Session I: 29 June, 13.30-14.45 (CEST)

Harmonicity plays a fundamental role in auditory perception because most natural sounds exhibit harmonic structure, making them more predictable and perceptually salient. Thus, harmonicity may support the formation of precise auditory predictions and robust neural responses to expectancy violations. Musical training may further modulate these processes, as prolonged experience with structured sound patterns refines internal predictive models and enhances sensitivity to statistical regularities in melodic sequences. Building on this framework, the present study investigates how harmonicity influences neural responses to melodic expectancy violations and whether musical training modulates this effect. We examine whether less probable tones in harmonic melodies elicit stronger brain potentials than comparable tones in inharmonic melodies, and whether musicians show enhanced neural responses relative to non-musicians. Understanding these mechanisms may clarify how the brain encodes auditory uncertainty and pitch structure in music perception. Hearing thresholds will be established using the Hughson–Westlake procedure, after which participants will undergo an EEG recording. During the experiment, participants will listen to monophonic melodies derived from instrumental works by Johann Sebastian Bach, divided into 150-second blocks. 50% of the melodies will be acoustically manipulated using the TANDEM-STRAIGHT algorithm to create inharmonic versions. Across 30 blocks (15 per condition), participants will rate the familiarity of each melody while neural responses to less probable tones, as estimated by the IDyOM model, are recorded and compared between harmonic and inharmonic conditions, as well as between musicians and non-musicians We expect that less probable tones in harmonic melodies will elicit stronger event-related potentials than those in inharmonic melodies, and that musically trained participants will demonstrate greater neural sensitivity to melodic expectancy violations. Data collection is currently in progress. The findings are expected to advance understanding of how the brain encodes auditory uncertainty on fine temporal scales relevant to music and speech perception. This research may also deepen knowledge of pitch processing mechanisms critical for language, music cognition, and broader auditory functions.

In animal models, experimental evidence has shown that auditory processing regions receive inputs from somatosensory processing regions, both in the ascending auditory pathway (Morest and Oliver, 1984; Jasmin et al., 2019) and in the cortex (Smiley et al., 2007; Hackett et al., 2007). In humans, caudal auditory cortex is involved in multisensory processing (Foxe et al., 2002) and supports spatial transformations in dorsal auditory streams (Rauschecker and Scott, 2009). Following theoretical work from Niven and Scott (2021), we did a functional neuroimaging experiment with 34 neurotypical individuals while they passively listened to human-generated action sounds (e.g., brushing hair, crinkling a crisp bag) that were behaviorally shown to induce tingling touch percepts. Auditory stimuli differed in a) being positively or negatively valenced and b) the degree to which they contained spatial information (i.e., spatially rich vs. spatially impoverished). We hypothesized that sounds with more (vs. less) spatial information would induce greater activity in bilateral posterior caudal auditory fields (Callan et al., 2013), and that activity in this region would be further modulated by valence. We also hypothesized that positively valenced sounds would induce activity in regions similar to those activated by pleasant music (Mas-Herrero et al., 2021), and that negatively valenced sounds would induce activity in misophonia-related regions (Kumar et al., 2017). Preliminary results from whole-brain analyses show that negatively (vs. positively) valenced sounds evoke widespread activation bilaterally in MTG and STG extending into the temporal pole, similar to activation previously found for unpleasant music and aversive sound (Zald and Pardo, 2002; Koelsch et al., 2006). Analysis regarding spatial information processing is ongoing. Implications of this work include: better understanding the functional role of human caudal auditory cortex, and developing a broader framework of pleasant sound-induced touch (i.e., bridging research from musical chills and ASMR).

tba

A joint neural–behavioral state can be defined as a stable neural activity pattern that evolves over time with behavioral demands and environmental inputs. Changes in neural dynamics should therefore inform changes in behavior, making time-resolved analyses of neural activity a powerful window into how cognitive processes are realized in biological and artificial systems. Most existing approaches to studying neural dynamics depend on training data, such as predefined behavioral labels. This reliance limits their applicability in naturalistic settings, where distinct behavioral epochs are not clearly defined. To overcome this limitation, we develop Universal Neural-state Identification through Temporal Embeddings (UNITE), a theory-driven, unsupervised method that detects state transitions directly from population dynamics. UNITE exploits the principle that neural dynamics vary minimally within a state but change sharply across transitions. It combines an efficient temporal embedding (based on Koopman operators) with a similarity analysis of nonlinear dynamics to identify these transitions in a time-resolved manner. UNITE recovers behaviorally relevant states in both biological and artificial systems. Applied to spiking activity from macaque area V4 during a selective attention task and to recurrent neural networks (RNNs) trained on a context-dependent decision-making task, UNITE identifies transitions corresponding to attention and decision epochs without using behavioral labels. Finally, UNITE detects behaviorally relevant states in naturalistic settings. When applied to RNNs trained to locate odor sources in simulated turbulent environments, it identifies search-like phases—tracking, loss, and recovery—directly from unlabeled population activity. Overall, our results demonstrate that UNITE can identify behaviorally relevant state transitions not only in classic epoch-based cognitive tasks but also in continuous naturalistic tasks, both in biological and artificial agents, and in an unsupervised manner.

The brain continuously generates predictions about upcoming sensory input, allowing perception to be shaped not only by incoming signals but also by contextual expectations. In audition, processing can begin prior to sound onset, as context prepares neural states for upcoming input. Here, we investigated how the processing, and as a results sensitivity to sensory deviants is modulated by naturally fluctuating contextual predictability in music. Twenty-five participants listened to musical melodies containing infrequent changes in sound location while EEG was recorded. In a behavioural task, participants indicated location changes via button press when cued. We quantified contextual predictability using the IDyOMS model (Information Dynamics of Multidimensional Sequences; Harley, 2022; Pearce, 2005), deriving entropy (precision of expectations) and information content (unexpectedness) for each note. We quantified the accuracy of location deviant detection as a function of contextual unexpectedness. Moreover, we modelled brain response to the location change using temporal response function (TRF), to observe the modulation of neural tracking of location deviant by contextual expectedness. Increased unexpectedness reduced location change detection accuracy (information content: OR = 0.85, 95% CI [0.74, 0.98], p = .027; entropy: OR = 0.87, 95% CI [0.77, 0.99], p = .028). While increasing unexpectedness alone enlarges the amplitude of N100 component (Page Trend Test = 7506.0, p = 0.005), the combination of high unexpectedness and location change showed a suppressed response, indicating a sub-additive interaction effect on the N100 amplitude (t(17) = 4.07, p < 0.001). We demonstrate that statistical learning of musical context modulates the early neural encoding of acoustic features, with contextual predictors interacting with sensitivity to the sensory deviants. This provides evidence that context and building predictions about incoming sounds shapes not only the late perception of sounds but their fundamental, neural representation at early processing stages.

Spatial processing is one of the most fundamental forms of human thought. It is not only used for navigating physical space, but also appears in abstract domains. People use spatial metaphors in many abstract concepts, for example a "high mood" or "low energy." However, the extent to which spatial cognition supports other high-level domains remains unclear. The goal of this project is to investigate whether people rely on non-verbal spatial reasoning when engaging in mathematical and musical thinking. These domains are well suited for exploring this topic because both are non-linguistic abilities with spatial representations and their notations reflect spatial structure. The aim is to investigate whether similar neural representations and transformations underlie cognition in these domains and in spatial processing itself. To explore this, the study uses a set of three carefully designed experimental tasks that vary in complexity and specificity, directly comparing spatial, mathematical, and musical processing. By analyzing both behavioral and neural data, the study checks whether similar cognitive and neural mechanisms underlie all three domains. The findings could help clarify the role of spatial cognition in abstract thought, offering new insights for spatial modes of reasoning and cognition. Additionally, the study has the potential to clarify the cognitive and neural relationships between math and music.

The taking of turns during conversation reaches a high temporal precision. Gaps between speakers average at around 200ms, which is less than the time that the brain requires to prepare speech production. Taking turns thus must necessitate the anticipation of turn completion points. It has previously been found that the brain affords temporal prediction through the synchronization of neural activity with rhythmic modulations in the speech stream at the level of syllables and intonational phrases. In a previous study we observed that rhythmicity at the syllabic rate preceding a turn boundary accelerated reaction times. To explore whether this effect was driven by higher neural synchronization to the more rhythmic stimuli we conducted an MEG study where 32 German-speaking participants played a Guess-Who game with two digital interlocutors. We are analyzing synchronization between the speech envelope and the neural signal before the turn boundary as a predictor of participants' response times. Data analysis for this study is currently underway.

The human lineage is the only one capable of communication through complex spoken language. Language is thus a unique and integral part of human behavior and cognition prominently involving several brain structures including but not limited to Broca’s area, Wernicke’s area and the Arcuate Fasciculus dorsal white matter fiber tract (AF). While neuroimaging and cytoarchitectonic studies have identified homologous regions of Broca’s area in non-human primates, these species still lack the capacity to produce language in the manner humans do. Although the microstructural and molecular organization of Broca’s area has been characterized in humans, it remains unclear to what extent this organization is present in our closest evolutionary relatives. This ongoing gap underscores the need for a detailed comparative analysis of these brain regions to understand the unique neuro-evolutionary foundations of Language. In this project we address this gap by employing a multimodal comparative analysis of Broca’s area in chimpanzees using histological methods. To this end, we investigate the cytoarchitectonic and molecular organization of Broca’s area in post mortem chimpanzee brains with reference to human data. In particular, we examine neurotransmitter receptor distribution, an approach that has not previously been performed in Broca’s area in the chimpanzee brain.

Cognitive impairment is a central feature of many neurological disorders and contributes substantially to disability and reduced quality of life. Understanding how individual differences in cognitive performance relate to brain structure is therefore an important goal of cognitive neuroimaging. Structural MRI provides a non-invasive window into neuroanatomical variation, yet conventional analyses often rely on predefined regions and may fail to capture distributed structural patterns associated with cognition. Recent advances in deep learning offer a data-driven approach to identify complex brain–cognition relationships directly from neuroimaging data . We investigated whether deep learning models can identify structural brain signatures associated with cognitive phenotypes using T1-weighted MRI from the ID1000 dataset of the Amsterdam Open MRI Collection (AOMIC) . A 3D CNN was trained to predict multiple psychometric measures including fluid intelligence, crystallized intelligence, memory performance, and personality traits from the NEO Five-Factor model. Model performance was evaluated using regression metrics and correlations between predicted and observed scores. Gradient-based explainability methods were applied to identify brain regions contributing to the predictions. Model predictions showed statistically significant associations with observed cognitive scores across several domains. The strongest correlation was observed for total intelligence (IST_intelligence_total; r = 0.27, p < 0.001), followed by memory performance (IST_memory; r = 0.20, p < 0.001). Smaller but significant associations were detected for crystallized intelligence (IST_crystallised; r = 0.14, p < 0.05) and fluid intelligence (IST_fluid; r = 0.12, p < 0.05). Among personality traits, only neuroticism showed an association (NEO_N; r = 0.12, p < 0.05). Explainability analyses highlighted contributions from medial and subcortical brain regions for intelligence-related measures and frontal–parietal regions for neuroticism. These findings demonstrate that deep learning applied to structural MRI can capture distributed neuroanatomical patterns associated with cognitive variation. Such approaches may provide new tools for studying the structural basis of cognition and developing quantitative neuroimaging markers relevant to cognitive decline. Future work will extend this framework by incorporating multimodal MRI data including diffusion and functional imaging.

Models trained to predict probability distributions over sequences generated by a latent world appear to develop internal representations of that world and its rules. This suggests a deep connection between the structure of the world and the geometry of probability distributions. In order to understand this link more precisely, we use a minimal stochastic process as a controlled setting: constrained random walks on a two-dimensional lattice that must reach a fixed endpoint after a predetermined number of steps. Optimal prediction of this process solely depends on a sufficient vector determined by the walker's position relative to the target and the remaining time horizon; in other words, the probability distributions are parametrized by the world's geometry. We train decoder-only transformers on prefixes drawn from the exact distribution of these constrained walks and analyze their internal representations. We find that hidden activations strongly align with the ground-truth sufficient vectors derived analytically. These results illustrate a concrete mechanism by which world-model–like representations can arise from the predictive structure of the data itself. Although demonstrated in a simplified toy system, the analysis suggests that geometric representations supporting optimal prediction may provide a useful lens for studying how neural networks internalize grammatical and structural constraints.

Neural networks can accurately forecast complex dynamical systems, yet how they internally represent underlying latent geometry remains poorly understood. We study neural forecasters through the lens of representational alignment, introducing anchor-based, geometry-agnostic relative embeddings that remove rotational and scaling ambiguities in latent spaces. Applying this framework across seven canonical dynamical systems—ranging from periodic to chaotic—we reveal reproducible family-level structure: multilayer perceptrons align with other MLPs, recurrent networks with RNNs, while transformers and echo-state networks achieve strong forecasts despite weaker alignment. Alignment generally correlates with forecasting accuracy, yet high accuracy can coexist with low alignment. Relative geometry thus provides a simple, reproducible foundation for comparing how model families internalize and represent dynamical structure.

Investigating human brain structure at the columnar and laminar level is possible with 7 Tesla MRI. However, the required sub-millimeter resolution typically results in lengthy scan times and higher susceptibility to motion artifacts. While traditional reconstruction methods like LORAKS facilitate accelerated acquisitions, they face limitations at high acceleration factors and are computationally expensive. Deep learning (DL) approaches may enable significantly higher acceleration. To explore this, we compared LORAKS and a series of supervised DL networks on ultra-high resolution MRI. To provide the fully-sampled ground truth required for supervised DL, we utilized 15 post-mortem chimpanzee brain MRI scans acquired at 7T (0.3 mm isotropic, PD/T1-weighted FLASH), ethically sourced as part of the Evolution of Brain Connectivity (EBC) project. Fully-sampled 3D k-space volumes were converted into 2D stacks and retrospectively undersampled 8-fold using variable-density Poisson-disc masks. We benchmarked 8 DL architectures implemented within the ATOMMIC framework, including image-domain (U-Net), cross-domain (KiKiNet), and physics-informed unrolled networks (VarNet, MoDL, JointICNet, RVN, RIM, and LPDNet). Data were split 7:1:2 (training/validation/testing), and networks were trained using a combined L1 and SSIM loss under a supervised learning pipeline. Reconstruction quality was assessed via NMSE, PSNR, and SSIM metrics. Quantitative analysis favored DL methods over traditional approaches. While U-Net yielded the highest PSNR/NMSE scores, visual inspection revealed significant over-smoothing and loss of microstructural details. VarNet achieved the best balance, attaining the highest SSIM (0.869) while faithfully preserving anatomical details. Conversely, LORAKS visually retained intricate anatomical details in low-CNR regions but at the cost of significant noise amplification. At 8-fold acceleration, LORAKS is constrained by its linear assumptions, struggling to resolve aliasing from extremely sparse data. This resulted in lower quantitative scores (SSIM: 0.542) compared to VarNet, which effectively leverages non-linear priors to resolve aliasing while offering faster inference. These findings confirm the feasibility of DL for accelerating ultra-high resolution MRI. Future work will focus on translating this framework to in vivo human neuroimaging and developing quality metrics that better reflect visual texture fidelity.

The Social Brain Hypothesis proposes that the demands of maintaining cohesion and navigating increasingly complex social groups were major drivers of cognitive evolution. Consistent with this account, both inter-species and inter-individual variation in structural properties of fronto-temporal regions (MPFC, TPJ/MSTS), involved in social cognition, have been shown to covary with diverse metrics of sociality in humans and non-human primates. More recently, the cerebellum has emerged as a structure implicated in social cognitive processing. In particular, the posterior lateral cerebellum in humans has undergone marked evolutionary expansion and is thought to contribute to social cognition through its interactions with the fronto-temporal social cognition network. This raises the possibility that cerebellar expansion was shaped by evolutionary pressures similar to those acting on fronto-temporal cortices. Building on this framework, my PhD will examine how fronto-temporo-cerebellar circuits covary with measures of social complexity in non-human primates.

Background: Functional neuroimaging studies consistently show co-activation of brain regions such as the anterior insula during tasks involving interoceptive and empathic processing. Growing evidence suggests a link between experiencing one’s own emotions and resonating with others’, positioning interoception as foundational to emotional experience and social behavior. This meta-analytic study investigates the neural convergence between empathy and interoception to delineate a brain network that integrates the two domains. Methods: We conducted two separate systematic literature searches for functional neuroimaging studies related to empathy and interoception. Using seed-based d mapping, we performed voxel-wise random-effects meta-analyses for each domain to identify consistent patterns of brain activation. We also divided the two faculties into their respective subdomains (e.g., observing pain, cardioception) and conducted additional analyses for each subgroup. We then applied a multimodal conjunction analysis to determine overlapping neural regions involved in both empathy and interoception. Results: Conjunction analysis of neural activity related to empathy and interoception revealed three key clusters of overlapping activation. The largest cluster revealed the right insula extending to the inferior frontal gyrus. The second largest cluster was found in the left hemisphere involving the left insula. The third cluster encompassed the right supplementary motor area extending to the left supplementary motor area. Conclusions: The observed neural overlap within the anterior insula, the inferior frontal gyrus, and supplementary motor regions suggests that empathy is deeply rooted in embodied processes, highlighting the integration of bodily states as a foundation for social interaction. This study provides grounds to further examine the direct role of interoception in shaping bodily self-awareness and resonating with others' emotions.

Introduction The Subiculum (Sub) is known as the output layer of the hippocampal formation, and contains boundary vector cells (BVCs), firing for boundaries at specific allocentric directions and distances (Lever et al., 2009). More recently it has been shown that vector cells in the distal Sub (dSub) can exhibit traces that persists for hours after boundary/object removal (Poulter et al., 2021). Prior studies suggest that such traces can be evoked by place cells (PCs), which index boundary presence at encoding. However, an account of proximo-distal differences remains absent. Here we propose a circuit-level model accounting for vector trace cell (VTC) coding. Methods: In our model, dSub neurons receive input from either direct sensory information (BVCs in proximal Sub) or mnemonic information (PCs in CA1), and the mismatch between these inputs updates CA1-dSub synapses. The inserted cues affect distal and proximal CA1 and their corresponding dSub units. Results: Our model accounts for known VTC properties, including: (i) the distribution of VTCs along the proximodistal axis, (ii) the percentage of VTCs across different cue types, and (iii) hours-long persistence of vector trace. (iv) By enriching CA1 representations, our model explains object-centered population coding in CA1 (Vandrey et al., 2021). (v) VTCs have longer tuning distances after cue removal. Discussion: Our study suggests that the mismatch detection for associative memory updating explains findings in the CA1-Sub pathway, predicting Sub’s functions in coordinating spatial encoding and memory retrieval. Our study provides a potential framework to extend the standard hippocampal model with a Sub component.

In natural conversations, humans constantly infer and respond to others’ feelings (empathy) and perspectives (mentalizing, also Theory of Mind/ToM). However, neural correlates of these psychological processes in real-time social interactions remain under-investigated. To gain a better understanding of neural and psychological activities in social contexts, we adapted an established EmpaToM paradigm and developed a novel interactive EmpaToM (iEmpaToM) paradigm. During the experiment, we implemented the hyperscanning technique with functional near-infrared spectroscopy (fNIRS) to simultaneously record real-time brain activities in dyads of interacting participants. The iEmpaToM paradigm includes two lab visits. First, participants prepare autobiographical stories of different emotional valence and cognitive demands with the help of experimenters. There are four conditions: high need for empathy but low need for mentalizing; low need for empathy but high need for mentalizing; high need for both; and low need for both. In the second visit, participants receive hyperscanning with a randomly assigned conversational partner. They complete a total of 16 trials, four for each condition, and take turns as the narrator. In each trial, the narrator tells the story for 35 seconds; the dyad then has a 60-second free conversation. For each participant, we use a 16x23 montage to record brain activity in three regions of interest (ROIs): the temporal parietal junction, superior temporal sulcus, and medial prefrontal cortex. This forms a total of 50 channels, including eight short-distance channels. The preliminary analysis included 12 dyads of German-speaking adults. Data preprocessing includes data type conversion, motion correction, spatial filtering, and block averaging. A linear mixed-effects ANOVA modeled the change in oxygenated hemoglobin during listening phases, including ROIs, conditions, and story phases (beginning, middle, and ending). The dyadic effect was controlled for. There was a significant effect of ROIs (F(2, 828) = 3.89, p = .021). This effect was driven by activity in the TPJ. The future data analysis plan includes analyses of interpersonal synchrony and brain-behavior correlations. Overall, iEmpaToM is a novel, interactive paradigm suitable for hyperscanning. It significantly improves ecological validity in social neuroscience research and has great potential to deepen our understanding of neural correlates underlying human social interaction.

Cognitive flexibility is commonly described as the ability to shift between different cognitive strategies to adapt to changing tasks or environments. It, hence, plays a vital role in problem-solving, creativity, and adaptive decision-making. Building on the foundational work of Boot et al. (2017), which explored the role of dopaminergic modulation and cognitive flexibility in creative cognition, our study aims to update these insights in light of developments over the past eight years. Initially a scoping review, we recently transitioned to a rapid review methodology to better suit our more focused research question and project timeline. The shift also allowed us to refine our inclusion criteria: we include only peer-reviewed, English-language primary research published from 2018 onwards, conducted with neuroimaging or brain stimulation methods in non-clinical human populations. The search was conducted across PsycArticles, PubMed, and Web of Science using the following terms: ((nucleus accumbens) OR (caudate nucleus) OR (caudatus nucleus) OR (putamen) OR (globus pallidus) OR (olfactory tubercle) OR (striat*) AND (dopamin*)) AND ((cognitive flexibility) OR (creativ*)). By synthesizing recent neuroimaging and brain stimulation studies, this review aims to provide a comprehensive understanding of how striatal regions and dopaminergic pathways contribute to cognitive flexibility and creativity and to inform methodological approaches for future experimental studies.

Cognitive aging is characterized by de differentiation, whereby previously distinct abilities become increasingly interrelated. Whether a similar process occurs at the neural level remains unclear due to limited longitudinal evidence. We investigated structural brain de differentiation by examining coordinated regional volume change in two longitudinal cohorts of healthy adults assessed over five years (N = 87) and two years (N = 167). Across both samples, multiple cortical and subcortical regions were manually traced with high reliability and analyzed using latent change score models to estimate baseline volumes and longitudinal change free of measurement error. Multivariate models quantified covariance among regional baseline levels and change scores, and tested the influence of chronological age, vascular and metabolic risk, and APOE ε4 status. Baseline regional volumes were positively correlated across all regions in both cohorts, indicating shared structural characteristics. Baseline change associations were generally weak. Crucially, longitudinal changes showed positive change change correlations, supporting structural de differentiation. Evidence was strongest in the five year cohort, where moderate coupling emerged among frontal, temporo limbic, and cerebellar regions, suggesting coordinated shrinkage across distributed systems. In the two year cohort, change change correlations were positive but smaller, likely reflecting reduced sensitivity to detect shared decline over shorter intervals. Chronological age and vascular metabolic risk explained substantial variance in baseline volumes and individual rates of shrinkage, particularly in frontal and neostriatal regions. However, adjustment for these factors had minimal impact on the covariance structure of regional change. APOE ε4 status showed negligible effects. These findings indicate that coordinated structural decline represents a general property of normative brain aging rather than a pattern driven primarily by systemic health differences. Using longitudinal structural equation modeling across distinct samples and intervals, this study provides evidence that neural de differentiation manifests as shared regional shrinkage over time, especially across longer follow up periods.

Atrial fibrillation (AF) disrupts normal cardiac afferent signalling to the brain, and emerging evidence suggests that this disruption may contribute to cognitive impairment commonly observed in this population. We hypothesise that successful catheter ablation, by restoring more regular cardiac signalling, will lead to measurable improvements in neurophysiological markers of cardiac interoception and cognitive performance.

Methods:

Participants with paroxysmal or persistent AF undergoing catheter ablation are assessed at three time points: pre-ablation (V1), immediately post-ablation (V2), and six-month follow-up (V3). At each visit, participants complete cognitive testing and undergo EEG recording to derive heartbeat-evoked potentials (HEPs), an index of cortical processing of cardiac signals. Structural and functional MRI data are also acquired at V1 and V3. The primary comparisons are between V1 and V3. Preliminary data from the first 20 participants are presented.

Preliminary results:

In the current sample, MoCA scores showed an observable trend towards improvement from V1 to V3. HEP amplitudes also changed following ablation, suggesting that restoration of sinus rhythm may modulate cortical processing of cardiac afferent signals. Full statistical analyses will be conducted once the target sample has been completed (n=50).

The cerebellum’s involvement in cognitive functions is increasingly recognized, yet its developmental contribution to cognition remains poorly understood. The cerebellum undergoes rapid development in early life, paralleling major cognitive and behavioral changes. Although clinical studies have linked early cerebellar disruptions to profound developmental deficits, it remains largely unclear how typical cerebellar maturation supports the development of cognitive functions and how it interacts with broader brain development. Here, we apply a normative modeling framework to map cerebellar volumetric growth from infancy to young adulthood (N = 751; ages 1-21 years). Using lobular and functional cerebellar parcellations, we comprehensively characterize typical cerebellar development and examine how it aligns with cerebral development and behavioral outcomes. Across parcellations, posterior higher association areas consistently show steeper growth trajectories than anterior sensorimotor areas. Cerebellar and cerebral areas with similar functional roles demonstrate coordinated maturation, and volumetric growth in the posterior cerebellum relates to individual differences in socio-linguistic behaviors. These findings establish a comprehensive reference for typical cerebellar development, highlight cerebellar co-maturation with the cerebral cortex, and underscore the cerebellum’s role in supporting emerging higher cognitive functions.

Multi-site psychiatric classification studies indicate moderate to high accuracy; nevertheless, the extent to which these estimates represent illness differentiation or confounding variables is hardly evaluated after training. We developed a four-class classifier using postmortem brain genetic expression data and used principal component analysis to ascertain the signal accessible by the classifier. We included nine GEO datasets, including n=607 from 471 patients: 171 with (SCZ), 134 with (BD), 75 with (MDD), and 227 (CTL); sourced from the prefrontal cortex, hippocampus, and striatum, using Affymetrix microarray with 10,678 probes. NeuroCombat (Fortin et al. 2018) eliminated batch effects across two platforms. SHAP-based consensus feature selection during cross-validation folds (Parvandeh et al. 2020) revealed 100 stable probes. XGBoost attained an F1 Macro score of 0.716 during 5-fold cross-validation. PCA indicated that brain region, rather than diagnosis, accounts for the majority of variance: PC1 (35.4%) is significantly correlated with brain region (F=1269, p<0.001) but not with diagnosis (F=0.24, p=0.86). Diagnosis is shown to be weakly significant in PC2 (6.8%, F=4.3, p=0.005), confounded by age and sex. Further confounding factors include the application of neuroCombat before to cross-validation, allowing test samples to impact batch parameters (Nygaard et al. 2016; Marzi et al. 2024); unregressed tissue pH influencing 24.7% of gene expression (Miyahara et al. 2023); and the absence of cross-dataset validation. Per-class SHAP analysis revealed potential disease-related patterns. Seven out of twenty leading probes were validated in published studies; thirteen needed substantiation and are potential candidates for confound-driven selection.We are executing corrections: batch correction inside cross-validation using neuroHarmonize (Pomponio et al. 2020), covariate regression for brain region, age, sex, and pH, Leave-One-Dataset-Out validation, and an analysis focused just on the prefrontal cortex (414 samples, 20,049 probes) to eliminate confounding by brain area. The revised results will be shown on the poster. These hurdles are analogous to multi-site neuroimaging classification, where ComBat encounters similar leakage and regional heterogeneity issues (Rosenblatt et al. 2024).

Millions of people suffer from long-term symptoms following COVID-19 infections, collectively known as post-COVID syndrome. Olfactory dysfunction is among its most prevalent and persistent manifestations. However, the interplay between olfactory and other symptom domains, inter-individual variability, and effective treatments remain poorly understood. In this cross-sectional observational study, we employed an ML-based analysis combining LASSO variable selection with voting across multiply imputed datasets and random forest variable importance ranking. These methods were applied to a comprehensive set of cognitive, neuropsychiatric, and inflammatory parameters. Olfactory function was assessed using the standardized Sniffin' Sticks test battery, and patients were recruited through the Post-COVID Center Erlangen over a period of 18 months. Hedonic perception was assessed using a visual analogue scale rating procedure during the odor identification task of the Sniffin' Sticks test, allowing examination of the upper hedonic range in hyposmic versus normosmic-range patients. We included 142 patients with confirmed post-COVID diagnosis in the final analysis. Cognitive domains, especially verbal fluency, emerged as the most robust predictors of reduced olfactory function. Additionally, hyposmic patients showed a reduced upper hedonic range, with significantly lower average ratings for complex odors including garlic and clove, which are among the most frequently reported parosmia triggers. These findings suggest common disruptions in olfactory and cognitive neural processing that go beyond the hypothesis of peripheral damage, consistent with the dense connectivity between piriform cortex, OFC, and limbic structures involved in both olfactory and cognitive processing. These results have implications for olfactory training, currently the best-evidenced rehabilitation approach, raising the hypothesis that it may yield broader cognitive benefits. A planned study examining smartphone-based smell training as a scalable rehabilitation approach will further address the translational potential of these findings. Neuroimaging studies tracking long-term changes in olfactory and cognitive networks and possible modifications with olfactory training could inform targeted rehabilitation strategies. In particular, the link between verbal fluency and olfaction identified here points toward frontotemporal and OFC network integrity as candidate targets for investigation.

This longitudinal study will investigate the relationships between sleep, nutrition, physical activity, and mental well-being in adults. Particular focus will be placed on energy intake, meal timing, nutrient composition, sleep quality and regularity, nutritional choices, and magnesium intake. Participants will be monitored over four weeks in their home environment. Questionnaires will provide information on daily lifestyle and wellbeing, while a wearable smartwatch will be used to assess physical activity and collect nocturnal heart rate variability (HRV) data. Sleep activity will be recorded for one week using wearable electroencephalography (EEG), with one night including simultaneous full polysomnography to validate wearable recordings. Daily food and beverage intake will be tracked using a smartphone application specifically designed for the study. By combining objective physiological measurements with detailed behavioral monitoring, the study aims to provide insight into the complex interactions between sleep, nutrition, physical activity, and wellbeing in everyday life.