Neuropsychological assessment and virtual reality training of social prediction in patients with cerebellar malformation

It has been proposed that impairments of the predictive function exerted by the cerebellum may account for social cognition deficits. Here, we integrated cerebellar functions in a predictive coding framework to elucidate how cerebellar alterations could affect the predictive processing of others’ behavior. Experiment 1 demonstrated that cerebellar patients were impaired in relying on contextual information during action prediction, and this impairment was significantly associated with social cognition abilities. Experiment 2 indicated that patients with cerebellar malformation showed a domain-general deficit in using contextual information to predict both social and physical events. Experiment 3 provided first evidence that a social-prediction training in virtual reality could boost the ability to use context-based predictions to understand others’ intentions. These findings shed new light on the predictive role of the cerebellum and its contribution to social cognition, paving the way for new approaches to the rehabilitation of the Cerebellar Cognitive Affective Syndrome.


Introduction
. Demographic and clinical information of the three groups in experiment 1. 19

7
TD participants showed significantly higher accuracy for the 90% action-context co-occurrence 8 (86.00±2.10%) compared to the 10% one (71.23±3.21%; p<0.011). This latter condition was also 9 lower than both the 40% (81.08±2.97%; p=0.011) and 60% (78.87±3.53%; p=0.040) conditions. 10 Similarly, CND patients were less accurate in the 10% condition (66.00±4.68%) compared to the 11 90% (83.12±2.62%; p<0.001), 40% (79.27±3.52%; p<0.001) and 60% (74.65±4.30%; p=0.024) ones. 12 Furthermore, their accuracy was higher in the 90% than in the 60% condition (p=0.031). Conversely, 13 no difference emerged within the CM group (10%: 68.08±3.69%, 40%: 62.81±3.41%, 60%: testing phases were treated, respectively, with between-subjects and mixed model ANOVA designs. 5 Asterisks indicate significant between-group comparisons for the familiarization (p=0.001) and for 6 the highest-probability condition of the testing phase (p<0.001). 7 To correlate performance at the task and social perception abilities, in line with previous across trials of the testing phase that represents the modulatory effect of the probabilistic associations, 10 thus providing a measure of the strength of the contextual priors. Correlation analyses (  In this experiment, we aimed to qualify the results of Experiment 1 by investigating whether 8 the deficits of CM patients in using contextual priors are specific for action prediction or, 9 alternatively, reflect the impairment of a domain-general predictive mechanism. We, thus, 10 administered 18 CM patients and two control groups of 18 CND patients without cerebellar 11 alterations and of 18 TD peers (see Table 3) with the same task as in Experiment 1 and with a shape 12 prediction task developed to assess the use of contextual priors for predicting physical events (Bianco 13 et al., 2020). The structure of the shape prediction task was similar to the action prediction task, but 14 participants were required to predict the unfolding of moving geometrical shapes (see Fig. 1c,d and 15 Material and Methods for detailed information). In line with the UCT hypothesis (Schmahmann, prediction of both social and physical events. However, previous research reported that cerebellar 1 patients may present stronger difficulties in processing social stimuli than inanimate objects (Cattaneo 2 et al., 2012; Sokolov et al., 2010). Accordingly, we anticipated that CM patients could perform worse 3 in the action prediction task than in the shape prediction task. Furthermore, in this experiment we 4 directly controlled for the influence of intelligent quotient (IQ) on the performance in the two tasks.  These findings support a domain-general prediction deficit due to cerebellar alterations as expected 2 by the UCT hypothesis (Schmahmann, 2019). The ANOVA also yielded significant effects of task 3 (F1,51=4.24, p=0.045, η 2 p=0.08), with better performance at the action than shape prediction task, but 4 its interactions with group and probability were non-significant (all p>0.194). Partially in contrast to ruled out that CM patients had worse performance with social stimuli than with physical events. 7 Moreover, in the follow-up ANCOVA, nor the main effect of task neither its interactions were 1 Figure 4. Results of Experiment 2. Eighteen patients per group were administered with the action 2 prediction task and the shape prediction task. Data processing and statistical analyses were performed 3 following the same design of Experiment 1 with the addition of task as within-subject variable. Then, 4 two follow-up ANCOVAs were used to partial out the effects of IQ. In the familiarization phase, the 5 ANCOVA showed a significant effect of the covariate IQ (p<0.001), while all other effects were non-6 significant (all p>0.053). In the testing phase, the main effect of group (p=0.020) and the group x 7 probability interaction (p<0.001) were still significant after partialling out the effects of IQ, but nor 8 the main effect of task neither its interactions were significant (all p>0.204). In this experiment, we tested how social prediction can be boosted in CM patients by a VR 11 training specifically developed to improve predictive abilities in a social scenario (VR-SPIRIT; 12 (Butti, Biffi, et al., 2020)). In the VR-SPIRIT, the participants were immersed in a playground scenario and, in each of the 80 total trials, they were asked to compete with one of four avatars for 1 reaching one of three recreational objects. Specific features of the scenario forced the participants to 2 anticipate the behavioral preference of each avatar, which, crucially, was associated to the objects 3 with pre-established probabilities. The use of predictive and random strategies in anticipating avatars' 4 intentions was computed respectively, by the mean percentage of scores obtained when the 5 probabilistic avatar-object association gave clues on avatar's intention (prediction score) and by the 6 mean percentage of scores obtained when context did not provide reliable information (random 7 score). CM patients were randomly assigned to the VR-SPIRIT or to a VR-based motor training and 8 then exposed for eight 45-minute sessions to one of these two interventions, which were run on the 9 same VR platform. Before and after the training, CM patients were administered a VR evaluation 10 session exploiting the same probabilistic design of the training sessions, but presenting a different 11 scenario, and the action prediction task adopted in the former experiments. This way, we evaluated 12 the transferability of a VR experience to the use of contextual priors for predicting social events in 13 patients with cerebellar alterations, paving the way to new rehabilitative approaches for CCAS 14 (Argyropoulos et al., 2020). 15 Preliminary analyses confirmed that participants assigned to the VR-SPIRIT and to the motor 16 training were comparable for age (Z=- r=0.63). Furthermore, the beta index at the action prediction task was higher after the VR-SPIRIT 5 than after the motor training (Z= -1.97, p= 0.049, r= 0.44), thus suggesting that training predictive 6 abilities in a social VR scenario also boosted the implicit learning and use of contextual priors during 7 action perception ( Fig. 5 and Table 4).
8 Table 4. Between-groups comparisons for the action prediction task and the VR evaluation sessions 9 before and after the two rehabilitative interventions. Significant results are reported in bold. In this work, we tested the effects of cerebellar alterations on the predictive processing of 12 incoming social and physical events. In line with the UCT hypothesis (Schmahmann, 1996, 2019), 13 our results showed that cerebellar patients showed a domain-general prediction deficit. In a recent 14 ALE meta-analysis, Siman-Tov and colleagues found a consistent activation of the cerebellum during alterations impaired the processing of contextual information, resulting in less precise predictions 3 when the available sensorial information was not sufficient to inform on the incoming events (Friston,4 2012a). The anatomo-functional organization of the cerebellum seems optimal for processing the co-5 occurrence between contextual and sensorial cues and integrating them into internal prediction inputs, crucially contributing to selecting the most probable event within a specific context. Thus, 9 within a Bayesian paradigm of brain functioning, the cerebellum could modulate the interactions 10 between cortical nodes of specific cognitive networks by supplying contextual priors that constrain 11 stimulus processing at any stage (Sokolov et al., 2017). 12 Our result of comparable deficits in the prediction of social and physical events suggests that 13 the cerebellum may apply its predictive computation regardless of the social nature of the processed 14 information. This finding is partially in contrast to a previous study on adult patients with acquired Despite the cerebellar coding of context-based predictions may operate across different 1 domains, an impairment in using contextual priors is likely to have major consequences in the social 2 domain, in which the context is mostly crucial to disambiguate others' behavior (Brown & Brüne, 3 2012). Accordingly, we showed that CM patients were more impaired than CND patients not only in 4 using contextual priors to predict the unfolding of others' actions, but also in discriminating the most 5 likely facial expression within a specific social scenario, as assessed by the NEPSY-II non-verbal 6 ToM subtest. Furthermore, the context-based prediction abilities of CND patients were associated 7 with their performance at the NEPSY-II subtests assessing affect recognition and non-verbal ToM     The results of this study should be discussed considering its limitations. Even though we 15 calculated a-priori the sample size for Experiments 1 and 2, we cannot exclude that the absence of 16 differences in executing the two prediction tasks might reflect a power issue also due to the  In conclusion, the present study provided evidence that a general deficit in using contextual 5 priors to predict incoming sensorial information might underlie the social perception deficits reported 6 in patients with cerebellar alterations. However, not only did we document these deficits and probe 7 their relevance for general social cognition abilities; we also provided evidence that a short-lasting, 8 but intense (i.e., eight sessions in two weeks), VR training designed to boost the learning of statistical 9 regularities in others' behavior can reinforce context-based predictions across different scenarios.      Action prediction task (Experiments 1, 2, 3) 3 We adopted the same task of Amoruso and colleagues (Amoruso et al., 2019). In this two-4 alternative forced choice (2AFC) task, participants were exposed to short videos that depicted a child 5 executing distinct grasping actions with two diverse intentions and they were asked to predict the 6 final outcome of the action (i.e., the motor intention). Before starting the experiment, participants 7 were introduced to the objects displayed in the videos, namely an apple and a glass, and were 8 informed about the different feasible object-manipulations associated with either individual (i.e., to 9 eat/drink) or interpersonal actions (i.e., to offer). Importantly, the apple and the glass were presented, 10 respectively, on a plate and on a tablecloth that could be of two different colors (i.e., orange or violet 11 for the plate, white or blue for the tablecloth). 12 The task consisted of two blocks, each comprising a familiarization phase (80 trials) 13 immediately followed by a testing phase (40 trials), for a total of 240 trials. Participants were 14 presented with the same videos in both phases, but in the testing one video duration was drastically 15 reduced. In the familiarization phase, videos were stopped just two frames before the hand-contact 16 with the object (25 frames, each frame lasting 33.33 ms for a total of 833 ms), thus displaying almost 17 in full the action unfolding. Importantly, during this phase the probability of presentation of a given 18 action in association with a specific contextual cue was biased by setting the probability of action- 19 contextual cue associations to 10%, 40%, 60%, or 90% of the total number of trials. In this way, the 20 prior expectation was implicitly manipulated (for further details see (Amoruso et al., 2019)). The type 21 of action-context association was counterbalanced between participants but remained constant in the 22 two blocks of familiarization within the same participant.

23
In the testing phase, video duration was drastically shortened (500 ms, 15 frames), thus 24 hindering kinematic information. Given the manipulation of the probability of co-occurrence between 25 the actions' intentions and the color of contextual cues in the familiarization phase, we expected that participants could implicitly rely on the contextual priors to overcome kinematics uncertainty, 1 presenting a probabilistic modulation in their responses. Notably, differently from the familiarization 2 phase, during the testing phase each of the 8 videos was presented for an equal number of trials (i.e., 3 five trials in each blocks). Moreover, no information about the probability of associations was 4 explicitly given to participants. 5 For both phases, each trial started with the presentation of a fixation cross lasting 2,000 ms, 6 followed by video presentation. Immediately after or during the videos (for the familiarization or the 7 testing phase, respectively) the Italian descriptors of the two possible actions (i.e., the verbs 8 "mangiare" or "bere" and "offrire", in English "to eat" or "to drink" and "to offer") were presented 9 until a response was recorded. The descriptors, written in white on a black background, were located 10 on the right and on the left side of the screen, with the location counterbalanced between participants 11 and consistent across trials and blocks for each participant. Participants had to respond by pressing 12 with their right or left index finger, respectively, the "m" (right) or the "z" (left) computer key of a 13 QWERTY keyboard. The experiment was ran in a single session lasting ~40 min, using E-Prime V2 14 software (Psychology Software Tools, Inc., Pittsburgh, PA, United States) on a 15.4-inch LCD 15 computer screen (resolution 1600 X 900 pixels, refresh rate 60 Hz). Participants were asked to sit in 16 front of the monitor at 60 cm. Short breaks were allowed between blocks and phases. 17 Shape prediction task (Experiment 2) 18 This 2AFC task was developed by following the same administration procedure and structure 19 of the action prediction task, but moving geometrical shapes were instead displayed. Participants were  as "short" and the unequal length side polygons as "long". Thus, for both the familiarization and the 13 testing phases, participants were asked to report whether the moving shape was a short (i.e., square 14 or parallelogram) or a long (i.e., rectangle or trapezoid) shape. Immediately after the videos, the 15 Italian descriptors of the two possible answers (i.e., "corta" o "lunga", in English "short" or "long") 16 were presented in the prompt response frame. 17 Importantly, each couple of polygons and the respective receptor could be differently colored, 18 using the same probabilistic manipulation as for the action perception task (Figure 2). During the 19 familiarization phase, the shapes fully appeared on the screen so that they could be easily identified 20 (933 ms, corresponding to 28 frames), while, in the testing phase, videos were interrupted one frame 21 after the halfway appearance of the horizontal segment (500 ms, corresponding to 15 frames), thus 22 providing minimal information about the specific shape and prompting the use of the contextual priors 23 to overcome sensory uncertainty.

24
A resume of the two predictive tasks is reported in Fig. 1.

VR rehabilitative interventions (Experiment 3)
The VR sessions were administered at the Grail Lab (Motek, Amsterdam, NL) at the 1 Scientific Institute, IRCCS E. Medea. In the experimental VR-SPIRIT, participants were immersed 2 in a playground scenario including a swing, a circular carousel and a rocking carousel. In each trial, 3 participants were asked to compete with one of four avatars for reaching one of these objects. To do 4 so, the participants had to anticipate the behavioral preference of each avatar, since they could not 5 pass the avatar after the direction of its movement was clear. Crucially, the avatars were associated 6 with pre-established probabilities to the objects, so that each of three avatars moved toward his 7 preferred object in the 80% of trials and chose each of the other two objects in the 10% of trials. 8 Conversely, the fourth avatar moved to each object with the same probability. The behavioral 9 preferences of the avatars remained constant within each session (80 trials), so that the participants 10 could learn the probabilistic avatar-object associations, but they were changed across sections. 11 The active control training consisted in a navigational game and in four GRAIL games 12 previously used in motor rehabilitation (Cesareo et al., 2017). Notably, no social agents were 13 presented, and no prediction abilities were required while playing these games.   (1-Beta). Accordingly, we enrolled 26 patients per group. Firstly, we 21 adopted a one-way ANOVA and a Chi-square test to verify that groups were comparable for age and 22 gender. For the action prediction task, response times (RT) were recorded, but not included in the between the correlations of the two clinical groups. Finally, we used two-tailed independent samples 10 t-test to analyze differences between groups in FSIQ and social perception abilities. 11 For Experiment 2, statistical analyses were performed following the same design of 12 Experiment 1 with the addition of task as within-subject variable. Furthermore, a follow-up 13 ANCOVA was used to partial out the effects of IQ. Given the 2 tasks x 4 probabilities x 3 groups  20 "as in SPSS" option. We reported ANOVA effect sizes as partial Eta squared (η 2 p), adopting 21 conventional cut-offs of η 2 p=.01, .06; and .14 for small, medium, and large effect sizes, respectively For Experiment 3, we used non-parametric tests for analyses due to small sample size 1 impacting on normal distribution of data. Preliminarily, we verified that the two groups had 2 comparable age, IQ, and gender (Mann-Whitney U/ Chi 2 tests). For the VR evaluation sessions, we 3 considered the total score and the mean duration of each trial. With the aim to weight the use of 4 predictive and random strategies, we computed, respectively, the mean percentage of scores obtained Data availability statement: The anonymized datasets generated and analyzed during the 1 current study (Experiment 1-2-3) are available from the Open Science Framework, at this link: