Predicting recovery following stroke: Deep learning, multimodal data and feature selection using explainable AI.

Machine learning offers great potential for automated prediction of post-stroke symptoms and their response to rehabilitation. Major challenges for this endeavour include the very high dimensionality of neuroimaging data, the relatively small size of the datasets available for learning and interpreting the predictive features, as well as, how to effectively combine neuroimaging and tabular data (e.g. demographic information and clinical characteristics). This paper evaluates several solutions based on two strategies. The first is to use 2D images that summarise MRI scans. The second is to select key features that improve classification accuracy. Additionally, we introduce the novel approach of training a convolutional neural network (CNN) on images that combine regions-of-interests (ROIs) extracted from MRIs, with symbolic representations of tabular data. We evaluate a series of CNN architectures (both 2D and a 3D) that are trained on different representations of MRI and tabular data, to predict whether a composite measure of post-stroke spoken picture description ability is in the aphasic or non-aphasic range. MRI and tabular data were acquired from 758 English speaking stroke survivors who participated in the PLORAS study. Each participant was assigned to one of five different groups that were matched for initial severity of symptoms, recovery time, left lesion size and the months or years post-stroke that spoken description scores were collected. Training and validation were carried out on the first four groups. The fifth (lock-box/test set) group was used to test how well model accuracy generalises to new (unseen) data. The classification accuracy for a baseline logistic regression was 0.678 based on lesion size alone, rising to 0.757 and 0.813 when initial symptom severity and recovery time were successively added. The highest classification accuracy (0.854), area under the curve (0.899) and F1 score (0.901) were observed when 8 regions of interest were extracted from each MRI scan and combined with lesion size, initial severity and recovery time in a 2D Residual Neural Network (ResNet). This was also the best model when data were limited to the 286 participants with moderate or severe initial aphasia (with area under curve = 0.865), a group that would be considered more difficult to classify. Our findings demonstrate how imaging and tabular data can be combined to achieve high post-stroke classification accuracy, even when the dataset is small in machine learning terms. We conclude by proposing how the current models could be improved to achieve even higher levels of accuracy using images from hospital scanners.

Quantifying error in effect size estimates in attention, executive function, and implicit learning.

Accurate quantification of effect sizes has the power to motivate theory and reduce misinvestment of scientific resources by informing power calculations during study planning. However, a combination of publication bias and small sample sizes (∼N = 25) hampers certainty in current effect size estimates. We sought to determine the extent to which sample sizes may produce errors in effect size estimates for four commonly used paradigms assessing attention, executive function, and implicit learning (attentional blink, multitasking, contextual cueing, and serial response task). We combined a large data set with a bootstrapping approach to simulate 1,000 experiments across a range of N (13-313). Beyond quantifying the effect size and statistical power that can be anticipated for each study design, we demonstrate that experiments with lower N may double or triple information loss. We also show that basing power calculations on effect sizes from similar studies yields a problematically imprecise estimate between 40% and 67% of the time, given commonly used sample sizes. Last, we show that skewness of intersubject behavioral effects may serve as a predictor of an erroneous estimate. We conclude with practical recommendations for researchers and demonstrate how our simulation approach can yield theoretical insights that are not readily achieved by other methods such as identifying the information gained from rejecting the null hypothesis and quantifying the contribution of individual variation to error in effect size estimates. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Transient Attention Gates Access Consciousness: Coupling N2pc and P3 Latencies Using Dynamic Time Warping.

The N2pc and P3 event-related potentials (ERPs), used to index selective attention and access to working memory and conscious awareness, respectively, have been important tools in cognitive sciences. Although it is likely that these two components and the underlying cognitive processes are temporally and functionally linked, such links have not yet been convincingly demonstrated. Adopting a novel methodological approach based on dynamic time warping (DTW), we provide evidence that the N2pc and P3 ERP components are temporally linked. We analyzed data from an experiment where 23 participants (16 women) monitored bilateral rapid serial streams of letters and digits in order to report a target digit indicated by a shape cue, separately for trials with correct responses and trials where a temporally proximal distractor was reported instead (distractor intrusion). DTW analyses revealed that N2pc and P3 latencies were correlated in time, both when the target or a distractor was reported. Notably, this link was weaker on distractor intrusion trials. This N2pc-P3 association is discussed with respect to the relationship between attention and access consciousness. Our results demonstrate that our novel method provides a valuable approach for assessing temporal links between two cognitive processes and their underlying modulating factors. This method allows to establish links and their modulator for any two time-series across all domains of the field (general-purpose MATLAB functions and a Python module are provided alongside this paper).

Hippocampal neurons code individual episodic memories in humans.

The hippocampus is an essential hub for episodic memory processing. However, how human hippocampal single neurons code multi-element associations remains unknown. In particular, it is debated whether each hippocampal neuron represents an invariant element within an episode or whether single neurons bind together all the elements of a discrete episodic memory. Here we provide evidence for the latter hypothesis. Using single-neuron recordings from a total of 30 participants, we show that individual neurons, which we term episode-specific neurons, code discrete episodic memories using either a rate code or a temporal firing code. These neurons were observed exclusively in the hippocampus. Importantly, these episode-specific neurons do not reflect the coding of a particular element in the episode (that is, concept or time). Instead, they code for the conjunction of the different elements that make up the episode.

Concealed identity information detection with pupillometry in rapid serial visual presentation.

The concealed information test (CIT) relies on bodily reactions to stimuli that are hidden in mind. However, people can use countermeasures, such as purposely focusing on irrelevant things, to confound the CIT. A new method designed to prevent countermeasures uses rapid serial visual presentation (RSVP) to present stimuli on the fringe of awareness. Previous studies that used RSVP in combination with electroencephalography (EEG) showed that participants exhibit a clear reaction to their real first name, even when they try to prevent such a reaction (i.e., when their name is concealed information). Because EEG is not easily applicable outside the laboratory, we investigated here whether pupil size, which is easier to measure, can also be used to detect concealed identity information. In our first study, participants adopted a fake name, and searched for this name in an RSVP task, while their pupil sizes were recorded. Apart from this fake name, their real name and a control name also appeared in the task. We found pupil dilation in response to the task-irrelevant real name, as compared to control names. However, while most participants showed this effect qualitatively, it was not statistically significant for most participants individually. In a second study, we preregistered the proof-of-concept methodology and replicated the original findings. Taken together, our results show that the current RSVP task with pupillometry can detect concealed identity information at a group level. Further development of the method is needed to create a valid and reliable concealed identity information detector at the individual level.

Branching Time Active Inference with Bayesian Filtering.

Branching time active inference is a framework proposing to look at planning as a form of Bayesian model expansion. Its root can be found in active inference, a neuroscientific framework widely used for brain modeling, as well as in Monte Carlo tree search, a method broadly applied in the reinforcement learning literature. Up to now, the inference of the latent variables was carried out by taking advantage of the flexibility offered by variational message passing, an iterative process that can be understood as sending messages along the edges of a factor graph. In this letter, we harness the efficiency of an alternative method for inference, Bayesian filtering, which does not require the iteration of the update equations until convergence of the variational free energy. Instead, this scheme alternates between two phases: integration of evidence and prediction of future states. Both phases can be performed efficiently, and this provides a forty times speedup over the state of the art.

Branching time active inference: Empirical study and complexity class analysis.

Active inference is a state-of-the-art framework for modelling the brain that explains a wide range of mechanisms such as habit formation, dopaminergic discharge and curiosity. However, recent implementations suffer from an exponential (space and time) complexity class when computing the prior over all the possible policies up to the time horizon. Fountas et al. (2020) used Monte Carlo tree search to address this problem, leading to very good results in two different tasks. Additionally, Champion et al. (2021a) proposed a tree search approach based on (temporal) structure learning. This was enabled by the development of a variational message passing approach to active inference (Champion, Bowman, Grzes, 2021), which enables compositional construction of Bayesian networks for active inference. However, this message passing tree search approach, which we call branching-time active inference (BTAI), has never been tested empirically. In this paper, we present an experimental study of the approach (Champion, Grzes, Bowman, 2021) in the context of a maze solving agent. In this context, we show that both improved prior preferences and deeper search help mitigate the vulnerability to local minima. Then, we compare BTAI to standard active inference (AcI) on a graph navigation task. We show that for small graphs, both BTAI and AcI successfully solve the task. For larger graphs, AcI exhibits an exponential (space) complexity class, making the approach intractable. However, BTAI explores the space of policies more efficiently, successfully scaling to larger graphs. Then, BTAI was compared to the POMCP algorithm (Silver and Veness, 2010) on the frozen lake environment. The experiments suggest that BTAI and the POMCP algorithm accumulate a similar amount of reward. Also, we describe when BTAI receives more rewards than the POMCP agent, and when the opposite is true. Finally, we compared BTAI to the approach of Fountas et al. (2020) on the dSprites dataset, and we discussed the pros and cons of each approach.

Branching Time Active Inference: The theory and its generality.

Over the last 10 to 15 years, active inference has helped to explain various brain mechanisms from habit formation to dopaminergic discharge and even modelling curiosity. However, the current implementations suffer from an exponential (space and time) complexity class when computing the prior over all the possible policies up to the time-horizon. Fountas et al. (2020) used Monte Carlo tree search to address this problem, leading to impressive results in two different tasks. In this paper, we present an alternative framework that aims to unify tree search and active inference by casting planning as a structure learning problem. Two tree search algorithms are then presented. The first propagates the expected free energy forward in time (i.e., towards the leaves), while the second propagates it backward (i.e., towards the root). Then, we demonstrate that forward and backward propagations are related to active inference and sophisticated inference, respectively, thereby clarifying the differences between those two planning strategies.

Modelling the simultaneous encoding/serial experience theory of the perceptual moment: a blink of meta-experience.

One way to understand a system is to explore how its behaviour degrades when it is overloaded. This approach can be applied to understanding conscious perception by presenting stimuli in rapid succession in the 'same' perceptual event/moment. In previous work, we have identified a striking dissociation during the perceptual moment, between what is encoded into working memory [Lag-1 sparing in the attentional blink (AB)] and what is consciously perceived (Lag-1 impairing in the experiential blink). This paper links this dissociation to what, taking inspiration from the metacognition literature, could be called meta-experience; i.e. how the ability to track and comment on one's visual experience with subjectivity ratings reflects objective performance. Specifically, we provide evidence that the information (in bits) associated with an encoding into working memory decouples from the experiential reflection upon that perceptual/encoding event and that this decoupling is largest when there is the greatest perceptual overload. This is the meta-experiential blink. Meta-experiential self-observation is common to many computational models, including connectionist interpretations of consciousness, Bayesian observers and the readout-enhanced simultaneous type/serial token (reSTST) model. We assess how our meta-experiential blink data could be modelled using the concept of self-observation, providing model fits to behavioural and electroencephalogram responses in the reSTST model. We discuss the implications of our computational modelling of parallel encoding but serial experience for theories of conscious perception. Specifically, we (i) inform theories of Lag-1 sparing during the AB and (ii) consider the implications for the global workspace theory of conscious perception and higher-order theories of consciousness.

Recovery after stroke: the severely impaired are a distinct group.

INTRODUCTION: Stroke causes different levels of impairment and the degree of recovery varies greatly between patients. The majority of recovery studies are biased towards patients with mild-to-moderate impairments, challenging a unified recovery process framework. Our aim was to develop a statistical framework to analyse recovery patterns in patients with severe and non-severe initial impairment and concurrently investigate whether they recovered differently. METHODS: We designed a Bayesian hierarchical model to estimate 3-6 months upper limb Fugl-Meyer (FM) scores after stroke. When focusing on the explanation of recovery patterns, we addressed confounds affecting previous recovery studies and considered patients with FM-initial scores <45 only. We systematically explored different FM-breakpoints between severe/non-severe patients (FM-initial=5-30). In model comparisons, we evaluated whether impairment-level-specific recovery patterns indeed existed. Finally, we estimated the out-of-sample prediction performance for patients across the entire initial impairment range. RESULTS: Recovery data was assembled from eight patient cohorts (n=489). Data were best modelled by incorporating two subgroups (breakpoint: FM-initial=10). Both subgroups recovered a comparable constant amount, but with different proportional components: severely affected patients recovered more the smaller their impairment, while non-severely affected patients recovered more the larger their initial impairment. Prediction of 3-6 months outcomes could be done with an R2=63.5% (95% CI=51.4% to 75.5%). CONCLUSIONS: Our work highlights the benefit of simultaneously modelling recovery of severely-to-non-severely impaired patients and demonstrates both shared and distinct recovery patterns. Our findings provide evidence that the severe/non-severe subdivision in recovery modelling is not an artefact of previous confounds. The presented out-of-sample prediction performance may serve as benchmark to evaluate promising biomarkers of stroke recovery.

The missing N1 or jittered P2: Electrophysiological correlates of pattern glare in the time and frequency domain.

Excessive sensitivity to certain visual stimuli (cortical hyperexcitability) is associated with a number of neurological disorders including migraine, epilepsy, multiple sclerosis, autism and possibly dyslexia. Others show disruptive sensitivity to visual stimuli with no other obvious pathology or symptom profile (visual stress) which can extend to discomfort and nausea. We used event-related potentials (ERPs) to explore the neural correlates of visual stress and headache proneness. We analysed ERPs in response to thick (0.37 cycles per degree [c/deg]), medium (3 c/deg) and thin (12 c/deg) gratings, using mass univariate analysis, considering three factors in the general population: headache proneness, visual stress and discomfort. We found relationships between ERP features and the headache and discomfort factors. Stimulus main effects were driven by the medium stimulus regardless of participant characteristics. Participants with high discomfort ratings had larger P1 components for the initial presentation of medium stimuli, suggesting initial cortical hyperexcitability that is later suppressed. The participants with high headache ratings showed atypical N1-P2 components for medium stripes relative to the other stimuli. This effect was present only after repeated stimulus presentation. These effects were also explored in the frequency domain, suggesting variations in intertrial theta band phase coherence. Our results suggest that discomfort and headache in response to striped stimuli are related to different neural processes; however, more exploration is needed to determine whether the results translate to a clinical migraine population.

Realizing Active Inference in Variational Message Passing: The Outcome-Blind Certainty Seeker.

Active inference is a state-of-the-art framework in neuroscience that offers a unified theory of brain function. It is also proposed as a framework for planning in AI. Unfortunately, the complex mathematics required to create new models can impede application of active inference in neuroscience and AI research. This letter addresses this problem by providing a complete mathematical treatment of the active inference framework in discrete time and state spaces and the derivation of the update equations for any new model. We leverage the theoretical connection between active inference and variational message passing as described by John Winn and Christopher M. Bishop in 2005. Since variational message passing is a well-defined methodology for deriving Bayesian belief update equations, this letter opens the door to advanced generative models for active inference. We show that using a fully factorized variational distribution simplifies the expected free energy, which furnishes priors over policies so that agents seek unambiguous states. Finally, we consider future extensions that support deep tree searches for sequential policy optimization based on structure learning and belief propagation.

Aging reduces EEG markers of recognition despite intact performance: Implications for forensic memory detection.

ERP-based forensic memory detection is based on the logic that guilty suspects will hold incriminating knowledge about crimes they have committed, and therefore should show parietal ERP positivities related to recognition when presented with reminders of their crimes. We predicted that such forensic memory detection might however be inaccurate in older adults, because of changes to recognition-related brain activity that occurs with aging. We measured both ERPs and EEG oscillations associated with episodic old/new recognition and forensic memory detection in 30 younger (age < 30) and 30 older (age > 65) adults. EEG oscillations were included as a complementary measure which is less sensitive to temporal variability and component overlap than ERPs. In line with predictions, recognition-related parietal ERP positivities were significantly reduced in the older compared to younger group in both tasks, despite highly similar behavioural performance. We also observed aging-related reductions in oscillatory markers of recognition in the forensic memory detection test, while the oscillatory effects associated with episodic recognition were similar across age groups. This pattern of results suggests that while both forensic memory detection and episodic recognition are accompanied by aging-induced reductions in parietal ERP positivities, these reductions may be caused by non-overlapping mechanisms across the two tasks. Our findings suggest that EEG-based forensic memory detection tests are less valid in older than younger populations, limiting their practical applications.

Breakthrough percepts of famous names.

Studies have shown that presenting own-name stimuli on the fringe of awareness in Rapid Serial Visual Presentation (RSVP) generates a P3 component and provides an accurate and countermeasure resistant method for detecting identity deception (Bowman et al., 2013, 2014). The current study investigates how effective this Fringe-P3 method is at detecting recognition of familiar name stimuli with lower salience (i.e., famous names) than own-name stimuli, as well as its accuracy with multi-item stimuli (i.e., first and second name pairs presented sequentially). The results demonstrated a highly significant ERP difference between famous and non-famous names at the group level and a detectable P3 for famous names for 86% of participants at the individual level. This demonstrates that the Fringe-P3 method can be used for detecting name stimuli other than own-names and for multi-item stimuli, thus further supporting the method's potential usefulness in forensic applications such as in detecting recognition of accomplices.

The Sync-Fire/deSync model: Modelling the reactivation of dynamic memories from cortical alpha oscillations.

We propose a neural network model to explore how humans can learn and accurately retrieve temporal sequences, such as melodies, movies, or other dynamic content. We identify target memories by their neural oscillatory signatures, as shown in recent human episodic memory paradigms. Our model comprises three plausible components for the binding of temporal content, where each component imposes unique limitations on the encoding and representation of that content. A cortical component actively represents sequences through the disruption of an intrinsically generated alpha rhythm, where a desynchronisation marks information-rich operations as the literature predicts. A binding component converts each event into a discrete index, enabling repetitions through a sparse encoding of events. A timing component - consisting of an oscillatory "ticking clock" made up of hierarchical synfire chains - discretely indexes a moment in time. By encoding the absolute timing between discretised events, we show how one can use cortical desynchronisations to dynamically detect unique temporal signatures as they are reactivated in the brain. We validate this model by simulating a series of events where sequences are uniquely identifiable by analysing phasic information, as several recent EEG/MEG studies have shown. As such, we show how one can encode and retrieve complete episodic memories where the quality of such memories is modulated by the following: alpha gate keepers to content representation; binding limitations that induce a blink in temporal perception; and nested oscillations that provide preferential learning phases in order to temporally sequence events.

Inflated Estimates of Proportional Recovery From Stroke: The Dangers of Mathematical Coupling and Compression to Ceiling.

The proportional recovery rule states that most survivors recover a fixed proportion (≈70%) of lost function after stroke. A strong (negative) correlation between the initial score and subsequent change (outcome minus initial; ie, recovery) is interpreted as empirical support for the proportional recovery rule. However, this rule has recently been critiqued, with a central observation being that the correlation of initial scores with change over time is confounded in the situations in which it is typically assessed. This critique has prompted reassessments of patients' behavioral trajectory following stroke in 2 prominent papers. The first of these, by van der Vliet et al presented an impressive modeling of upper limb deficits following stroke, which avoided the confounded correlation of initial scores with change. The second by Kundert et al reassessed the value of the proportional recovery rule, as classically formulated as the correlation between initial scores and change. They argued that while effective prediction of recovery trajectories of individual patients is not supported by the available evidence, group-level inferences about the existence of proportional recovery are reliable. In this article, we respond to the van der Vliet and Kundert papers by distilling the essence of the argument for why the classic assessment of proportional recovery is confounded. In this respect, we reemphasize the role of mathematical coupling and compression to ceiling in the confounded nature of the correlation of initial scores with change. We further argue that this confound will be present for both individual-level and group-level inference. We then focus on the difficulties that can arise from ceiling effects, even when initial scores are not being correlated with change/recovery. We conclude by emphasizing the need for new techniques to analyze recovery after stroke that are not confounded in the ways highlighted here.

How hot is the hot zone? Computational modelling clarifies the role of parietal and frontoparietal connectivity during anaesthetic-induced loss of consciousness.

In recent years, specific cortical networks have been proposed to be crucial for sustaining consciousness, including the posterior hot zone and frontoparietal resting state networks (RSN). Here, we computationally evaluate the relative contributions of three RSNs - the default mode network (DMN), the salience network (SAL), and the central executive network (CEN) - to consciousness and its loss during propofol anaesthesia. Specifically, we use dynamic causal modelling (DCM) of 10 min of high-density EEG recordings (N = 10, 4 males) obtained during behavioural responsiveness, unconsciousness and post-anaesthetic recovery to characterise differences in effective connectivity within frontal areas, the posterior 'hot zone', frontoparietal connections, and between-RSN connections. We estimate - for the first time - a large DCM model (LAR) of resting EEG, combining the three RSNs into a rich club of interconnectivity. Consistent with the hot zone theory, our findings demonstrate reductions in inter-RSN connectivity in the parietal cortex. Within the DMN itself, the strongest reductions are in feed-forward frontoparietal and parietal connections at the precuneus node. Within the SAL and CEN, loss of consciousness generates small increases in bidirectional connectivity. Using novel DCM leave-one-out cross-validation, we show that the most consistent out-of-sample predictions of the state of consciousness come from a key set of frontoparietal connections. This finding also generalises to unseen data collected during post-anaesthetic recovery. Our findings provide new, computational evidence for the importance of the posterior hot zone in explaining the loss of consciousness, highlighting also the distinct role of frontoparietal connectivity in underpinning conscious responsiveness, and consequently, suggest a dissociation between the mechanisms most prominently associated with explaining the contrast between conscious awareness and unconsciousness, and those maintaining consciousness.

Damage to Broca's area does not contribute to long-term speech production outcome after stroke.

Broca's area in the posterior half of the left inferior frontal gyrus has long been thought to be critical for speech production. The current view is that long-term speech production outcome in patients with Broca's area damage is best explained by the combination of damage to Broca's area and neighbouring regions including the underlying white matter, which was also damaged in Paul Broca's two historic cases. Here, we dissociate the effect of damage to Broca's area from the effect of damage to surrounding areas by studying long-term speech production outcome in 134 stroke survivors with relatively circumscribed left frontal lobe lesions that spared posterior speech production areas in lateral inferior parietal and superior temporal association cortices. Collectively, these patients had varying degrees of damage to one or more of nine atlas-based grey or white matter regions: Brodmann areas 44 and 45 (together known as Broca's area), ventral premotor cortex, primary motor cortex, insula, putamen, the anterior segment of the arcuate fasciculus, uncinate fasciculus and frontal aslant tract. Spoken picture description scores from the Comprehensive Aphasia Test were used as the outcome measure. Multiple regression analyses allowed us to tease apart the contribution of other variables influencing speech production abilities such as total lesion volume and time post-stroke. We found that, in our sample of patients with left frontal damage, long-term speech production impairments (lasting beyond 3 months post-stroke) were solely predicted by the degree of damage to white matter, directly above the insula, in the vicinity of the anterior part of the arcuate fasciculus, with no contribution from the degree of damage to Broca's area (as confirmed with Bayesian statistics). The effect of white matter damage cannot be explained by a disconnection of Broca's area, because speech production scores were worse after damage to the anterior arcuate fasciculus with relative sparing of Broca's area than after damage to Broca's area with relative sparing of the anterior arcuate fasciculus. Our findings provide evidence for three novel conclusions: (i) Broca's area damage does not contribute to long-term speech production outcome after left frontal lobe strokes; (ii) persistent speech production impairments after damage to the anterior arcuate fasciculus cannot be explained by a disconnection of Broca's area; and (iii) the prior association between persistent speech production impairments and Broca's area damage can be explained by co-occurring white matter damage, above the insula, in the vicinity of the anterior part of the arcuate fasciculus.