPLI, we used MNE-python’s brain diseases such as depression in Parkinson’s [56], and connectivity module [46]. This gives us the five different schizophrenia [57], the accuracy of working memory in coupling matrices (weighted adjacency matrix) each of adults [58] and encoding personality traits [59]. size (128×128) pertaining to each of the five frequency bands, for each participant and each Rasa. 4.3.1 Definitions of network metrics In this section, we define each of the network metrics 4.2 Thresholding of brain networks [28] used as features in this work. We use the NetworkX Functional networks mostly preserve weak and errone- Python library [60] to evaluate each of these metrics: ous connections, which may conceal the topology of cru- cial connections [21]. Thresholding is commonly used to 1. Average degree (AD): The degree of a node in a net- remove a percentage of the weakest links to retain a usa- work is the number of its neighbors or the number ble sparse network. We applied the thresholding process of nodes that it directly connects to. The average as implemented in the paper [47]: the network should be of this number over all the nodes in the average 97% connected, and the average degree should be greater degree. than 2 ∗ log(n) , while maintaining the highest threshold 2. Maximum degree (MD): It is the maximum of all value for edge weights, where n is the number of nodes. the node degrees in a network. Pandey et al. Brain Informatics (2022) 9:15 Page 7 of 20 3. Average edge weight (AEW): Edge weight is the lies in a large number of such shortest paths, it has strength of an edge connecting given two nodes in a high node betweenness centrality. Average NBC a network. Average edge weight is the mean of edge is the average over all nodes. weights over all the edges in the network. 1 4. Average edge betweenness centrality (EBC): Simi- 4. Maximum edge weight (MEW): It is the maximum larly, for an edge, the edge betweenness centrality of all the edge weights in the network. In other measures the number of shortest paths on the net- words, it is the strongest connection present in the work to which this edge belongs. Average EBC is network. average over all network edges. 5. Density (D): It is the ratio of the total number of edges present in the network to the number of pos- sible edges in the network. 4.4 Random Forest (RF) classifiers 6. Average clustering coefficient (ACC): The clustering Network metrics from different networks were used as coefficient of a node measures the fraction of trian- features for the classification. In this study, we trained gles involving that node. In other words, it meas- binary classifiers between two given Rasa s. We selected ures the extent to which its neighbors tend to form Random Forest (RF) classifier for this research due to a complete graph. The average clustering coeffi- its well-established theory and easy interpretability [61]. cient is the average of this quantity over all nodes. RF predicts the class based on a number of fitted deci- 7. Local efficiency (LE): For a network node, it is sion tree classifiers on various sub-samples of the data- defined as the inverse of the average shortest path set. Features are used to build decision trees, where a length of all its neighbors among themselves. It feature denotes a node, and a threshold is used to split measures how robust the network is to the failure the node into two children nodes. The quality of the split of this particular node in terms of its communica- is decided using the Gini criteria. Once the trees are fit- tion efficiency. ted, and optimum thresholds are identified, the final class 8. Global efficiency (GE): Similarly, global efficiency is selected by the majority vote. RF controls over-fitting measures the efficacy of distant information trans- and averaging improves the predictive accuracy. We fer in a network. It is defined as the inverse of the performed validation using the tenfold stratified tech- average characteristic path length between all node nique. The classifier’s performance was evaluated using pairs present in the network. accuracy, precision, recall, and f1-score. Models were 9. Number of communities (NC): A community in the developed using scikit-learn python [64]. The complete network is the set of nodes that have denser con- process of construction of networks to classification is nections or a higher number of edges within this shown in Fig. 3. The input to the random forest was the node-set, than to other nodes or communities in number of subject samples × 14 features. The ‘number of the network. A modular network is organized into the tree’ was set to 100 trees in the forest, the ‘quality of clearly identifiable communities. the split’ was measured by Gini impurity, ‘max depth’ of 10. Modularity (M): Modularity is the measure of the the nodes of the trees were spread until all leaves were extent to which a network is divided into commu- pure or leave had minimum split samples. The ‘min sam- nities. This measure is often used as a quantity that ples’ were set to 2 for the minimum number of samples is optimized, in various community detection algo- required to split an internal node. ‘Min sample leaf ’ was rithms. set one for the minimum number of samples required to 11. Transitivity (T): Transitivity is the ratio of thrice be at a leaf node. ‘Min weight fraction’ on the leaf was set the number triangles to the number of connected to equal weight, and the ‘max features’to consider when triples of nodes in the network. looking at the split was sqrt(number of features). 12. Average degree centrality (ADC): Centrality is the Previous studies have suggested employing permu- measure of the importance of the node in a net- tation-based p-values for assessing the competence of work, or how central is the node to overall network a classifier [62, 63]. This test is proposed to measure the connectivity. The degree centrality of a node is a real connection between the data and the class labels, fraction of the number of links a node has to the and learning signifies a real class structure. We used the total number of potential links it can have in the permutation test with 10,000 rounds with fivefold cross- network. validation to examine the statistical significance of the 13. Average node betweenness centrality (NBC): classifier. The permutation test shuffles the labels of the Betweenness centrality measures how often a node instances to evaluate the significance of the classifier. This bridges the connections between any two pairs of test [63] has been utilized extensively in the literature and nodes in a network via the shortest path. If a node Pandey et al. Brain Informatics (2022) 9:15 Page 8 of 20 Fig. 3  The workflow of the present paper: (Box 1) EEG data acquisition is performed when a participant watches movie clips. The subsequent step involves preprocessing and segmentation of EEG signals into epochs of 5 s. Then the extracted segments are passed for frequency decomposition into five frequency bands comprising delta (1–4 Hz), theta (4–7 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–35 Hz) bands. This is followed by the construction of brain networks with a threshold that retains the significant connections. Network properties are computed from the thresholded functional networks. These network properties are then used as features to build binary classifiers between Rasa s. Resultant models are selected based on the significance of the permutation test (Box 2). From the selected models, we identify the distinguishable and indistinguishable pairs and the frequency bands in which these pairs appear the results discussed via the permutation test are effective. result in significantly different results, hence, we gener- A small p-value suggests that there is a real dependency ated two-dimensional embedding using seven values: [5, between features and targets, which has then been used 10, 15, 20, 30, 40, 50]. by the estimator to give good predictions. A high p-value may indicate little or no relationship between the fea- 4.6 Statistical analysis tures and targets or that the estimator could not use the To test for gender and age effects on the features relationship to make good predictions. Majorly, the per- extracted, we averaged the network features extracted for mutation test procedure assesses how likely a particular male and female participants for each emotion and found accuracy score would be observed by chance. We have that the measures for the two genders were strongly cor- used the implementation of sklearn [64]. related ( R2 > 0.95 ). Similarly, we did not find any cor- relation between age and any of the features extracted 4.5 Visualization ( p > 0.05 ). Hence, we do not consider these two factors The obtained feature matrix was high dimensional, in further analyses. which limits the visualization in two-dimensional space; therefore, we applied t-distributed Stochastic Neigh- 5 Results bor Embedding (t-SNE) to generate lower-dimensional 5.1 Findings from classification embedding [65]. t-SNE is a manifold learning unsuper- We developed binary classification models between pairs of vised approach for non-linear dimensionality reduction. Rasa s (emotions) across five bands. There were 36 models It transforms the data into a low-dimensional space for built for each band comprising a total of 180 (36 × 5) mod- visualization. EEG signals contain non-linearity and rep- els. In Fig. 4, the first column depicts the test accuracy score resent manifold brain processes, therefore we applied for each model across bands and the second column men- this technique to observe the manifolds that can retain tions the respective significance scores (p-values). Based on the non-linear relationship of the dataset. There is one the p-value, we segregated the Rasa pairs as either indis- parameter, perplexity, which defines the variance of the tinguishable or distinguishable. Indistinguishable refers to Gaussian distribution. Different values of perplexity (See figure on next page.) Fig. 4  [Left column] Matrices represent the test accuracy between each pair of Rasa s across five frequency bands. Order of frequency bands from top to bottom is: delta, theta, alpha, beta, and gamma. [Right column] The corresponding p-value indicates the statistical significance of test scores between Rasa pairs Pandey et al. Brain Informatics (2022) 9:15 Page 9 of 20 Fig. 4  (See legend on previous page.) Pandey et al. Brain Informatics (2022) 9:15 Page 10 of 20 the pair whose p-value was greater than 0.1, whereas the Bibhatsam formed a discrimination group ( p < 0.001 ) with distinguishable pair had p-value less than 0.01. Santam, Veeram, Karunayam, and Sringaram, with accura- cies of 88%, 82%, 85%, and 82%, respectively. Theta band Definition  Indistinguishable pair implies that the clas- showed distinguishable pairs of Sringaram with six other sification model was unable to discriminate between Rasa s with accuracy approximately above 90%, except for features of Rasa s. In contrast, distinguishable pair rep- Bibhatsam and Bhayanakam. The alpha band for Sringa- resents that the model determined the discriminating ram formed only two discriminating pairs ( p < 0.001 ) properties between Rasas. with Santam and Hasyam. For Sringaram, beta and gamma bands showed a similar relationship as depicted in the delta band. In the beta band, Bibhatsam formed two pairs ( p < 0.001 ) with Hasyam and Adbhutam. Gamma band 5.1.1 Indistinguishable pairs revealed the same pairs as delta. Bhayanakam with Karu- We selected the indistinguishable pairs based on the two nayam depicted significant discrimination across the delta, thresholds on the p-value, i.e., p > [0.1, 0.5] , and plotted beta, and gamma bands. In Table 2, the classifier’s perfor- them as shown in Fig.  5. The nine Rasa s are arranged in mance is shown for delta and gamma bands. a circular layout and the existing links between the Rasa We projected the feature matrix to a lower-dimensional pairs represent that they are indistinguishable in a given space, which made it easier for interpretation. We applied band. The maximum number of such pairs were found in an unsupervised t-SNE dimensionality reduction technique the alpha and theta bands; whereas, the delta, beta, and on the obtained distinguishable pairs ( p < 0.001 ) in the gamma bands showed lesser pairs. To illustrate the indis- delta, and gamma bands. We observed clear separation in tinguishable pairs more clearly, we constructed Venn dia- some pairs as shown in Fig.  8. For example, Sringaram’s grams based on the obtained relationships, as shown in data points clustered mostly in a corner of the 2-dimen- Fig.  6. The overlap between any two Rasa s depicts that sional feature space separated from the other Rasa s. Sec- the pair is indistinguishable. One such example is that of ondly, Karunayam with Bibhatsam and Bhayanakam Bibhatsam and Bhayanakam, a pair that is largely indistin- reflected a clear separation of data points in delta and guishable, except in the beta band. However, for p > 0.5 gamma bands. Similarly, Bibhatsam showed spatial separa- they formed a indistinguishable pair only in delta and alpha tion with Santam and Veeram Rasa s. We rendered the 2D bands. view using t-SNE, but there might be better separability in Key finding: Theta and alpha bands formed maximum the higher dimensions. indistinguishable pairs. Key finding: Slow wave (delta band) and fast wave (beta and gamma bands) formed maximum distinguishable 5.1.2 Distinguishable pairs pairs. The smaller the p-value, the stronger the evidence to have the discriminating features between two classes. Therefore, 5.2 Interpreting outcome of classifiers using network we selected two thresholds ( p < [0.01, 0.001] ) and, respec- metrics tively, plotted the distinguishable pairs in Fig. 7. The alpha In this section, we aim to obtain an intuitive understand- band formed the minimum distinguishable pairs followed ing of the classification results obtained in the previ- by the theta band, whereas the delta, beta, and gamma ous sub-sections by analyzing the network properties of bands revealed the maximum distinguishable pairs. The the different brain networks. In this pursuit, we take two delta and gamma bands showed a similar set of distinguish- approaches, one where the network metrics are averaged able pairs when p < 0.001 . Sringaram reflected the sig- across Rasa s for each frequency bands, and second, where nificant distinction from other Rasa s across bands, and in the network metrics are averaged across frequency bands the delta band it showed a classification accuracy of above for each Rasa. 90% ( p < 0.001 ) with other Rasa s except for Bibhatsam. (See figure on next page.) Fig. 5  A connection between two Rasa s represents an indistinguishable pair. The top and bottom rows represent the connections obtained with a p-value greater than 0.1 and 0.5, respectively. Indistinguishable pair implies that the model was unable to distinguish between characteristics of Rasa s. [From the top, in the anticlockwise direction the Rasa s are in order: Santam (pink), Hasyam (red), Bibhatsam (green), Sringaram (yellow), Adbhutam (cyan), Bhayanakam (orange), Karunayam (purple), Veeram (blue), and Raudram (dark green).] Pandey et al. Brain Informatics (2022) 9:15 Page 11 of 20 Fig. 5  (See legend on previous page.) Pandey et al. Brain Informatics (2022) 9:15 Page 12 of 20 Fig. 6  The top and bottom rows represent the Venn diagrams obtained with a p-value greater than 0.1 and 0.5, respectively. The presence of more than one Rasa in a set indicates similar indistinguishable connections 5.2.1 Analyzing frequency bands after averaging network 5.2.2 Analyzing Rasa s after averaging network metrics metrics across Rasas across frequency bands For each frequency band, we averaged the magnitude of For each Rasa , the magnitude of network metrics after network metrics over all Rasa s. The averaged metrics are averaging over the five frequency bands is shown in shown in Table  3, with maximum and minimum values Table  4, with the minimum and maximum values high- across the bands shown in bold fonts. From these values, lighted in bold. The minimum and maximum aver- we examine the similarities and differences between bands. age degrees were indicated by Sringaram (16.28) and Gamma band showed the minimum average degree, fol- Raudram (24.04). The maximum degree was found in lowed by the delta and beta bands. The maximum degree three sets that had magnitude above 50, 60, and 70. Srin- is observed in theta and alpha bands. Gamma had the garam had the least maximum degree of 56.11, Bibhat- minimum average edge weight, whereas alpha had the sam (64.02) and Bhayanakam (67.31) formed another maximum value. The network density was minimum in group of above 60. Hasyam (70.01), Adbhutam (71.87), the gamma band, followed by delta, beta, alpha, and theta Santam (73.4), Karunayam (73.83), Veeram (73.98), and bands. Delta band had the minimum average clustering Raudram (74.76) were above 70. The average edge weight coefficient, whereas the maximum was in the alpha band. between 0.40 and 0.46 comprised Karunayam (0.424), Similar observations were repeated for the rest of the net- Santam (0.429), Hasyam (0.442), Veeram (0.449), and work metrics. Delta or gamma band indicated the mini- Raudram (0.45). Adbhutam (0.46), Bhayanakam (0.47), mum magnitudes of network metrics, whereas alpha or and Bibhatsam (0.50) observed within 0.52. And the max- theta band maintained the maximum value. There were imum was for sringaram (0.58). Density ranges from 0.17 only two exceptions where the role was reversed—gamma to 0.19 included Raudram (0.189), Santam (0.182), Karu- band showed a maximum, and theta band a minimum. In nayam (0.181), Hasyam (0.1789), Adbhutam (0.1787), contrast, average node and edge betweenness centrality and Veeram (0.1782). Bhayanakam (0.16), Bibhatsam (NBC, EBC) showed a descending order of magnitudes (0.15), and Sringaram (0.12) had the least three values. from gamma, delta, beta, alpha, and theta bands. We drew For the remaining metrics (before ADC, as shown in the top 5% of the network connections in Fig. 9. Table 4), we found that Sringaram had minimum magni- Key finding: Delta and gamma bands have lower magni- tude, whereas Raudram and Karunayam had maximum. tudes of network metrics, whereas theta and alpha bands In contrast, average node and edge betweenness centrali- retained higher magnitudes, except for NBC and EBC. ties showed minimum values for Raudram and maximum for Sringaram. (See figure on next page.) Fig. 7  A connection between two Rasa s represents a distinguishable pair. The top and bottom rows represent the connections obtained with a p-value less than 0.01 and 0.001, respectively. Distinguishable pair implies that the model was able to distinguish between characteristics of Rasa s. [From the top, in the anticlockwise direction the Rasa s are in order: Santam (pink), Hasyam (red), Bibhatsam (green), Sringaram (yellow), Adbhutam (cyan), Bhayanakam (orange), Karunayam (purple), Veeram (blue), and Raudram (dark green).] Pandey et al. Brain Informatics (2022) 9:15 Page 13 of 20 Fig. 7  (See legend on previous page.) Pandey et al. Brain Informatics (2022) 9:15 Page 14 of 20 Fig. 8  Network features of a distinguishable pair ( p < 0.001 ) are projected in lower-dimensional 2D space using t-SNE. The top and bottom rows represent the features extracted from delta and gamma bands, respectively Table 2  The classifier’s performance is evaluated using accuracy, precision, recall, and F1-Score Class1 Class2 Band Accuracy Precision Recall F1-Score p-value Bibhatsam Santam Delta 0.875 0.9 0.85 0.8666 0.00009 Bibhatsam Veeram Delta 0.825 0.8833 0.8 0.81 0.0006 Bibhatsam Karunayam Delta 0.85 0.8833 0.85 0.8433 0.0002 Bhayanakam Karunayam Delta 0.85 0.8833 0.85 0.8433 0.0009 Bibhatsam Santam Gamma 0.85 0.9 0.84 0.84 0.00009 Bibhatsam Veeram Gamma 0.825 0.8833 0.8 0.81 0.0002 Bibhatsam Karunayam Gamma 0.875 0.9666 0.8 0.8466 0.0004 Bhayanakam Karunayam Gamma 0.8 0.8 0.9 0.8233 0.0006 p-value is obtained by permutation test Key findings: happiness and sadness, the gamma band has been the optimal band for generating discriminating features [15]. • Ten out of fourteen network properties suggested A recent study [33] presents that the beta and gamma that the Rasa s Sringaram and Raudram limit the are more effective brain rhythms in identifying emotions magnitude of network features. Based on this, we than the theta and alpha. Furthermore, some neurosci- inferred a magnitude scale as shown in Fig. 10, where ence studies reveal that neural encodings of emotional Sringaram determined the one side limit of the scale, information are stored primarily in higher frequency while Raudram maintained the other side. bands [69, 70]. Another recent paper by Yang and col- • In contrast to other network metrics, node, and edge leagues reports that long-distance connections noted in betweenness centralities are maximum in the Sringa- the high-frequency bands, especially in the high gamma ram and minimum in Raudram. bands, showed significant differences among emotional • The network properties of Bibhatsam and Bhay- states [9]. Brain activities in the high-frequency band (> anakam were nearly close to each other. 30Hz) are known to be associated with emotional inte- gration and play a role in cognitive control of emotions [71, 72]. Several studies have looked at those high-fre- 6 Discussion quency responses to affective pictures, most of which Higher frequency has been consistently reported to reported enhanced responses to negative images [73–75]. be crucial for the classification of different emotions Zheng and colleagues observe that the delta band per- [66–68]. In the previous study for the classification of formed better than the theta and alpha bands for emotion Pandey et al. Brain Informatics (2022) 9:15 Page 15 of 20 Fig. 9  Connectivity graphs of Rasa s depict 5% of the strong connections across bands. The node’s size indicates the degree, and the width and color of the edges denote the connection strength measured using wPLI index (averaged over all the 20 participants). Blue and red colors indicate the minimum and maximum strength, respectively. The visualizations are generated using the ’BrainNet Viewer’ (www.nitrc.org/projects/bnv/) Table 3  Frequency bands after averaging across all Rasas Band AD MD AEW MEW D ACC​ GE LE NC M T ADC NBC EBC Delta 19.1262 63.4888 0.4605 0.7311 0.1506 0.2593 0.5050 0.4550 4.2611 0.0110 0.2434 0.1506 0.0087 0.0019 Theta 29.5588 75.3388 0.5033 0.7966 0.2327 0.3262 0.5625 0.5495 4.2444 0.0119 0.3319 0.2327 0.0071 0.0012 Alpha 26.1502 75.8944 0.5395 0.8440 0.2059 0.3955 0.5444 0.5956 4.2777 0.0270 0.3609 0.2059 0.0077 0.0014 Beta 20.9482 70.3 0.4262 0.7958 0.1649 0.3873 0.5137 0.5708 4.2388 0.0191 0.3453 0.1649 0.0086 0.0019 Gamma 12.6756 62.3611 0.416 0.8448 0.0998 0.3253 0.4661 0.4872 3.9333 0.0239 0.2392 0.0998 0.0101 0.0030 Maximum and minimum values are highlighted recognition of three categories (positive, neutral, and Interestingly, our result on the alpha band resonated negative) [76]. They observe this outcome from the fea- with the previous research on Indian Rasa s [32]. This tures of differential asymmetry and rational asymmetry. study reports that the community structure of differ- Delta band is less studied in the literature, and a recent ent Rasa networks in the alpha band is the most simi- study on event-related emphasizes the research on delta lar. Similar observation about the indistinguishability activity patterns and alterations in delta energy, which of (two) emotions in alpha band was also reported in might improve our understanding of emotional process- [77]—a study aimed at discriminating multiple emo- ing by focusing on the slow waves (delta band) [19]. tional states using EEG data collected from subjects watching emotion-inducing video clips. According to Pandey et al. Brain Informatics (2022) 9:15 Page 16 of 20 Table 4  Average network metrics for each Rasa obtained after averaging across all bands Rasa AD MD AEW MEW D ACC​ GE LE NC M T ADC NBC EBC Raudram 24.0454 74.76 0.4527 0.8040 0.1893 0.3630 0.5334 0.5653 4.18 0.0099 0.3235 0.1893 0.0079 0.0016 Santam 23.2392 73.4 0.4291 0.7757 0.1829 0.3586 0.5298 0.5589 4.08 0.0108 0.3217 0.1829 0.0081 0.0017 Karunayam 23.0126 73.83 0.4245 0.7712 0.1812 0.3601 0.5290 0.5564 3.97 0.0070 0.3201 0.1812 0.0082 0.0017 Hasyam 22.7248 70.01 0.4425 0.7894 0.1789 0.3548 0.5241 0.5534 4.12 0.0177 0.3136 0.1789 0.0083 0.0018 Adbhutam 22.6998 71.87 0.4608 0.8037 0.1787 0.3529 0.5239 0.5446 4.15 0.0089 0.3155 0.1787 0.0083 0.0018 Veeram 22.6351 73.98 0.4492 0.8048 0.1782 0.3608 0.5263 0.5594 4.17 0.0136 0.3170 0.1782 0.0082 0.0018 Bhayanakam 20.6726 67.31 0.4749 0.8079 0.1627 0.3262 0.5108 0.5180 4.28 0.0146 0.3025 0.1627 0.0086 0.0019 Bibhatsam 19.9148 64.02 0.5010 0.8190 0.1568 0.3139 0.5052 0.5007 4.28 0.0215 0.2897 0.1568 0.0088 0.0020 Sringaram 16.2820 56.11 0.5878 0.8463 0.1282 0.2582 0.4827 0.4274 4.49 0.0631 0.2336 0.1282 0.0096 0.0025 Maximum and minimum values are highlighted another recent study, emotional stimulus processing is conducted in various countries and regions, this model is associated with a decrease of power in the alpha and mostly accurate and consistent [82, 83]. beta bands across studies and task conditions [78]. A summary of our results obtained on distinguish- Most EEG studies include 15-20 participants because of able pairs in different frequency bands is presented the complexity in EEG setup and data collection. There- in Table  5. We observe that Bibhatsam (disgust), an fore, with a small number of samples, some techniques unpleasant emotion, was distinguishable from Santam are proposed to identify the significance of the Machine (peace) and Veeram (heroic), both pleasant emotions, Learning performance estimates. The recent article [79] in the delta and gamma bands. We also find that Bib- highlights the strong biases observed using solely K-fold hatsam (disgust) and Karunayam (sorrow), both repre- cross-validation, and therefore it is significant to use senting unpleasant emotions, formed a distinguishable rigorous methods for analysis. Hence, our study mainly pair. On noticing the activeness scale, however, Bibhat- used fivefold cross-validation with permutation test of sam and Karunayam indicate high and mild intensity 10,000 rounds that produced robust and unbiased per- emotions, respectively, and hence this pair although formance estimates regardless of the sample size. Based similar on the pleasant dimension, it is dissimilar on on our results from classifiers and magnitudes of network activeness scale. In beta band, Hasyam (comic) and metrics, we observe that Bhayanakam (fear) and Bibhat- Adbhutam (astonishment) were distinguished from sam (disgust) exhibited high similarity. They both signify Bibhatsam. Similarly, Bhayanakam and Karunayam unpleasant emotions as per the circumplex model. Rus- formed a distinguishable pair in delta, beta, and sell and James proposed a circumplex model for emo- gamma bands, indicating high and mild intensity. tions classification [80]. This is more related to dimension space theory, which refers to emotion as continuous and relevant [81]. The circumplex model describes emotion 7 Limitations and future scope into two dimensions: pleasure and activeness, as shown We would like to mention a few limitations of this in Fig.  11. Activeness is categorized into mild and high study. Our results are based on the scalp electrodes, intensity, while pleasure is classified into pleasant and which do not have clearly defined source mapping unpleasant. Based on the results of previous research inside the brain, and therefore we confine the findings Fig. 10  Network Rasa scale: Sringaram and Raudram form the limiting boundaries for the magnitude of network properties, and all the other Rasa s fall within those limits. (*) over a set of Rasa s denotes that their order is not necessarily the same as shown, and it may slightly vary with frequency bands. For some network metrics, they may share the properties or may differ. Bibhatsam and Bhayanakam are consistently close to each other across network metrics and also across the bands Pandey et al. Brain Informatics (2022) 9:15 Page 17 of 20 for extending the Rasa analysis on different races as well, and explore the similarities and differences (if any) from the results reported in this paper. This research contributes to the pioneering work on Indian Rasa s , reporting network-based similarity and differ- ences in brain responses collected through EEG. 8 Conclusion In this work, we computed the functional connectiv- ity networks, corresponding to nine Rasa s , that repre- sented the correlations between the activities of brain regions while a person was watching emotional movie clips. In order to identify distinguishable and indistin- guishable pairs of Rasa s , the network features from the corresponding functional networks were employed for the classification task. Our binary classification result (accuracy) between a given Rasa pair, were re-affirmed with a permutation test. The two key findings of our study are as follows: Fig. 11  A circumplex model of emotions classification. The model has two dimensions encompassing pleasure and activeness 1. Slow (delta band) and fast (beta and gamma bands) brain waves generated the maximum number of dis- tinguishable pairs. 2. Theta and alpha rhythms exhibited higher number of in the signal space rather than source space. We used indistinguishable Rasa s pairs. a single set of film clips (corresponding to different Rasa s) which was selected based on the ranking from Our classification results also highlighted the role of a group of participants who confirmed the evoking of frequency bands in examining the differences between these particular emotions. The study could be carried emotions. We found that the delta, beta, and gamma out on more such film clip sets for nine Rasa s . Hence, produced the maximum number of distinguishable this study motivates building a benchmark dataset of pairs, whereas theta and alpha waves resulted in more audio-visual stimuli corresponding to Rasa s for EEG indistinguishable pairs, for which the classifiers failed studies. We acknowledge that thresholding on p-value to generate discrimination with statistical significance. can also vary based on pairs. However, the main objec- In addition, to gain interpretability of the obtained two tive of the article is to present the significance of bands groups of frequency bands, we analyzed network prop- by utilizing network features. Hence, future works erties and observed that the magnitudes of the delta, would have ample opportunity to see the pair-wise beta, and gamma networks were mostly lower than differences and similarities in more depth, includ- theta and alpha bands. ing the role of network features. The EEG experiment In the delta band, a pair between Bibhatsam and San- involved only Indian students, hence there is a scope tam obtained the maximum accuracy of 87.5% with precision, recall, and f1-score of 0.9, 0.85, and 0.86, respectively. A pair between Bibhatsam and Karunay- Table 5 Distinguishable pairs (p < 0.001) of Rasa s in different ama showed an accuracy of 85% with precision, recall, frequency bands and f1-score of 0.88,0.85, and 0.84, respectively, and similar performance was achieved for Bhayanakam Rasa 1 Discriminated ( Rasa s 2) Band and Karunayam. The classification accuracy between Bibhatsam Santam, Veeram and Karunayam Delta and Gamma Bibhatsam and veeram was 82.5% with precision, Bibhatsam Hasyam and Adbhutam Beta recall, and f1-score of 0.88, 0.85, and 0.84, respec- Bhayanakam Karunayam Delta, Beta and Gamma tively. We obtained a similar relationship as the delta Sringaram All Delta, Beta and Gamma in the gamma band, with the highest accuracy of 87.5% Sringaram except Bibhatsam and Bhay- Theta between Bibhatsam and Karunayama with precision, anakam recall, and f1-score, of 0.96,0.8,0.84, respectively. Bib- Sringaram Santam and Hasyam Alpha hatasam and Santam showed an accuracy of 85%. In the Pandey et al. Brain Informatics (2022) 9:15 Page 18 of 20 beta band, we obtained a maximum accuracy of 85% Author details 1  Computer Science and Engineering, Indian Institute of Technology Gandhi- between Bibhatsam and Hasyam. Among all Rasa s , we nagar, 382355 Gandhinagar, India. 2 Center for Advanced Systems Understand- obtained a maximum classification accuracy of 97% in ing (CASUS), Helmholtz-Zentrum Dresden-Rossendorf, Görlitz, Germany. the delta band between Sringaram and Adbhutam, fol- 3  Centre for Cognitive and Brain Sciences, Indian Institute of Technology Gandhinagar, 382355 Gandhinagar, India. lowed by beta and gamma bands with 95% and 94% with Raudram and Santam, respectively. Received: 14 December 2021 Accepted: 28 June 2022 Based on the magnitudes of the network metrics, we observe that the Raudram (for 10 network met- rics) and Sringaram (for all network metrics) Rasa s are the extreme emotions, i.e., one of them has a minimum References (maximum) value, while the other has a maximum (mini- 1. Ekman P (1993) Facial expression and emotion. Am Psychol 48(4):384 2. Parrott WG (2001) Emotions in social psychology: essential readings. mum) magnitude. The other seven Rasa s, lie in the range Psychology Press, London set by these two for all the network metrics. Based on this 3. Frijda NH et al (1986) The emotions. Cambridge University Press, observation, we approximate a Rasa scale, where they are Cambridge 4. Tomkins SS (1984) Affect theory. Approaches to emotion placed next to each other on a one-dimensional line. The 163(163–195):31–65 ordering goes as Sringaram, Bibhatsam, Bhayanakam, 5. Chakravorty P (1998) Hegemony, dance and nation: the construction of Adbhutam, Veeram, Hasyam, Karunayam, Santam, the classical dance in India. South Asia J South Asian Stud 21(2):107–120 6. Sharma PB (2013) Painting: a tool of non-verbal communication. Lang Raudram, where the last six can interchange their posi- India 13(7):312–318 tions depending on the frequency bands. 7. Mehta T (1995) Sanskrit play production in ancient India. Motilal Banarsi- dass Publ, New Delhi Acknowledgements 8. Kothare SV (2014) Atlas of EEG Patterns, 2nd edn. Neurology 83(7):668. We thank Science and Engineering Research Board (SERB) and PlayPower https://​doi.​org/​10.​1212/​WNL.​00000​00000​000696 Labs for supporting the Prime Minister’s Research Fellowship (PMRF) awarded 9. Yang K, Tong L, Shu J, Zhuang N, Yan B, Zeng Y (2020) High gamma band to Pankaj Pandey. We thank Federation of Indian Chambers of Commerce & EEG closely related to emotion: evidence from functional network. Front Industry (FICCI) for facilitating this PMRF. This work was partially funded by Hum Neurosci 14:89 the Center of Advanced Systems Understanding (CASUS), which is financed 10. Fries P (2005) A mechanism for cognitive dynamics: neuronal communi- by Germany’s Federal Ministry of Education and Research (BMBF) and by the cation through neuronal coherence. Trends Cogn Sci 9(10):474–480 Saxon Ministry for Science, Culture and Tourism (SMWK) with tax funds on the 11. Salinas E, Sejnowski TJ (2001) Correlated neuronal activity and the flow of basis of the budget approved by the Saxon State Parliament. We thank Dyuti- neural information. Nat Rev Neurosci 2(8):539–550 man Mukopadhyay for the experimental design and data collection. 12. Landau AN, Esterman M, Robertson LC, Bentin S, Prinzmetal W (2007) Different effects of voluntary and involuntary attention on EEG activity in Author contributions the gamma band. J Neurosci 27(44):11986–11990 PP was mainly responsible for the conceptualization, methodology, imple- 13. Tallon-Baudry C, Bertrand O, Hénaff M-A, Isnard J, Fischer C (2005) Atten- mentation, production of charts, and manuscript writing of this study. RT tion modulates gamma-band oscillations differently in the human lateral was primarily involved in conceptualization, methodology, reviewing, and occipital cortex and fusiform gyrus. Cereb Cortex 15(5):654–662 writing of the draft. KPM was responsible for data curation, conceptualization, 14. Peng Y, Qin F, Kong W, Ge Y, Nie F, Cichocki A (2021) Gfil: a unified reviewing, and editing of the draft. All authors read and approved the final framework for the importance analysis of features, frequency bands and manuscript. channels in eeg-based emotion recognition. IEEE Trans Cogn Dev Syst 15. Li M, Lu B-L (2009) Emotion classification based on gamma-band EEG. Authors’ information In: 2009 Annual international conference of the IEEE engineering in medi- Pankaj Pandey is PhD scholar in Computer Science and Engineering at Indian cine and biology society, pp. 1223–1226. IEEE Institute of Technology Gandhinagar. Richa Tripathi is a post-doctoral fellow 16. Zheng W-L, Lu B-L (2015) Investigating critical frequency bands and at Center for Advanced Systems Understanding (CASUS), Görlitz, Germany. channels for EEG-based emotion recognition with deep neural networks. Krishna Prasad Miyapuram is associate professor in Cognitive and Brain Sci- IEEE Trans Auton Ment Dev 7(3):162–175 ence at Indian Institute of Technology Gandhinagar. 17. Ray WJ, Cole HW (1985) EEG alpha activity reflects attentional demands, and beta activity reflects emotional and cognitive processes. Science Funding 228(4700):750–752 Not applicable. 18. Pan C, Shi C, Mu H, Li J, Gao X (2020) EEG-based emotion recognition using logistic regression with gaussian kernel and Laplacian prior and Availability of data and materials investigation of critical frequency bands. Appl Sci 10(5):1619 The dataset used for the current study will be made available on request. 19. Klados MA, Frantzidis C, Vivas AB, Papadelis C, Lithari C, Pappas C, Bamidis PD (2009) A framework combining delta event-related oscillations (EROS) and synchronisation effects (ERD/ERS) to study emotional processing. Declarations Comput Intell Neurosci 2009:549419 20. Li X-J, Yang G-H (2016) Graph theory-based pinning synchronization of Ethics approval and consent to participate stochastic complex dynamical networks. IEEE Trans Neural Netw Learn Institute Ethical Committee (IEC) of Indian Institute of Technology, Gandhina- Syst 28(2):427–437 gar approved this study. Prior to conducting experiments, all of the partici- 21. Sun S, Li X, Zhu J, Wang Y, La R, Zhang X, Wei L, Hu B (2019) Graph theory pants provided informed consent. analysis of functional connectivity in major depression disorder with high-density resting state EEG data. IEEE Trans Neural Syst Rehabil Eng Consent for publication 27(3):429–439 This paper has the consent for publication from all authors. 22. Friston KJ (1994) Functional and effective connectivity in neuroimaging: a synthesis. Hum Brain Mapp 2(1–2):56–78 Competing interests The authors declare that they have no competing interests. Pandey et al. Brain Informatics (2022) 9:15 Page 19 of 20 23. Bassett DS, Bullmore E (2006) Small-world brain networks. Neuroscientist 50. Bullmore E, Sporns O (2009) Complex brain networks: graph theoreti- 12(6):512–523 cal analysis of structural and functional systems. Nat Rev Neurosci 24. Bassett DS, Gazzaniga MS (2011) Understanding complexity in the 10(3):186–198 human brain. Trends Cogn Sci 15(5):200–209 51. Sporns O, Chialvo DR, Kaiser M, Hilgetag CC (2004) Organization, 25. van Straaten EC, Stam CJ (2013) Structure out of chaos: functional brain development and function of complex brain networks. Trends Cogn network analysis with EEG, MEG, and functional MRI. Eur Neuropsychop- Sci 8(9):418–425 harmacol 23(1):7–18 52. McMenamin BW, Langeslag SJ, Sirbu M, Padmala S, Pessoa L (2014) 26. Zhang J, Zhao S, Huang W, Hu S (2017) Brain effective connectivity Network organization unfolds over time during periods of anxious analysis from EEG for positive and negative emotion. In: International anticipation. J Neurosci 34(34):11261–11273 Conference on Neural Information Processing, pp. 851–857. Springer 53. Lee Y-Y, Hsieh S (2014) Classifying different emotional states by means 27. Li P, Liu H, Si Y, Li C, Li F, Zhu X, Huang X, Zeng Y, Yao D, Zhang Y et al of EEG-based functional connectivity patterns. PLoS ONE 9(4):95415 (2019) EEG based emotion recognition by combining functional 54. Vecchio F, Miraglia F, Rossini PM (2017) Connectome: graph theory appli- connectivity network and local activations. IEEE Trans Biomed Eng cation in functional brain network architecture. Clin Neurophysiol Pract 66(10):2869–2881 2:206–213 28. Rubinov M, Sporns O (2010) Complex network measures of brain con- 55. Klimesch W (1999) EEG alpha and theta oscillations reflect cogni- nectivity: uses and interpretations. Neuroimage 52(3):1059–1069 tive and memory performance: a review and analysis. Brain Res Rev 29. He Y, Evans A (2010) Graph theoretical modeling of brain connectivity. 29(2–3):169–195 Curr Opin Neurol 23(4):341–350 56. Song X, Hu X, Zhou S, Xu Y, Zhang Y, Yuan Y, Liu Y, Zhu H, Liu W, Gao J-H 30. Pessoa L, McMenamin B (2017) Dynamic networks in the emotional (2015) Association of specific frequency bands of functional MRI signal brain. Neuroscientist 23(4):383–396 oscillations with motor symptoms and depression in parkinson’s disease. 31. Wang J, Zuo X, He Y (2010) Graph-based network analysis of resting- Sci Rep 5(1):1–7 state functional MRI. Front Syst Neurosci 4:16 57. Moran LV, Hong LE (2011) High vs low frequency neural oscillations in 32. Tripathi R, Mukhopadhyay D, Singh CK, Miyapuram KP, Jolad S (2019) schizophrenia. Schizophr Bull 37(4):659–663 Characterization of functional brain networks and emotional centers 58. Bassett DS, Bullmore ET, Meyer-Lindenberg A, Apud JA, Weinberger using the complex networks techniques. In: International conference DR, Coppola R (2009) Cognitive fitness of cost-efficient brain functional on complex networks and their applications, pp. 854–867. Springer networks. Proc Natl Acad Sci 106(28):11747–11752 33. Li X, Song D, Zhang P, Zhang Y, Hou Y, Hu B (2018) Exploring EEG fea- 59. Wei L, Duan X, Zheng C, Wang S, Gao Q, Zhang Z, Lu G, Chen H (2014) tures in cross-subject emotion recognition. Front Neurosci 12:162 Specific frequency bands of amplitude low-frequency oscillation 34. Kaneshiro B, Nguyen DT, Norcia AM, Dmochowski JP, Berger J (2020) encodes personality. Hum Brain Mapp 35(1):331–339 Natural music evokes correlated EEG responses reflecting temporal 60. Hagberg A, Swart P, S Chult D (2008) Exploring network structure, dynam- structure and beat. Neuroimage 214:116559 ics, and function using networkx. Technical report, Los Alamos National 35. Thirumalai M (2001) An introduction to Natya Shastra-gesture in aes- Lab.(LANL), Los Alamos, NM (United States) thetic arts. Lang India 1(6):27–33 61. Breiman L (2001) Random forests. Mach Learn 45(1):5–32 36. Uppal C (2018) Rasa: natyashastra to bollywood. Western indology on 62. Molinaro AM, Simon R, Pfeiffer RM (2005) Prediction error estimation: a Rasa: a purvapaksha, 201–225 comparison of resampling methods. Bioinformatics 21(15):3301–3307 37. Kumar CB (2014) The popularity of the supporting cast in Hindi cin- 63. Ojala M, Garriga GC (2010) Permutation tests for studying classifier perfor- ema. South Asian Popul Cult 12(3):189–198 mance. J Mach Learn Res 11(6) 38. Uppal C (2018) Rasa: natyashastra to bollywood. Western indology on 64. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Rasa: a Purvapaksha, 179–199 Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, 39. Uppal C (2018) Rasa: natyashastra to bollywood. Western indology on Cournapeau D, Brucher M, Perrot M, Duchesnay E (2011) Scikit-learn: Rasa: a Purvapaksha, 89–177 machine learning in Python. J Mach Learn Res 12:2825–2830 40. Beitmen LR (2014) Neuroscience and Hindu aesthetics: a critical 65. Van der Maaten L, Hinton G (2008) Visualizing data using t-sne. J Mach analysis of vs ramachandran’s “science of art”. FIU electronic theses and Learn Res 9(11) dissertations. 1198. https://​digit​alcom​mons.​fiu.​edu/​etd/​1198 66. Lin Y-P, Wang C-H, Jung T-P, Wu T-L, Jeng S-K, Duann J-R, Chen J-H (2010) 41. Ghosh M (1967) The Natyasastra ascribed to Bharata-Muni EEG-based emotion recognition in music listening. IEEE Trans Biomed 42. Chang C-Y, Hsu S-H, Pion-Tonachini L, Jung T-P (2019) Evaluation of Eng 57(7):1798–1806 artifact subspace reconstruction for automatic artifact components 67. Soleymani M, Pantic M, Pun T (2011) Multimodal emotion recognition in removal in multi-channel EEG recordings. IEEE Trans Biomed Eng response to videos. IEEE Trans Affect Comput 3(2):211–223 67(4):1114–1121 68. Zheng W-L, Zhu J-Y, Lu B-L (2017) Identifying stable patterns over time for 43. Vinck M, Oostenveld R, Van Wingerden M, Battaglia F, Pennartz CM emotion recognition from EEG. IEEE Trans Affect Comput 10(3):417–429 (2011) An improved index of phase-synchronization for electro- 69. Müller MM, Keil A, Gruber T, Elbert T (1999) Processing of affective physiological data in the presence of volume-conduction, noise and pictures modulates right-hemispheric gamma band EEG activity. Clin sample-size bias. Neuroimage 55(4):1548–1565 Neurophysiol 110(11):1913–1920 44. Hardmeier M, Hatz F, Bousleiman H, Schindler C, Stam CJ, Fuhr P (2014) 70. Kortelainen J, Väyrynen E, Seppänen T (2015) High-frequency electroen- Reproducibility of functional connectivity and graph measures based cephalographic activity in left temporal area is associated with pleasant on the phase lag index (pli) and weighted phase lag index (WPLI) emotion induced by video clips. Comput Intell Neurosci 2015 derived from high resolution EEG. PLoS ONE 9(10):108648 71. Matsumoto A, Ichikawa Y, Kanayama N, Ohira H, Iidaka T (2006) Gamma 45. Lau TM, Gwin JT, McDowell KG, Ferris DP (2012) Weighted phase lag band activity and its synchronization reflect the dysfunctional emotional index stability as an artifact resistant measure to detect cognitive EEG processing in alexithymic persons. Psychophysiology 43(6):533–540 activity during locomotion. J Neuroeng Rehabil 9(1):1–9 72. Tang Y, Li Y, Wang J, Tong S, Li H, Yan J (2011) Induced gamma activity in 46. Gramfort A, Luessi M, Larson E, Engemann DA, Strohmeier D, Brodbeck eeg represents cognitive control during detecting emotional expres- C, Goj R, Jas M, Brooks T, Parkkonen L et al (2013) Meg and EEG data sions. In: 2011 Annual International Conference of the IEEE Engineering in analysis with MNE-python. Front Neurosci 7:267 Medicine and Biology Society, pp. 1717–1720. IEEE 47. Bassett DS, Meyer-Lindenberg A, Achard S, Duke T, Bullmore E (2006) 73. Güntekin B, Basar E (2007) Emotional face expressions are differentiated Adaptive reconfiguration of fractal small-world human brain functional with brain oscillations. Int J Psychophysiol 64(1):91–100 networks. Proc Natl Acad Sci 103(51):19518–19523 74. Onton JA, Makeig S (2009) High-frequency broadband modulation of 48. Barabási A-L, Albert R (1999) Emergence of scaling in random net- electroencephalographic spectra. Front Hum Neurosci 3:61 works. Science 286(5439):509–512 75. Martini N, Menicucci D, Sebastiani L, Bedini R, Pingitore A, Vanello N, 49. Watts DJ, Strogatz SH (1998) Collective dynamics of ‘small-world’ net- Milanesi M, Landini L, Gemignani A (2012) The dynamics of EEG gamma works. Nature 393(6684):440–442 responses to unpleasant visual stimuli: from local activity to functional connectivity. Neuroimage 60(2):922–932 Pandey et al. Brain Informatics (2022) 9:15 Page 20 of 20 76. Zheng W-L, Dong B-N, Lu B-L (2014) Multimodal emotion recognition using eeg and eye tracking data. In: 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 5040–5043. IEEE 77. Tonoyan Y, Looney D, Mandic DP, Van Hulle MM (2016) Discriminating multiple emotional states from EEG using a data-adaptive, multiscale information-theoretic approach. Int J Neural Syst 26(02):1650005 78. Schubring D, Schupp HT (2021) Emotion and brain oscillations: high arousal is associated with decreases in alpha-and lower beta-band power. Cereb Cortex 31(3):1597–1608 79. Vabalas A, Gowen E, Poliakoff E, Casson AJ (2019) Machine learning algorithm validation with a limited sample size. PLoS ONE 14(11):0224365 80. Russell JA (1980) A circumplex model of affect. J Pers Soc Psychol 39(6):1161 81. Tomkins SS (1962) Affect, imagery, consciousness: cognition: duplication and transformation of information, vol 4. Springer, Berlin 82. Qi C, Li M, Wang Q, Zhang H, Xing J, Gao Z, Zhang H (2018) Facial expres- sions recognition based on cognition and mapped binary patterns. IEEE Access 6:18795–18803 83. Feldman Barrett L, Russell JA (1998) Independence and bipolarity in the structure of current affect. J Pers Soc Psychol 74(4):967 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub- lished maps and institutional affiliations.