Statistical Learning Impairments as a Consequence of Stroke

Useless, describes a problem but offers NO solution. 

Statistical Learning Impairments as a Consequence of Stroke 

Albulena Shaqiri^1*, James Danckert², Lauren Burnett² and Britt Anderson^2*

• ¹Laboratory of Psychophysics, Brain Mind Institute, School of Life Sciences, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland
• ²Department of Psychology, University of Waterloo, Waterloo, ON, Canada

Statistical learning is the implicit learning of the contingencies between sequential stimuli, typically from mere exposure. It is present from infancy onward, and plays a role in functions from language learning to selective attention. Despite these observations, there are few data on whether statistical learning capacity changes with age or after brain injury. In order to examine how brain injury affects the ability to learn and update statistical representations, we had young control and healthy elder participants, as well as participants with either left or right brain injury, perform an auditory statistical learning task. Participants listened to two languages with made-up words that were defined by the transition probability between syllables. Following passive listening, learning was assessed with a two-alternative forced choice test for the most familiar word. As in previous studies, we found that young controls have a learning capacity limitation for statistical learning; a second language is less well learned than the first, and this statistical learning capacity limit is attenuated with age. Additionally, we found that brain damaged patients, whether with left or right hemispheric damage, showed impaired statistical learning. This impairment was not explained by aphasia or cognitive deficits. As statistical learning is a critical skill for daily life, a better appreciation of the nature of this impairment will improve our understanding of the cognitive effects of brain injury and could lead to new rehabilitation strategies.

Introduction

Our ability to quickly learn consistent relationships between sequential stimuli is called statistical learning (Turk-Browne, 2012). One prominent example of statistical learning is our ability to group sounds presented in a consistent order. Present in infancy, the statistical learning of word borders is held to be critical to normal language development (Saffran and Kirkham, 2018). Statistical learning is not, however, restricted to infants and children. This ability persists into adulthood and operates across multiple domains. In the spatial domain, we can learn that some events are more likely to happen in one location than another (Druker and Anderson, 2010; Cort and Anderson, 2013; Jiang et al., 2013). In the temporal domain, we learn that some sequences are more likely than others (Maljkovic and Nakayama, 1996; Fiser and Aslin, 2002), and we learn to predict interval durations (Danckert and Anderson, 2015). Statistical learning does not depend on an active, deliberate search for structure, though it may potentially be aided by such strategies (Gebhart et al., 2009).

Many different brain areas are involved in statistical learning. In a study using a word segmentation task, similar to what is to be reported here, Karuza et al. (2013) found significant changes in metabolic activity in the pars opercularis and pars triangularis of the left frontal lobe. However, when the statistical learning tasks are broadened beyond language based tasks, other brain regions are also highlighted. In their 2015 review, Schapiro and Turk-Browne (2015) reported that the superior temporal gyrus (important for sequential analysis) and the temporal-parietal junction (when regularities are violated) are also frequently found in functional imaging work on statistical learning tasks.

While statistical learning is functionally pervasive, has restricted anatomical correlates, and is present into adulthood, little is known about statistical learning capacity as we grow older, or as a result of brain injury. If statistical learning were to contribute, as seems likely, to implicit learning capacity in elders, then it is likely that events such as strokes would impact it. This would also have an impact on rehabilitation, as current stroke rehabilitation commonly emphasizes learning through repetition.

The hypothesis that brain injury might impair statistical learning after right hemisphere damage is consistent with our prior data on tracking environmental regularities in patients with stroke (Shaqiri and Anderson, 2012; Shaqiri et al., 2013; Stöttinger et al., 2014a,b). However, other data suggests that either hemisphere could be important. Wolford et al. (2000) studied two split-brain participants and a cohort of people with unilateral brain damage. Participants made predictions for two independent sequences of events that could appear in either their right or left visual fields, and thus, be processed by either the left or right hemisphere of the split-brain participants. Depending on which visual field the stimuli were presented in, it was found that both hemispheres were able to form, independently, statistical representations for the stimuli.

Thus, we know that statistical learning is present into adulthood, that it is supported by both hemispheres, and that it is functionally important. What we do not know is how statistical learning is impacted by brain injury, whether there are unique hemispheric effects, nor precisely which brain areas are critical for brain damage to impact statistical learning. We also do not know if statistical learning deficits after brain injury, should such occur, can be re-mediated by massed repetition, breaks, or information about the material to be learned, all of which have been suggested to improve the statistical learning capacity of healthy young individuals (Gebhart et al., 2009; Franco et al., 2011).

We initiated these experiments with several expectations: based on our prior probability learning work (Shaqiri and Anderson, 2013), we expected the presence of neglect to interact with statistical learning deficits. Therefore, we partitioned our right brain damaged (RBD) participants into subgroups with (+N) and without (−N) neglect. In some prior work on updating mental models, we had found differences between left (LBD) and right hemisphere stroke patients (Danckert et al., 2012), and so we also included a LBD group. While many types of learning decline with age, not all do, and this may be particularly true for some forms of implicit memory (Howard and Howard, 2015). Thus, we included both old (OC) and young controls (YC) to compare the stroke patients to participants of the same age, and to address age effects. However, our initial results showed a generally poor performance of all the brain damaged groups. We therefore undertook a more exploratory, less hypothesis driven effort to characterize the extent and nature of the impairment, its sensitivity to manipulations [that had been shown to work in healthy adults (Gebhart et al., 2009)], and a broader investigation of healthy participants to confirm that we were able to reproduce the basic effects reported previously with these testing materials. In the end, we ran several small studies in which all participants were tested on their ability to learn a single language and that allowed us to do a large omnibus test of statistical learning ability. In addition, we could also test the effects of the secondary manipulations in a broad way. Because these secondary manipulations are exploratory, there might be a greater risk of type II error than type I, thus, we have not adjusted for multiple statistical comparisons. Readers should be aware of the exploratory natures of our studies and the risk that any statistical significance may be inflated by multiple comparisons.

The outline of the paper is as follows. We first present the methods for all experiments and for all groups: healthy young, healthy elders, RBD +N, RBD −N, and LBD. We begin the results section with omnibus analyses. We show that our young adult data replicates prior work and that brain damage impairs statistical learning. We report the specific experiments in more detail to look for hints of any beneficial effects of increased exposure, different length rest breaks, or information about the nature of the learning task. Lastly, we discuss the fact that, despite our expectations, damage to either the left or right hemisphere appears equally likely to produce statistical learning impairments. We conclude with an exploratory voxel lesion symptom mapping analysis to ask, regardless of hemisphere, which brain region is most predictive of statistical learning impairments.