Detecting Mind-Wandering with Machine Learning
Jin, Christina
DOI:
10.33612/diss.171835555
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date: 2021
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Jin, C. (2021). Detecting Mind-Wandering with Machine Learning: Discovering the Neural Correlates of Mind-Wandering Through Generalizable Machine Learning Classifiers with EEG. University of Groningen. https://doi.org/10.33612/diss.171835555
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
1
INTRODUCTION
Have you ever wondered why there are moments that we find ourselves end up thinking about something else, for example, holiday plans, novel plots, or some impressive life events, when our attentional focus was initially on our job? This divergence between our mental activity and the current environment refers to a mental phenomenon called mind-wandering, absence of mind, inattention, or daydreaming. In this thesis, we use mind-wandering as the term to describe these phenomena, and investigate its origin and neural correlates.
The introduction will start with the definition of mind-wandering, and contrast it with similar phenomena such as distraction. Cognitive models of mind-wandering are reviewed so that the readers can understand how mind-wandering is supported by basic cognitive functions such as attention. We reviewed a substantial amount of physiological and neural evidence indicating mind-wandering is supported by a complex network of brain areas. Two key neural processes can be derived from the studies done so far – mind-wandering involves both sensory decoupling and memory retrieval processes. Finally, we reviewed multiple factors that can influence the occurrence of mind-wandering, such as working memory, task loads, meta-awareness, personal concerns, etc.
To better understand how mind-wandering arises, it would be helpful to know moment-by-moment whether someone is mind-wandering. To that end, the general goal of this thesis is to develop methods to track mind-wandering using electroencephalography (EEG) data by training machine learning classifiers. Importantly, these classifiers should eventually be able to generalize across tasks (in the same study), participants, and studies (experiments). Apart from providing us a moment-by-moment assessment of the mind-wandering state of an individual, efficiently trained mind-wandering classifiers also give us clues about the brain areas that are crucial for mind-wandering to occur. Three independent studies included in the thesis demonstrate how such classifiers can be developed.
1 . 1 C O N C E P T O F M I N D - W A N D E R I N G
1.1.1 Heterogeneity and definition in the current research
Mind-wandering is a common mental phenomenon: it has been estimated to account for 46.9% of the waking time (Killingsworth & Gilbert, 2010). The content of mind-wandering is heterogeneous, as many dimensions can be used to classify thoughts that occur during mind-wandering into different subtypes, e.g., their temporal orientation (to the past/future/present), modality (in a verbal or auditory form), self-reference, stickiness, etc. (Stawarczyk et al., 2011; van Vugt & Broers, 2016; Wang et al., 2018). While the range of the mind-wandering concept is under vigorous debate (Christoff et al., 2018; Seli, Kane, et al., 2018) the research community did reach consensus that mind-wandering is always task-irrelevant – otherwise it is part of the current task.
From the generation perspective, mind-wandering has been defined as a independent mode of thinking that occurs spontaneously. This stimulus-independent constraint rules out another form of task-unrelated thoughts (TUTs): thoughts that are triggered by distractors in the environment (Robison & Unsworth, 2015), referred to as ‘distraction’. The spontaneous characteristic further frames mind-wandering as out of deliberate cognitive control in the current thesis (for discussion about intentional/deliberate mind-wandering, see Seli, Ralph, et al., 2017). Thus, we can conclude that mind-wandering is a thought generation process with three characteristics:
• task-irrelevant
• spontaneous, automatic, without deliberate control • stimulus-independent and self-generated
1.1.2 (Mal)functions of mind-wandering
Mind-wandering is thought to be one of the main causes of decreased efficiency in everyday life situations. For example, road crashes are significantly related to exhibiting high trait mind-wandering (i.e., having a personality that shows a predisposition to mind-wander, Gil-Jardine et al., 2017). Research also found that temporal perception is disturbed while mind-wandering (Terhune et al., 2017), indicating that jobs requiring precise temporal judgment could be harmed by mind-wandering.
C O G N I T I V E F R A M E W O R K S O F M I N D - W A N D E R I N G 3
Yet, the common occurrence and the spontaneous nature of mind-wandering indicates its important role in humans’ mental world. In general, it is known that thinking activity helps to consolidate memory and to extract general knowledge from recent experiences (Wamsley, 2013). As a result, mind-wandering may contribute to problem-solving and creativity (Smallwood and Schooler, 2015). Researchers found that a moderate amount of mind-wandering is beneficial in encoding complex virtual scenes (Blondé et al., 2020). In studies of experience sampling that examine the occurrence of thoughts in the course of everyday life, mind-wandering has been shown to affect mood positively, especially when the thought content was pleasant (Welz et al., 2018).
Some subtypes of wandering, for example, rumination (sticky mind-wandering, i.e., being difficult to disengaging from mind-wandering) might be more maladaptive than non-sticky mind-wandering and is often related to some psychological pathology like depression (van Vugt & Broers, 2016). However, others are of the opinion that rumination differs from mind-wandering. The two processes share various characteristics: rumination and mind-wandering both are automatically generated, irrelevant to the task, and independent of the outside stimuli. Yet, mind-wandering allows thoughts to move freely between topics, while rumination contains too many automatic constraints to fix thoughts on a certain topic (Christoff et al., 2016). Given the current definition, we do not discriminate rumination from mind-wandering and regard it as a particularly sticky form of mind-wandering.
1 . 2 C O G N I T I V E F R A M E W O R K S O F M I N D -W A N D E R I N G
There are multiple cognitive frameworks that incorporate mind-wandering. First, mind-wandering is incorporated into the executive models of attention, in which attention functions as a resource that supports other cognitive functions such as reading, memory, and encoding stimuli. Mind-wandering competes with the attention required for performing the primary task. While standard attentional executive models assume controlled processes are directed towards a specific goal (e.g., the ongoing task), the goal of mind-wandering is initiated automatically, oriented to unsolved problems like salient personal goals (Klinger, 1999; Smallwood & Schooler, 2006).
Another framework for explaining wandering focuses on the fact that mind-wandering reflects a mode of thinking that is decoupled from the external environment (Smallwood & Schooler, 2015). This is supported by substantial neural and physiological evidence (Baird et al., 2014; Baldwin et al., 2017; Barron et al., 2011; Broadway et al., 2015; Gouraud et al., 2018; Huijser et al., 2018; Kam et al., 2011; Kam et al., 2012; Konishi et al., 2017; Smallwood et al., 2008; Unsworth & Robison, 2016; Xu et al., 2018). The term ‘decoupling’ means that the cortex shows evidence for reduced processing of incoming stimuli when mind-wandering occurs. Decoupling happens both in the earlier perceptual processing stage and the later cognitive processing stages (Kam & Handy, 2013). Pupillometric evidence also supports that mind-wandering withdraws attention from the current task (Smallwood et al., 2011). In the section ‘Behavioral and neural evidence’, we will review this evidence in more detail.
1 . 3 B E H A V I O R A L A N D N E U R A L E V I D E N C E F O R M I N D - W A N D E R I N G
1.3.1 Experimental paradigms and associated behavioral patterns
One method that has been employed frequently for measuring participants’ momentary attention states – including mind-wandering – is the experience sampling method. Participants are required to give a self-report of their current attentional state or thinking content when presented with probe questions on the screen (see Weinstein, 2018, for a summary of probe questions). Often the probes are randomly embedded in a trial sequence, while the time between two consecutive probes is jittered to prevent anticipation. The main drawback of this method is that the trials are encoded by participants’ self-reports, which imposes the risks of noisy labelling considering the reliability of the self-reports and inter-rater consistency. In fact, to preview, in all three experiments of this thesis, this pitfall of labelling using self-reports may have limited the classifiers’ performance. Besides, probes seem to have a framing effect. Asking participants if they are currently mind-wandering would cause them to report more mind-wandering than asking them if they are focusing on the task (Weinstein et al., 2018). However, given the spontaneous and generated nature of mind-wandering, using self-reports to obtain data is almost inevitable (see the discussion of Chapter 4). Mind-wandering can be triggered and measured in multiple experimental paradigms. One commonly employed paradigm is the
sustained-attention-to-B E H A V I O R A L A N D N E U R A L E V I D E N C E F O R M I N D - W A N D E R I N G 5
response task (SART). This paradigm is an adapted form of a go-nogo task in which participants have to respond to frequently-presented go stimuli but withhold their response to rare nogo stimuli. When participants experience attentional lapses, they will easily miss the nogo stimuli and make a commission error (Manly et al., 1999). Cheyne et al. (2006) further verified the SART and indicated that the nogo commission error is caused by a failure of sustained attention, which was visible in relatively fast response times before participants made a nogo error.
While a commission error in the SART is an obvious indicator of mind-wandering, it was later also revealed that response time variability also positively correlated with how frequently an individual engaged in task-unrelated thoughts (TUTs) in the SART (McVay & Kane, 2009). It is likely that this increase in response time variability reflects behavioral variation due to the attentional disengagement from the current task. This stimulated the development of another paradigm – the metronome response task (MRT), which yields a more sensitive measure of response time variability. In the MRT, participants were instructed to respond synchronously by pressing buttons with the rhythmic presentation of tones. Researchers found that participants had larger response variability before probe-caught mind-wandering compared to a probe-probe-caught on-task state (Seli et al., 2013). The metronome counting task (MCT) was an adaptation of the MRT, in which the rhythmic tones were replaced by a specified counting goal (e.g., from 1 to 20) and participants pressed a button as soon as they reached the goal (e.g., 20). Similar to the MRT, the response time variability served as the main measure. Researchers found that not only the response time variability became larger before an incorrect than a correct response, but also that it reached the maximum before a self-caught response failure. Their findings indicate that the MRT can also be used to measure the meta-awareness level during mind-wandering (Anderson & Farb, 2020).
Finally, in a reading task, researchers had participants report the number of the words re-read following a self-caught or probe-caught mind-wandering. They found re-reading was frequent when participants disengaged from mind-wandering and re-focused on the task (Varao-Sousa et al., 2017). This could be a direction to examine when studying mind-wandering in reading tasks by using eye-tracking.
1.3.2 Eye-tracking studies
Ocular activity can give information about the attentional focus and the focusing level of the individual. There are multiple relevant measures that can be derived from eye-tracking data. For example, pupil diameter can be used as an indication of the current attention level. Research has shown that participants had smaller pupil size in a two-second pre-trial interval when they were mind-wandering in comparison to when they were on-task (Unsworth & Robison, 2016) as well as during the trial (Huijser et al., 2018; Unsworth & Robison, 2016). Pelagatti et al. (2018) found that pupil size became larger in the 2-6 seconds following words or phrases that triggered mind-wandering compared to other emotional verbal cues which did not trigger mind-wandering. In addition, they measured a smaller pupil size around 4-6 seconds preceding a self-report of mind-wandering compared to on-task when the pupillometry was time-locked to the presentation of the probes. In addition to pupil size, Krasich et al. (2018) found that mind-wandering was associated with fewer but longer fixations, greater fixation dispersion and more frequent eyeblinks, forming the gaze-based signature of mind-wandering.
1.3.3 EEG markers
Mind-wandering has been associated with specific changes in both event-related and frequency-based components of EEG.
Smallwood et al. (2008) found a reduced P3 at the parietal region while participants were mind-wandering compared to when they were on-task. Barron et al. (2011) further indicated not only the posterior P3 (P3b) was reduced but also the frontal P3 (P3a) was inhibited during wandering, suggesting mind-wandering was associated with a general decoupling from the processing of the external stimuli. Kam et al. (2011) found further evidence for perceptual decoupling using both visual and auditory stimuli and demonstrated that evidence consistent with decoupling can be observed as early as 100ms after stimulus onset. A similar reduction in ERP components during mind-wandering was found in later studies (e.g., P1, Baird et al., 2014; N1, Broadway et al., 2015; P3a, Baldwin et al., 2017). Xu et al. (2018) found P2 and slow waves at the parietal regions to be attenuated during periods before probe-caught mind-wandering, suggesting that not only the initial stages of perceptual processing, but also the deep processing of stimuli is reduced when mind-wandering. Kam et al. (2012) further found reduced error-related negativity (ERN) during mind-wandering. This may mean that mind-wandering is associated with a decrease of cognitive monitoring.
B E H A V I O R A L A N D N E U R A L E V I D E N C E F O R M I N D - W A N D E R I N G 7
Using time-frequency analysis, Baldwin et al. (2017) found that mind-wandering was associated with a power increase in the alpha band (8.78Hz) during a simulated driving task. Consistent with that finding, Compton et al. (2019) found that participants exhibited increased alpha (8-12Hz) power before they gave a self-report of mind-wandering compared to before they reported being on-task. Ros et al. (2013) trained participants to voluntarily reduce their alpha (8-12Hz) power using neurofeedback and uses the same oddball task for pre- /post-training measures. They found that the amount of alpha power change during the neurofeedback training was positively related to their resting state alpha power change before and after the training. They also found that the resting state alpha power change was positively related to the probe-caught mind-wandering change during an oddball task. This suggests that alpha power reduction has a direct effect on participants’ mind-wandering in the post-training cognitive task. Braboszcz and Delorme (2011) found EEG activity increases in the theta (4-7Hz) and delta (2-3.5Hz) bands and decrease in alpha (9-11Hz) and beta (15-30Hz) bands at posterior sites in mind-wandering. A possible explanation for this discrepancy is that the authors of this study used focusing on the breath rather than focusing on a cognitive task as a comparison with mind-wandering (Baldwin et al., 2017; Compton et al., 2019; Ros et al., 2013). Since focusing on breath indicates attentional shift from the outside to the inner experience, alpha enhancement reflects the reduced sensory processing in this process.
1.3.4 Brain regions and networks
While EEG underpins the neural evidence of the decoupling process underlying wandering, a more precise mapping of the cortical corelates of mind-wandering comes from studies with cortical thickness analysis, functional neural imaging, cortical lesions and neuromodulation. The studies will be summarized in the following four categories: the default mode network (DMN) structure, the non-DMN structure, the within-DMN connectivity and the between-network connectivity.
DMN structure
The default mode network (DMN) was originally defined as a group of regions deactivated during cognitive tasks requiring externally-oriented attention (Raichle et al., 2001), presumably accounting for the spontaneous mental processes
appearing during the resting state (Christoff et al, 2004). The DMN has three functionally distinct subsystems. The first subsystem serves as the hubs of the neural connectivity underlying internally-oriented cognition, including the anterior mPFC (amPFC), posterior cingulate cortex (PCC), and posterior inferior parietal lobule (pIPL). The second subsystem centered around the medial temporal lobe (MTL), which includes important memory-related brain areas including the hippocampus and parahippocampal cortex, stressing that memory retrieval processes are a key to identifying mind-wandering at a cognitive level. This subsystem also includes other ventral structures – the retrosplenial cortex (RSC), the ventral mPFC (vmPFC), and the pIPL. The third subsystem covers dorsal structures, including the dorsomedial PFC (dmPFC), the lateral temporal cortex (LTC), temporo-polar cortex (TPC), and parts of the inferior frontal gyrus (IFG). This subsystem is involved in a variety of brain functions like mentalizing, conceptual processing, and emotional processing (reviewed by Christoff et al., 2016).
Bonnelle et al. (2011) found that patients suffering from traumatic brain injury had impaired sustained attention and the degree of impairment was associated with an increase in DMN activation, especially within the precuneus and PCC. Mittner et al. (2014) showed that increased activity within the DMN can be one of the neural signatures to train an efficient mind-wandering classifier; the signal is strong enough to predict the occurrence of mind-wandering on a single-trial level. Durantin et al. (2015) found significant activation of the mPFC during mind-wandering using functional near-infrared spectroscopy (fNIRs) and confirmed that the activity of the DMN regions can be used to train a machine learning classifier of mind-wandering. Bernhardt et al. (2014) found an increased thickness of the medial prefrontal cortex (mPFC) and anterior/midcingulate cortex (ACC/MCC) is associated with a higher tendency to engage in task-unrelated thoughts under low-demanding conditions but not under high-demanding conditions (i.e., in a choice reaction task but not a working memory task). This implies that these brain structures are involved in mind-wandering, although their precise role is not yet clear.
Another way of studying mind-wandering is through lesion studies or by manipulating the brain areas involved in this process. Bertossi and Ciaramelli (2016) observed that patients with a lesion to the vmPFC showed reduced mind-wandering rates in experiments and less daydreaming in their daily life. In another lesion study, researchers revealed that reduced mind-wandering was associated
B E H A V I O R A L A N D N E U R A L E V I D E N C E F O R M I N D - W A N D E R I N G 9
with lesions of the mPFC, IPL, and IFG (Philippi et al., 2021). McCormick et al. (2018) found that hippocampal damage did not reduce the daily mind-wandering amount but influenced the contents of the mind-wandering thoughts to be more present-oriented with more verbally-mediated semantic knowledge, instead of the autobiographical material that is the source of mind-wandering in healthy individuals. Finally, Bertossi et al. (2017) showed that transcranial direct current stimulation (tDCS) over mPFC reduced the mind-wandering propensity— surprisingly this was only true for male participants, but not for females.
Non-DMN structures
Mind wandering cannot only be observed in the DMN. Christoff et al. (2004) found that mind-wandering and on-task thoughts depend both on the activation of regions including visual areas, medial temporal lobe, and lateral cortical association areas. This indicates that mind-wandering and on-task states might not strictly be distinguishable. In fact, mind-wandering may require some similar cognitive strategies as a cognitive task, such as organizing information into a coherent narrative and shielding that narrative from other tasks. For example, Christoff et al. (2009) identified the involvement of both the default mode network (DMN) and the executive network associated in mind-wandering, and these networks are also involved in various task-related processes such as prospection and planning. They found mind-wandering to be associated with greater activation in the temporopolar cortex, parahippocampus, rostrolateral prefrontal cortex, parietal, and visual cortex areas.
Researchers found that more mind-wandering was reported during automatic compared to manual driving, and this was accompanied by a decreasing OxyHb concentration in right temporal-parietal and occipital regions (Hidalgo-Muñoz, et al., 2019). The medial orbitofrontal cortex (mOFC) was shown to be able to encode the affective tone of the TUTs (Tusche et al., 2014).
Within-DMN connectivity
Godwin et al. (2017) showed that resting-state DMN connectivity was positively related to participants’ trait mind-wandering. The within-DMN connectivity pattern also gives insights into the subtypes of mind-wandering. Smallwood et al.
(2016) found that the variations in connectivity from the medial and lateral temporal lobe memory systems to the PCC differentiated the thinking content of mind-wandering. Similarly, Karapanagiotidis et al. (2017) used the right hippocampus as the seed and revealed that increased functional coupling to the anterior regions of the DMN is associated with increased mental time travel during tasking.
Between-network connectivity
Finally, we can look at between-network connectivity, instead of connectivity within the DMN. Fox et al. (2015) conducted a meta-analysis with 24 neuroimaging studies, which not only confirmed the key role of connectivity within the DMN structures like mPFC, PCC, MTL, and bilateral IPL in mind-wandering but also indicated the coupling between the DMN and the frontoparietal control network. Zhang et al. (2019) found that the DMN was decoupled more from the medial visual cortex in participants who mind-wandered more. The connectivity between the core of the DMN and the primary visual cortex is correlated with the amount of detail of the thinking content (Ho et al., 2019). Negative functional connectivity between the DMN-salience network and the DMN-executive network and positive connectivity between the salience network and the executive network was observed more frequently during on-task than mind-wandering (Denkova et al., 2019). Mittner and colleagues (2014) showed that the connectivity reduced more between the DMN and the anti-correlated network (task positive-network) during mind-wandering compared to being on-task.
1 . 4 C O N T R I B U T I N G F A C T O R S
Multiple factors have been found to be associated with the occurrence of mind-wandering. We will briefly describe each of these below.
Working memory
Researchers showed that working memory capacity was negatively correlated with task errors, response time variance and TUTs. This is explained by a higher
C O N T R I B U T I N G F A C T O R S 11
level of cognitive control supported by the working memory capacity to suppress mind-wandering and keep attention on the task (McVay and Kane, 2009). Corroborating these results, Randall et al. (2019) found an interaction between working memory and task demands when predicting mind-wandering. Specifically, their research showed that higher working memory capacity prevented individuals from engaging in mind-wandering when the task demands became extremely high. In a reading task, having a higher working memory capacity has been shown to limit the occurrence of mind-wandering and to benefit the reading performance (Soemer & Schiefele, 2020).
Task load
Vannucci et al. (2019) showed in a visual detection task that the amount of mind-wandering reduced when task loads increased. Randall et al. (2019) further suggested mind-wandering was most frequent in extreme task conditions – either very low or very high load – with its occurrence shaped like a V-curve. They also indicated an interaction between task demands and working memory capacity, such that higher working memory capacity prevents the increase of mind-wandering in an extremely difficult task. This was interpreted using the concept of the task’s resource sensitivity. Performance on tasks higher in resource sensitivity (e.g., medium load) are more likely to be modulated by the amount of attentional resources available – higher performance with higher attentional allocation, so participants are more likely to spend mental effort on such tasks. On the contrary, performance of tasks low in resource sensitivity (e.g., very low or very high load) is less likely to be interfered with when adjusting the mental investment so participants are less likely to spend more attentional resources. The ‘saved’ resources allow the generation of mind-wandering. Given the task load is relative to the individual cognitive capacity required in the task (e.g., working memory in a math solving task; high load for individuals with low working memory might be medium for those with high working memory), it explains the interaction between task load and working memory in mind-wandering rate (Randall et al., 2019).
Krimsky et al. (2017) found a time by task load interaction on mind-wandering. Participants showed more mind-wandering in the low load than the high load condition at the beginning of the task. When time passed, they exhibited a more
rapid mind-wandering increase in the high load than the low load condition and reported more mind-wandering in the high load condition at the end of the task.
Time
Time spent on the task is associated with increased mind-wandering (Krimsky et al., 2017). This might be explained by the reduced alertness or vigilance across time, manifesting as a sustained attention decline (Oken et al., 2006). Robison and Unsworth (2017) showed that alertness uniquely predicted mind-wandering independent of working memory capacity. In Chapter 3, we also discriminated the vigilance decrement from mind-wandering.
Meta-awareness
In their review, Schooler et al. (2011) suggested a role for meta-awareness in mind-wandering. Meta-awareness refers to the explicit noticing of the content of experience. In general, when one is aware that mind-wandering is occurring, one can decide to continue it or abort it (Schooler et al., 2001). Evidence for this comes from research that distinguishes so-called ‘self-caught’ from so-called ‘probe-caught’ mind-wandering. Probe-caught mind-wandering is mind-wandering that is identified by inserting regular self-report prompts in the task, while self-caught mind-wandering is identified when participants press a pre-specified key when they notice they are mind-wandering. Self-caught mind-wandering places a higher demand on an individual’s capacity for meta-awareness (Seli, Smilek, et al., 2018)—when participants do not notice they are mind-wandering, they cannot report it. Interestingly, the study Allen et al. (2013) showed that it was not the overall frequency of wandering but the frequency of shifting between mind-wandering and on-task states related to the meta-cognition.
Personal concerns
Using personal concern related words, McVay and Kane (2013) showed that the mind-wandering rate could be raised by 3-4%. Another study by Jordano and Touron (2017) found a similar concern priming effect on the self-reported mind-wandering rate.
C O N T R I B U T I N G F A C T O R S 13
Aging
Aging is associated with a reduction of mind-wandering. McVay et al. (2013) found that older participants reported fewer TUTs compared to the younger adults across multiple experiments. Staub et al. (2014) suggested the reduction of mind-wandering in the aged group might be explained by their higher motivation and increased cognitive control when performing tasks. Martinon et al. (2019) showed reduced connectivity between the left anterior temporal lobe and the prefrontal regions of the DMN, indicating a neural correlate of the aging effect on mind-wandering.
Sleepiness
A higher level of daytime sleepiness and sleep-related disturbances not only influences the tendency to experience mind-wandering in both daily life and the lab experiments but also predicts the SART performance independent of mind-wandering (Stawarczyk & D'Argembeau, 2016). Sleep disturbances were found to predict participants’ mind-wandering scores on the questionnaire (Marcusson-Clavertz et al., 2019), as well as the reported amount of the task-unrelated thoughts caught by probes (Marcusson-Clavertz et al., 2020).
Mood and needs
Koelsch et al. (2019) found that the affective tone of music influences the valence of mind-wandering in the same direction (i.e., music with a positive affective tone increased the amount of positively-valenced mind-wandering), suggesting its therapeutic values in treating maladaptive forms of mind-wandering, such as depressive ruminations. Rummel et al. (2017) studied how physical needs influenced mind-wandering. They found that hunger but not sexual arousal induced more mind-wandering.
Stimulus properties
The properties of task-related external stimuli influence mind-wandering as well, even though mind wandering is defined specifically as internally-directed spontaneous self-generated thoughts. Faber et al. (2018) found that the complexity
of the audiovisual features of a presented video predicted the likelihood of mind-wandering. Mind-wandering was less likely to occur when there were more event changes. In Choi et al. (2017), the visual or auditory modality of mind-wandering was shown to be consistent with the visual or auditory modality of the stimuli.
Personality
An individual’s personality influences their mind-wandering experience. The frequency of mind-wandering has been shown to be positively related to levels of narcissism, in other words, narcissistic people tend to mind-wander more. Moreover, individuals with a higher level of narcissism tend to exhibit a type of mind-wandering that is more self-related, positive and future-oriented (Kanske et al., 2017). Also, neurotic individuals tend to report more mind-wandering and have poorer attentional control (Robinson, Gath, et al., 2017).
Clinical pathologies
Many clinical pathologies are linked to mind-wandering, especially its maladaptive forms. Such pathologies include attentional deficit hyperactivity disorder (ADHD), depression, obsessive-compulsive disorder (OCD), unsolved-trauma related symptoms, etc. It has been found that an ADHD-inattention group mind-wanders more in both their daily life and in laboratory experiments. When in a dysphoric mood, they are more likely to engage in intrusive ruminations (Jonkman et al., 2017). Mind-wandering was found to induce emotional lability, while both mind-wandering and emotional lability are known to contribute to the ADHD symptom severity (Helfer et al., 2019). Meanwhile, a higher rate of mind-wandering is also associated with more severe OCD symptoms (Seli, Risko, et al., 2017). Rumination happens when negative memories are retrieved repeatedly for over-processing, contributing to the depression symptoms (van Vugt et al., 2018). Unsolved-representations of individual childhood trauma predicted the negative valence of mind-wandering (Marcusson-Clavertz et al., 2017).
T H I S T H E S I S 15
Researchers have explored ways to interfere with our spontaneous, self-generated thinking process. This is achieved by controlling some of the contributing factors (e.g., behavioral training using meditation to enhance self-awareness), or performing neuromodulations using the neural markers of mind-wandering (e.g., neurofeedback to modulate the magnitude of alpha oscillations). Strengthening cognitive control and meta-awareness is regarded as a main direction to regulate mind-wandering in our daily lives and constrain its negative impact during working (Smallwood et al., 2015). To achieve this, mindfulness interventions have been suggested to be effective (Mrazek et al., 2013). Expert meditation practitioners reported less mind-wandering with reduced depth and reduced theta and alpha rhythms compared to non-expert practitioners (Brandmeyer & Delorme, 2016). Bastian et al. (2017) found participants with inner speech suppression had less awareness of their mind-wandering. After priming them with verbal materials to activate their inner speech, they gained enhanced retrospective awareness of mind-wandering. Ros et al. (2013) used neurofeedback training in suppressing their alpha power. A successful reduction of alpha power led to a reduction of mind-wandering occurrences during a cognitive task after training compared to before.
1 . 6 T H I S T H E S I S
The previous sections have identified many different neural correlates of mind-wandering, as well as many factors that influence mind-wandering. However, almost all these studies depend on probes to detect mind wandering. Unfortunately, it is less clear how we could track mind-wandering from moment to moment. If we were able to do that, it would allow us to identify much more precisely how different factors influence mind-wandering. The research into the neural and physiological correlates of mind-wandering is particularly important considering its ‘hidden’ nature—it cannot be observed directly but is only identifiable through self-report. What’s more, an efficient neural maker of mind-wandering is not only usable as a measure of treatment effectiveness but could fulfill training and therapeutical purposes when combined with neuromodulation methods like neurofeedback.
The research presented in this thesis consists of three experiments with one common purpose – training effective mind-wandering classifiers with EEG using machine learning models. EEG was selected as our main data source because of
its high temporal resolution—it operates on a millisecond scale. As our mental states shift between mind-wandering and on-task constantly, the temporal detail of EEG allows it to capture such dynamics more precisely than other measures of neural activity such as magnetic resonance neuroimaging. However, there is a temporal-spatial precision tradeoff here when proposing EEG as the methodological basis. Scalp EEG is not good at localizing the relevant brain regions especially when the neural activity is related to deeper cortical structures like the DMN, salience network, or the limbic system – which we know are involved in mind-wandering. In Chapter 3, you will see how we tackled this dilemma with source localization techniques – to trace the signal back to its most likely cortical generator with mathematical models. Other considerations of choosing EEG are financial cost and portability. We hope the research results can inspire applications in the domain of neurofeedback. Both cost and portability are vital for such applications.
More specifically, Chapter 2 makes the first attempt at predicting mind-wandering with EEG, using several EEG candidate markers derived from previous findings, including ERP, band power and channels connectivity. These features were used to train a support vector machine (SVM) classifier to predict the individual mind-wandering reports (intra-subjective modelling). We included two cognitive tasks and used the classifier based on the data of one task to predict the data from the other. Interestingly, the classifiers we developed were able to generalize across tasks (across-task predictions).
In Chapter 3, we used machine learning classifiers similar to those used in Chapter 2 to distinguish vigilance decrement and low task demands from mind‐wandering. In this study, we tried to dissociate three relevant mental phenomena – mind-wandering, low vigilance, and low task demands by training SVM classifiers with source-localized EEG signals. Previous findings showed positive correlations between the three (see also section ‘contributing factors’), and they have similar impacts on task performance and the amplitude of parietal-occipital alpha power. We therefore examined whether a classifier originally trained to discriminate vigilance or task demands was also be able to predict mind-wandering. When this generalization across the three phenomena failed, we used the source-localized alpha oscillations as a feature to show that each of the three phenomena had separate brain locations, with some overlaps between mind-wandering and low task demands. In this study, we also showed that similar accuracy of
mind-T H I S mind-T H E S I S 17
wandering prediction could be achieved from a between-subject classifier, compared to the within-subject classifier in Chapter 2.
In Chapter 4, we examined whether we could further improve mind-wandering predictions by relying on deep neural networks. Specifically, we relied on convolutional neural network (CNN), which is a powerful form of neural networks to handle larger data sizes with no assumptions on features. We fed the network with raw EEG, single-trial ERP patterns from all channels, and power/phase coherence information from all bands and all channels. To further verify the classifier’s generalizability, we trained the network on the data from Chapter 2 and tested the same classifier on the data from Chapter 3 –independent experiments with different participants and different tasks.
In Chapter 5, I will review and summarize the main conclusions from the three studies. I will review the current most popular hypothetical models of mind-wandering and present my opinion based on the data presented in this thesis. Last but not least, I will put forward possible directions to explore in mind-wandering research, in relation to the findings presented in this thesis.
