Abstract
Machine Learning and Altered States of Consciousness: What Will We Think of Next?
Gloria Y. Yeh, MD MPH, Osher Center for Integrative Health, Harvard Medical School and Brigham and Women’s Hospital, Boston, MA.
Machine learning is a subfield of artificial intelligence (AI), which is broadly defined as the capability of a machine to imitate intelligent human behavior. While traditional machine learning has been around for decades, more recent advances in complex neural networks and deep learning have allowed further sophistication in modeling, pattern recognition, and prediction analyses of complex phenomena. Given this power, the applications of machine learning technology appear endless. Nearly every industry is exploring new ways to leverage machine learning, and research in mind–body medicine is no exception. Within academia, this ubiquitous fascination with the promise of AI has created new dialogues around expanded applications to probe novel research questions that we could not have considered before.
One such area is that of human consciousness, or in this example, understanding the neural correlates of altered states of consciousness (ASC). Historians have reported human interest in tapping into ASC perhaps as early as 30,000 years ago, from use of mind-altering plants, alcohol or excessive dancing to achieve ecstatic states, early religious use like in the rites of Dionysus, or various shamanic traditions for therapeutic purposes. In modern day psychiatry, there has been renewed interest in this aspect of meditation, hypnosis, and now psychedelic drugs.
In a recent study by Moujaes et al., we see an interesting example in neuroscience of applying machine learning to functional MRI data in a way that pushes us into this new horizon. While applying machine learning to analyze neuroimaging data in the mind–body field is not new, the application of more advanced methods to understand and predict states of consciousness shift induced by different pharmacologic and nonpharmacologic methods is quite novel.
Prior research has suggested that on a phenomenological level, ASC induced by psychedelics and meditation are similar. Not unlike with advanced meditative practice, psychedelic experiences are often described as deeply personal, spiritual, or transformative. In fact, this overlap has led to further work suggesting that one can support or enhance the other and is the rationale for the current renaissance of psychedelics in psychiatry and its potential intersection with mind–body medicine. Despite this interest, there is a lack of rigorous data that fully describe the neural correlates of these ASC. In this study, investigators sought to do just that, specifically, to investigate what are the common and/or distinct neural correlates involved, and what is the predictive value of the functional connectivity data in distinguishing ASC from psychedelics versus mind–body therapies in a whole-brain data-driven approach.
Neuroimaging data from pharmacologic induction was acquired from a crossover RCT where healthy volunteers underwent a session of psychedelic-induced ASC with psilocybin (dose = 0.2 mg/kg; n = 23), or LSD (dose = 100 mg; n = 25) followed by resting state functional connectivity MRI. A placebo session occurred 2 weeks later. Data from nonpharmacological induction was acquired from mind–body experts in hypnosis (induction of standardized hypnotic state called “Esdaile”; n = 30) and Buddhist meditation (mindful open awareness; n = 29) while they practiced in the scanner. Each served as their own control with separate scans during a nonhypnosis or nonmeditation condition. Both traditional statistical methods and machine learning were utilized to analyze region-of-interest to region-of-interest connectivity matrices. For machine learning, investigators utilized binary support vector machine classification and multilevel Gaussian process classification with a 5000 times permutation test for accuracy in order to determine whether the ASC condition could be predicted based on these matrices. The findings revealed that pharmacologic and nonpharmacologic methods of inducing ASCs exhibited distinct connectivity patterns that were predictive at the individual level. Psilocybin- and LSD-induced ASC showed more overlap in neural correlates of functional connectivity when compared with each other, while hypnosis and meditation were more distinct. Machine learning methods were able to predict pharmacologic- versus nonpharmacologic-induced ASC with total accuracy of 85%, and distinguish hypnosis- and meditation-induced ASC with accuracy of 66%. Interestingly, there was lack of a common neural network in all four ASC groups, a surprising finding given the observed overlap in phenomenology. One of the most discriminative pathways was the connection between V1 (primary visual cortex) and the somatomotor network, thought to be related to the different ways visual perceptions are altered in the four methods. For example, in hypnosis, visual perceptual alterations may be the result of top-down mental representations translated into perceptual states, whereas there is not such a focus on visual imagery in open awareness meditation; in psychedelics, visual perceptual alterations are the result of internally driven excitation of the visual pathway.
This study represents just one small glimmer where AI and machine learning might contribute to the vast universe of human understanding. As we enter a renaissance of psychedelics in medicine and probe the relationship with meditation, have advanced machine learning methods enabled new horizons of understanding? Perhaps it may add insight to age old questions about the nature of consciousness? At the least, these investigations may help us better understand possible mechanistic pathways of ASC and how to leverage/combine our mind–body therapies for specific individuals and specific therapeutic states of mind within and beyond psychiatry. For consciousness, it helps to move the phenomenon out from the sole purview of philosophers, theologians, anthropologists or even psychologists, and opens the door for neuroscience to complement prior scholars in new data-driven ways not possible before.
Being “Scared Stiff”: Using Machine Learning to Understand How Body Movement and Posture Encode Acute Psychosocial Stress
Peter M. Wayne, Osher Center for Integrative Health, Harvard Medical School and Brigham and Women’s Hospital, Boston, MA.
A fundamental premise underlying mind–body research and its therapeutic application is that human patterns of physical movement can influence mental states, and conversely, mental states can influence somatic phenomena. At the clinical and epidemiological level, this mind–body dialectic underlies the established observations that both the total amount and diurnal pattern of physical activity (and sedentariness) correlate with risk of depression, and conversely, being depressed or anxious can alter exercise behavior. However, there remains a paucity of more granular research evaluating how acute mental states are reflected in patterns of body movement and posture. Such knowledge could provide novel biomarkers for noninvasively tracking mental states, inform mechanisms of cross-system mind–body interactions, and lay the foundation for mind–body movement training strategies that can more effectively mitigate the impact of acute emotional stressors.
Robert Richer and colleagues based in Friedrich-Alexander University in Erlangen, Germany, conducted an elegant laboratory-based study to evaluate the impact of acute psychosocial stress on body movements and posture. Two unique features of the study were the use of inertial measurement unit (IMU)-based motion capture suits to characterize overall movement, as well as movements of different body segments, and machine learning methods to classify stressed versus nonstressed suites of movement information.
This commentary focuses on the larger of two coordinated and similar experiments reported by the study team.1 The study population included 39 healthy men and women (19–33 years old; 46% female). Acute mental states were experimentally altered using the Trier Social Stress Test (TSST), a laboratory stress induction gold standard, and the friendly-TSST (f-TSST), a control condition that is as similar as possible to the TSST but does not activate the hypothalamic–pituitary–adrenal axis. The sequence of exposures was random and implemented on two consecutive afternoons. Empirical biological measures of stress were characterized on both days, immediately prior to and following (f-) TSST exposures, with sequentially collected samples of salivary cortisol (7 samples over 90 min). Participants also completed the State and Trait Anxiety Inventory to assess self-reported state levels of anxiety and the Positive and Negative Affect Schedule to measure positive and negative affect before (f-) TSST exposures. Body movement and positions of body segments (e.g., head, limbs, and trunk) were quantified with a motion capture system consisting of 17 IMU sensors allowing full-body motion capture. This system outputs a rich matrix of data features for 23 body segments, including position and orientation in a global orientation system, and velocity, acceleration, angular velocity, and angular acceleration for all segments.
Results of cortisol assays and measures of state anxiety and negative affect confirmed that the TTST markedly increased psychosocial stress relative to the control f-TTST, with robust and statistically significant between-group differences for all outcomes. Motion sensor data suggest that that acute stress induction leads to a reproducible freezing behavior, characterized by less overall motion as well as more and longer periods of no movement. After statistical testing and Bonferroni correction, 43 out of 587 extracted IMU features showed significant differences between TSST and f-TSST. While not statistically significant, the average velocity of total body and trunk were lower when exposed to acute psychosocial stress. Examples of features that were significantly different in the stress-induced group include lower head acceleration and head movement entropy, and longer periods and percentage of time with no movement for the head, trunk, and upper extremities.
Machine learning-based procedures applied to IMU features were successfully used to classify the presence or absence of induced acute psychosocial stress. The classification experiments trained on features extracted over the complete (f-)TSST data achieved a high level of classification accuracy.
These findings confirm several previous studies that investigated the relationship between acute psychosocial stress and single aspects of body posture and movements, but are the first to extend this to the full body. The authors concluded that their results “suggest a close interplay between behavioral, endocrinological, and motoric systems during situations of acute stress and motivate the use of motion information as an additional digital biomarker to the established biopsychological markers for acute stress assessment to obtain a more holistic picture of the human stress response.”
The strengths of this study lie in its approach and future potential, rather than the specific findings. I do not think that any mind–body researcher, practitioner, or teacher would consider the finding that acute social anxiety can induce a “freezing” state as a new insight, as “fight, flight, or freeze” are well-established psychophysiological responses to stress. Visual and manual inspection of movement patterns and postures, which encode an individual’s relationship to gravity, are a key diagnostic element used by practitioners of Tai Chi, yoga, Feldenkrias, Rolfing, and Alexander techniques. However, the extension of subjective observation to objective, precise, and multidimensional aspects of movement opens up many new possibilities. In combination with varying experimental set-ups, e.g., different forms of acute stress (e.g., social, mechanical, verbal, and visual), and or in populations with varying baseline health conditions (e.g., PTSD, depression, chronic pain, and fear of falling), understanding whole and/or component body movement responses could lead to new insights in links between behavioral, neuroendocrine, and motor systems. Prior research in the field of embodied cognition suggests that experimentally altered patterns of arms swinging and trunk and head dynamics during gait can impact affect, and conversely, different affective states are associated with different gait patterns. Deploying noninvasive wearable motion capture systems could not only provide more detailed insight into motor and biomechanical responses to experimental stressors, but could be used as a form of biofeedback for training more functional responses and resilience to stressors. Knowing that exposure to a threat (e.g., an emotional challenge) consistently leads to a contraction of neck muscles, head immobility, leg stiffening and holding of breath, could lead to mind–body training strategies that replace reflective detrimental responses with more functional ones. However, given the complexity and nuances of human movement, body-wide arrays of movement (and other physiological) sensors will need to be coupled with machine learning and other emerging artificial intelligence-related quantitative methods. While it is unlikely that the insights and instructions of a skilled mind–body teacher will ever be replaced by wearable feedback systems or robotic instructors, emerging technologies may afford unique tools for both mechanistically understanding links between mind and body and adding high-tech tools to complement traditional soft-touch training.
Chatbots, Virtual Therapists, and Stressed-Out Students: A Digital Potpourri of Mindfulness
Darshan H. Mehta, Osher Center for Integrative Health, Harvard Medical School and Brigham and Women’s Hospital, Boston, MA, USA.
By nearly every metric, student mental health is worsening. Almost three-quarters of students reported moderate or severe psychological distress in another national survey conducted by the National College Health Assessment. Even before the pandemic, colleges and universities faced a surge in demand for mental health care that far outpaced capacity. The traditional counseling center model has proven ill-equipped to address this challenge. While positive trends such as reduced stigma around mental health issues have led more students to seek help, today’s college students grapple with a dizzying array of challenges—from coursework and relationships to economic strain, social injustice, mass violence, and various forms of loss related to COVID-19. Innovative approaches are needed to equip faculty and staff better to support distressed students.
Manualized stress management and mind–body interventions (e.g., mindfulness-based stress reduction) can be helpful; however, they face many barriers, including time and cost. Its typical delivery has also been limited, as it requires a live human being to deliver the intervention. On the other end, fully autonomous digital interventions need more engagement. Conversational agents, like chatbots, have greater engagement but lack facial expressions or tonal nuances. A newer technology, virtual humans, appears promising since they use facial expressions and body gestures through a realistic humanoid experience.
In this study by Karhiy et al.,1 the investigators delivered a mindfulness program to university students using three different delivery modes: a virtual human, a chatbot, and a live human being. Participants were randomly allocated to one of these three groups to receive the intervention. The virtual human was designed to resemble a young professional adult male and was programmed to behave like a therapist. Participants interacted with the virtual human through an autonomous animation on a website. The virtual human delivered the intervention script to the participants. In the chatbot intervention, participants in this group interacted with a text-based chatbot programmed to provide the same intervention script as the virtual human. The chatbot did not have a human-like appearance but communicated with participants through text messages. Last, participants in the human intervention group received stress management intervention from a male health psychology intern via Zoom. The individual delivered the mindfulness instructions to the participants similarly to an in-person therapy session. The intervention consisted of an initial 15-minute in-person session and instructions to engage in online homework sessions at least twice weekly for the next four weeks. The mindfulness instructions delivered to all three groups were identical to ensure consistency in the intervention content.
The intervention aimed to increase participants’ attention to and acceptance of their present experiences, ultimately reducing stress and enhancing mindfulness. By comparing the effectiveness of the virtual human, chatbot, and live human therapist in delivering the mindfulness intervention, the study sought to evaluate the potential of digital platforms, particularly virtual humans, in promoting mental well-being and stress reduction among university students.
The participants in the study completed a baseline questionnaire and a postintervention questionnaire. The mean age of the participants was 25 years, primarily females (78%) and various ethnic backgrounds represented. The primary outcome was perceived stress using the 10-item Perceived Stress Scale. Secondary outcomes included self-reported current stress, physiological measures of stress, mindfulness, perceptions of empathy and quality of interactions, and homework adherence and engagement.
Results indicated that all groups experienced significant improvements in perceived stress from baseline to follow-up. There was no significant interaction effect among the groups, suggesting that all interventions effectively decreased stress and enhanced mindfulness levels. The effect sizes for stress reduction and mindfulness improvement were moderate to large across the teletherapy, virtual human, and chatbot groups. There were no significant differences in heart rate or heart rate variability between groups or over time.
Furthermore, the study examined homework adherence and engagement among participants. The virtual human group exhibited significantly higher adherence to homework tasks than the teletherapy and chatbot groups. In addition, there was a significant difference in homework engagement and satisfaction ratings among the groups, with the chatbot group reporting lower engagement and satisfaction levels than the other two groups. Not surprisingly, the participants perceived the live therapist as more empathic than the chatbot or virtual human. The most significant critique of the virtual human was his robotic voice.
Overall, the study intervention provided participants with access to mindfulness training through innovative digital platforms, offering a convenient and potentially effective way to deliver mental health interventions to student populations. The findings suggest that virtual humans can effectively deliver mindfulness interventions and enhance homework engagement and adherence compared with other digital formats. These results are promising for developing future interventions utilizing virtual humans in mental health support. The study also highlights the importance of participant feedback in improving virtual human interventions, particularly in enhancing perceived empathy. Simply said, virtual humans can deliver mindfulness interventions effectively!
This research is provocative as it compares different forms of digital delivery of mindfulness interventions, including teletherapy, and measures their effects on self-reported and physiological measures of stress reduction and mindfulness. It underscores the potential of virtual humans to strengthen engagement and adherence to mental health interventions, paving the way for innovative approaches to supporting mental well-being. This type of research provides valuable insights into the advantages of using virtual humans to deliver mindfulness interventions and suggestions for enhancing empathy in virtual human interactions. Future studies must explore integrating artificial intelligence models to improve empathy in virtual human interventions; however, data privacy and ownership considerations would need to be addressed, primarily as they deal with sensitive health considerations, including mental health.