New approaches to investigating social gestures in autism spectrum disorder
© Kishida et al; licensee BioMed Central Ltd. 2012
Received: 10 February 2012
Accepted: 27 April 2012
Published: 24 May 2012
Skip to main content
© Kishida et al; licensee BioMed Central Ltd. 2012
Received: 10 February 2012
Accepted: 27 April 2012
Published: 24 May 2012
The combination of economic games and human neuroimaging presents the possibility of using economic probes to identify biomarkers for quantitative features of healthy and diseased cognition. These probes span a range of important cognitive functions, but one new use is in the domain of reciprocating social exchange with other humans - a capacity perturbed in a number of psychopathologies. We summarize the use of a reciprocating exchange game to elicit neural and behavioral signatures for subjects diagnosed with autism spectrum disorder (ASD). Furthermore, we outline early efforts to capture features of social exchange in computational models and use these to identify quantitative behavioral differences between subjects with ASD and matched controls. Lastly, we summarize a number of subsequent studies inspired by the modeling results, which suggest new neural and behavioral signatures that could be used to characterize subtle deficits in information processing during interactions with other humans.
The challenges of social exchange are shared throughout the animal kingdom [1–3]. Humans engage in repeated reciprocal interactions with kin and non-kin; this has required our species to develop a particular capacity to track these interactions and assign credit and blame accordingly [3–6]. The underlying neurobiological mechanisms underlying these abilities in humans remain an open area of investigation, which is generating insight into a number of mental illnesses including autism spectrum disorders (ASD) [7–18].
One important challenge for an agent engaged in social exchange is the ability to generate models of others’ mental states, a capacity referred to as ‘theory of mind’ [19, 20]. Human neuroimaging experiments have implicated a consistent set of brain regions hypothesized to be involved in this process [21–24]. The role of this kind of computation and others are exemplified in even the simplest exchanges. For instance, in a fair trade the brains of the interacting agents must be able to: (1) compute norms for what is considered fair; (2) detect deviations from such norms; and (3) select appropriate actions based on these deviations.
Reciprocating social exchanges and their related computations occur in our daily lives and are the basis for a number of staged interactions in clinical practice. However, the computations themselves likely occur well below our threshold for conscious experience. For example, in the diagnostic procedures for ASD a trained professional interacts with a suspected patient and navigates a predetermined series of give-and-take scenarios (for example, the Autism Diagnostic Observation Scale (ADOS) ); during this interaction the clinician tries to detect typical and atypical social gestures based on trained-up (for example, experience-based) expectations of behavior. Additionally, insight into the patient’s social cognitive process is sought using interviews of the patient and family members [26, 27].
Computational approaches in combination with quantitative probes of behavior (for example, game theoretic paradigms [28, 29]) promise to reveal underlying dimensions of healthy human decision-making and thereby will provide a basis for which to compare against in psychiatric populations [30, 31]. These approaches promise to excel, over traditional approaches, in capturing quantitative normative behavior and associated neurobiological responses. In these tasks participants are required to make decisions under incentivized conditions; the incentive is typically a monetary one and the decision spaces are restricted by the particular game . The games may be against a ‘roll of the dice’ or against real or simulated agents (for example, other humans or computer simulations) and can be employed though a computer interface. These conditions can augment traditional diagnostic procedures by providing additional behavioral and neurobiological measurements. In computer-based game play, the behaviors elicited and signals exchanged between players are all captured within the patient-computer interface. These measurements can then be directly ported into tests over competing computational models about hypothesized cognitive processes underlying social behavior. In addition, variables within the games are designed to be parametric, which provides increased specificity in the analysis of associated brain responses. The use of the internet and technologies like hyperscanning, which synchronizes fMRI scanners and study participants over the internet , has already and will continue to allow more complex social exchange to be studied while allowing experimenters to have control over which signals are exchanged. These methods take highly complex social interactions and reduce the exchanges to simpler, more tractable, dimensions for quantitative analysis; and, have recently been applied to the neurobiological investigation of social exchange [33–38]. Fairness and cooperation games dominate these recent efforts because they assess a subject’s internal norm for what is fair in an exchange, and they require that each subject model their partner’s mental state [28, 39–45]. These games are excellent experimental probes because they are simple, there is an existing body of behavioral data employing them across a variety of contexts, and there are known solution concepts for how they ‘should’ be played by a rational self-interested agent [28, 40, 41]. Importantly, these games all require participants to model their partner. Such simplified behavioral probes provide a starting point for extracting quantitative descriptions of social signaling and its pathologies. These games are beginning to prove valuable in early work in clinical populations such as borderline personality disorder  and ASD [45, 46].
The multi-round trust game (Figure 1B, [34, 46, 47]): (1) there are two players, an anonymous ‘investor’ and an anonymous ‘trustee’; both players have full knowledge of what the other player’s options are given the rules of the game; (2) the investor is given 20 points to invest with the trustee (should they decide to); (3) once an investment is made (range of possibilities = 0 to 20 points in 1-point increments) the trustee receives three times the investment and the investor keeps what they did not invest; (4) now the trustee decides how much to reciprocate to the investor (the range of possibilities = 0 to 3 times the investment, in 1-point increments); (5) the trustee keeps what they did not reciprocate and the amount reciprocated is added to the investor’s total; (6) at the end of each round the totals for each player are shown to both players and ‘deposited’ into each of their respective banks; (7) this game is repeated for a total of 10 rounds where each round starts with a fresh 20 points and the totals at the end of each round are put aside to be tallied up at the very end of the game; (8) at the end of 10 rounds the totals in the bank for each player determines how much real money they will be paid and is how the game is incentivized: the more points one earns the more real money they will take home. To earn the most points possible the players must cooperate; however, generous signals of cooperation can be taken advantage of and a player may cheat the other out of an equal distribution of the profits.
The stag-hunt game also engages two participants and also requires participants to cooperate to achieve maximal gains. The goal of the game is to ‘hunt’ either a ‘stag’ (big profits) or a ‘rabbit’ (smaller profits). The players see a game board where they can move one square at a time and do so sequentially. In this manner, the participants observe the behavior of their partner and attempt to infer whether the move was intended to hunt the stag (cooperate) or a rabbit (defect). The rabbits do not move and are positioned on the board such that they can be hunted by the efforts of a single player, but provide a small payoff; on the other hand, two players must cooperate to hunt the stag and are rewarded for doing so with a bigger point value. Each move in the game that does not collect either a rabbit or a stag is viewed as a cost and diminishes a player’s point total. In the implementation of this game by Yoshida and colleagues the stag moved first, then the human participant, then a computer agent partner . Like the multi-round trust game the players point totals are directly tied to the amount of real money they will take home at the end of the experiment, thus incentivizing participants to make truthful expressions of their preferences and assessments of the game state.
Yoshida and colleagues’ results from applying the stag-hunt game to investigate hidden cognitive processes such as belief inference in autistic populations is particularly promising for our understanding of ASD and for computational and game theoretic approaches to investigating the heterogeneity known to exist within the spectrum . The application of their ‘game theory of mind’  demonstrated an ability to differentiate ASD symptom severity by a parameter in their model related to the patient’s ability to infer the strategic sophistication of their partner; the authors clarify this as the ability to infer the partner’s mindreading strategy. Additionally this parameter was distinct from the models estimate of patients’ ability to plan iteratively, which was not related to symptom severity, but rather showed a strong correlation with intellectual ability as assessed by verbal and non-verbal intelligence tests .
The multi-round trust game has been used to investigate social exchange in pairs of healthy individuals [34, 46–48] and pairs consisting of a healthy investor and a trustee diagnosed with a psychiatric disorder including: ASD [46, 48], borderline personality disorder [48, 49], major depression , and attention deficit hyperactivity disorder . Tomlin and colleagues identified a specific response pattern along the anterior to posterior axis of the cingulate cortex . They identified that a spatial pattern of activity that corresponded with ‘self’ and ‘other’ phases of the trust game exchange. This agent-specific response did not modulate with the number of points exchanged, the character of the gesture (benevolent or malevolent [34, 47]), or role (investor or trustee) of the player; the response pattern was specific to who was acting at that particular stage of the game (that is, ‘me’ or ‘not me’). Additionally, the agent-specific pattern of activity was absent when the players participated in a control experiment where the partner was absent and the players knowingly engaged in a computer-driven task . This result suggested that the cingulate ‘self’ and ‘other’ response patterns may be affected in special populations where these signals would be important for social exchange.
Where x i is the amount of money acquired by player i and similarly for player j. The α i and β i parameters model envy (player j gets more than player i) and guilt (player i gets more than player j). In Ray et al. , these parameters were taken as the ‘type’ of each player in the exchange and hence stratifies types of players according to their reaction to inequitable exchanges. That model used the observed patterns of monetary exchange to infer the type of each player and also included an estimation of the most likely depth-of-thought used by each player in the exchange - that is, how deep into the interaction each player simulated in order to decide on their next move. This is a potentially important estimate in the context of ASD since this kind of parameter may provide a new way to characterize one part of afflicted individual’s differences in socio-emotional reciprocity. The model provided a full generative model that was able to capture broad classes of behavior observed in subjects carrying out the trust game, but it has yet to be applied directly to subjects with ASD playing the game although these data exist .
The number of rounds back (‘t-n’), the order of the polynomial, and the number of clusters to be identified were all left as free parameters to be discovered in their approach. Here, a Bayesian classification algorithm is applied to the regression coefficients (β). One important feature of this approach is that it focuses on how the history of the game impacts the decisions of the healthy investors. In all cases where psychopathology groups were used as subjects, these players were always in the trustee role (Figure 6A). Hence, studying the impact of such groups on investor behavior uses the investor brain as a kind of biosensor for features of the traditional psychopathology group behavior as expressed through this reduced two-person exchange. Figure 6B shows the outcome of using this method on a very large population of dyads where a healthy investor played a trustee drawn from a number of psychopathology groups. The approach is different from traditional cluster analysis in that a probability distribution is estimated for all clusters so that ASD subjects have a probability measure of being in each cluster. ASD subjects were over-represented in cluster 2 as shown in Figure 6B.
Work on these and similar models is underway and will require large amounts of normative data to be able to develop individual differences metrics useful in the real-world of clinical applications. So while these results are very preliminary, they are promising.
Social impairments are among the defining characteristics of ASD (from DSM IV-TR) . Several influential theories have been posited to explain these social impairments seen in autism. The most prominent hypotheses implicate anomalies in theory-of-mind, joint attention, and functional cortical organization that contribute to social deficits in ASD. Together, studies across these domains highlight that individuals with autism perform quite well in some tasks that require inferring the beliefs and intentions of others, yet show marked deficits in other social inferences. The current state of the art in ASD research suggests there is significant value for probing ASD subjects using staged social exchange games paired with modeling approaches for estimating objective parameters associated with deficits in social computations.
To the authors’ knowledge, Yoshida and colleagues are the first to report a computational model of theory-of-mind processing applied to an ASD cohort . In this work they take advantage of the staged interaction of the stag-hunt game. Like the trust game the stag-hunt game incentivizes cooperation, but players must infer from the behavior emitted from their partner whether cooperation is being reciprocated and whether it remains the most efficacious strategy. If the players infer a defection from cooperation then the behavioral strategy must be flexibly adapted. To develop their ‘game theory of mind’  the authors represent the goals of each player with value functions derived from the machine learning literature. Given two players in the stag-hunt game the authors develop an estimation for their joint value function and recursively iterate: my representation of your value functions, your representation of my value functions, my representation of your representation of my value functions, and so on. The authors assume that this recursion is bounded; in other words, that it does not not go on ad infinitum and they estimate the depth of thought (that is, sophistication) from the expressed behavior of the players. Importantly, this mathematically expressed theory-of-mind is only one form of the kind of computational expression that may capture the kinds of inferences that occur during processes such as mentalizing or that may reveal aberrant computations carried out in the behavior and neural processes in individuals with autism. Other models have been developed [50, 53, 54], but have not been applied to the ASD population, which may capture or highlight other interesting computations necessary for adaptive mentalization. These kinds of approaches are likely to yield interesting quantitative characterizations of behavioral strategies and associated brain responses. Indeed, Yoshida and colleagues have reported interesting neural responses associated with parameters in their model in healthy volunteers  and have separately reported significant and interesting behavioral findings in a cohort of participants diagnosed with ASD . In the latter account, the group revealed quantitative evidence for deficits in theory-of-mind processes, but also revealed heterogeneity in their ASD cohort. This kind of heterogeneity is to be expected in the broader ASD population and this finding is an important demonstration of the quantitative power that these games and applied computational approaches may have going forward.
These early applications of game theory to probe the behavioral space of social exchange have already revealed interesting aspects that may be critical for healthy social interaction. The advantages of these approaches will also be derived from subsequent investigations into the neural responses engaged during these games where scientists may further reduce the stimulus space to test the role of specific brain responses observed to be important. For example, Chiu and colleagues utilized the multi-round trust game to reveal reduced activity in the MCC during live social exchange that correlated with increased symptom severity on two social subscales and the overall score of the revised Autism Diagnostic Interview, Revised [26, 46]. Lack of a neural signal for the first-person point-of-view in the cingulate cortex is suggested by the follow-up work presented here investigating neural responses to perspective taking in a experiment that controlled for effects of expertise familiarity (Figure 4). The results and current interpretation are consistent with a major subset of the clinical features of ASDs including deficits in imaginative play, communication, and social exchange. In addition, other literature suggests that the cingulate cortex may be a key anatomical region that differentiates humans and great apes from other primate species [55, 56] and is a major site of cognitive integration serving many robust functions [57, 58]. The MCC has specifically been implicated in the attribution of ownership of limbs and actions [47, 59, 60], willful production of speech and action [61, 62], and more recently agent discrimination and perspective-taking during social exchange [46, 47]. Our results suggest that the visual response in MCC demonstrated here may be related to an automatically generated representation akin to that elicited during first-person perspective-taking. The various roles suggested for the cingulate cortex imply that the function of this region is complex and may serve highly integrated and complex representations required for normal human cognition. Its role in mentalizing or theory-of-mind computations has yet to be investigated using a computationally rigorous model, suggesting an opportunity for future work in this domain.
Game theoretic approaches for measuring deviations from healthy decision-making behavior and the associated neural responses are a novel, but rapidly developing area in quantitative neurobiological approaches to understand mental disorders . Neuroeconomic studies of healthy humans engaged in social decision-making tasks has revealed a number of interesting insights into brain responses associated with healthy human behavior and are beginning to make inroads to understanding ASDs. Early investigations into the possible genetic underpinnings have revealed that parameters associated with economic probes may show some significant heritability [63, 64]. An advantage that the economic probes carry over traditional DSM criteria is in the objective and quantitative nature of their assessments of behavior. As these probes and computational descriptions of estimable model-based parameters develop so will the potential to quantify and decompose specific dimensions of social exchange. This is ultimately a task of reducing behavior and neural responses associated with mental illness to sub-components with the hope of providing refined targets for diagnostic specificity and novel (or improved) treatment strategies.
The work presented here represents the earliest developments of the use of computational and game theoretic approaches to understand ASD. We present computationally obtained results from hyperscanned interactions between healthy participants and participants diagnosed with ASD. These interactions were social, however they were reduced to quantitative exchanges governed by the rules of the multi-round trust game. From these results a region of interest in the middle cingulate cortex was determined that was shown to differentiate self and other perspective-taking in an eyes-closed mental imagery task. Additionally we demonstrate that passively viewing pictures of oneself or another person elicits responses in this same region, which suggests the capacity to develop passive picture viewing assays for middle cingulate function using fMRI. Clustering algorithms applied to the behavioral gestures elicited during the multi-round trust game and Bayesian modeling approaches to social interaction were discussed to highlight the impact quantitative approaches to ASD are beginning to have. This work is clearly in its earliest stages, but is showing promise towards having objective and mathematically explicit theories of social cognition that hope to elucidate associated neurobiological mechanisms at multiple levels of investigation.
This work was funded by the Kane Family Foundation (PRM), Wellcome Trust Principal Research Fellowship (PRM), and the National Institutes of Health: R01-DA11723 (PRM), R01 MH085496 (PRM), and T32-NS43124 (KTK).
This article is published under license to BioMed Central Ltd. .This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.