Publications

Summative assessments (SAs) and formative assessments (FAs) fulfil complementary functions in the educational endeavour. SAs measure knowledge at the end of a unit in a standardised, high-stakes setting, while FAs evaluate student performance during daily classroom activities to tailor feedback and instruction. Computer-based FA (CBFA) systems enable collecting unprecedented amounts of data objectively and with minimal disruption for students, under conditions that more closely resemble real-life behaviour. Given concerns about student stress and ecological validity associated with SAs, potential biases in teacher judgements, and the high burden entailed by traditional classroom assessments, we investigated whether and how well FA outcomes can predict SA outcomes. Specifically, we estimated student abilities in a large sample of children evaluated at different time points during compulsory schooling and performed a systematic comparison of regression models trained to predict SA abilities on different subsets of features derived from FA abilities and auxiliary variables. A model that included mean abilities in different competence domains performed best, accounting for a considerable proportion of variance (30–48\%), although this was still below that explained by past SA measures. Most predictive FA features generally corresponded to abilities from the same or a similar competence domain as the predicted SA ability. We report systematic model biases that would warrant consideration when using the models for decision-making. Our findings provide valuable insights into how learning progress connects to future achievement, which can help teachers adapt instruction earlier and inform policies to reduce reliance on high-stakes testing.

School curricula guide the daily learning activities of millions of students. They embody the understanding of the education experts who designed them of how to organize the knowledge that students should acquire in a way that is optimal for learning. This can be viewed as a learning ’theory’ which is, nevertheless, rarely put to the test. Here, we model a data set obtained from a Computer-Based Formative Assessment system used by thousands of students. The student-item response matrix is highly sparse and admits a natural representation as a bipartite graph, in which nodes stand for students or items and an edge between a student and an item represents a response of the student to that item. To predict unobserved edge labels (correct/incorrect responses) we resort to a graph neural network (GNN), a machine learning method for graph-structured data. Nodes and edges are represented as multidimensional embeddings. After fitting the model, the learned item embeddings reflect properties of the curriculum, such as item difficulty and the structure of school subject domains and competences. Simulations show that the GNN is particularly advantageous over a classical model when group patterns are present in the connections between students and items, such that students from a particular group have a higher probability of successfully answering items from a specific set. In sum, important aspects of the structure of the school curriculum are reflected in response patterns from educational assessments and can be partially retrieved by our graph-based neural model.

Introduction: Delay discounting (DD), the preference for smaller and sooner rewards over larger and later ones, is an important behavioural phenomenon for daily functioning of increasing interest within psychopathology. The neurobiological mechanisms behind DD are not well understood and the literature on structural correlates of DD shows inconsistencies. Methods: Here we leveraged a large openly available dataset (n = 1196) to investigate associations with memory performance and gray and white matter correlates of DD using linked independent component analysis. Results: Greater DD was related to smaller anterior temporal gray matter volume. Associations of DD with total cortical volume, subcortical volumes, markers of white matter microscopic organization, working memory, and episodic memory scores were not significant after controlling for education and income. Conclusion: Effects of size comparable to the one we identified would be unlikely to be replicated with sample sizes common in many previous studies in this domain, which may explain the incongruities in the literature. The paucity and small size of the effects detected in our data underscore the importance of using large samples together with methods that accommodate their statistical structure and appropriate control for confounders, as well as the need to devise paradigms with improved task parameter reliability in studies relating brain structure and cognitive abilities with DD.

The cortical alterations underpinning the acquisition of motor skills remain debated. In this longitudinal study in younger adults, we acquired performance and neuroimaging (7 T MRI) measures weekly over the course of 6 weeks to investigate neural changes associated with learning sequences of simultaneous finger presses executed with the non-­dominant hand. Both the intervention group (n = 33), which practiced the finger sequences at home, and the control group (n = 30, no home practice) showed general performance improvements, but performance improved more and became more consistent for sequences that were intensively trained by the intervention group, relative to those that were not. Brain activity for trained sequences decreased compared with untrained sequences in the bilateral parietal and premotor cortices. No training-r­elated changes in the primary sensorimotor areas were detected. The similarity of activation patterns between trained and untrained sequences decreased in secondary, but not primary, sensorimotor areas, while the similarity of the activation patterns between different trained sequences did not show reliable changes. Neither the variability of activation patterns across trials, nor the estimates of brain structure displayed practice-­related changes that reached statistical significance. Overall, the main correlate of learning configural sequences was a reduction in brain activity in secondary motor areas.

Decision-making is a cognitive process of central importance for the quality of our lives. Here, we ask whether a common factor underpins our diverse decision-making abilities. We obtained 32 decision-making measures from 830 young people and identified a common factor that we call ‘‘decision acuity,’’ which was distinct from IQ and reflected a generic decision-making ability. Decision acuity was decreased in those with aberrant thinking and low general social functioning. Crucially, decision acuity and IQ had dissociable brain signatures, in terms of their associated neural networks of resting-state functional connectivity. Decision acuity was reliably measured, and its relationship with functional connectivity was also stable when measured in the same individuals 18 months later. Thus, our behavioral and brain data identify a new cognitive construct that underpins decision-making ability across multiple domains. This construct may be important for understanding mental health, particularly regarding poor social function and aberrant thought patterns.

With increasing age, functional connectomes become dissimilar across normal individuals, reflecting heterogenous aging effects on functional connectivity (FC). We investigated the distribution of these effects across the connectome and their relationship with age-related differences in dopamine (DA) D1 receptor availability and gray matter density (GMD). With this aim, we determined aging effects on mean and interindividual variance of FC using fMRI in 30 younger and 30 older healthy subjects and mapped the contribution of each connection to the patterns of age-related similarity loss. Aging effects on mean FC accounted mainly for the dissimilarity between connectomes of younger and older adults, and were related, across brain regions, to aging effects on DA D1 receptor availability. Aging effects on the variance of FC indicated a global increase in variance with advancing age, explained connectome dissimilarity among older subjects and were related to aging effects on variance of GMD. The relationship between aging and the similarity of connectomes can thus be partly explained by age differences in DA modulation and gray matter structure.

Learning, or the process of acquiring knowledge and skill, allows humans to shape and adapt to their environments during development. Researchers have long theorized that the principal brain processes behind learning resemble a recruitment process. The brain initially explores an expanded pool of candidate neural circuits. Based on outcomes, the most promising candidate circuit is selected for refinement. Partly fuelled by new methods, the last decade of research on learning-related functional and structural changes in rodents has supported this theory, and, more recently, related evidence has started to emerge from human studies. We emphasize the need for formal theories and neurocomputational modelling of cortical plasticity to guide work on open issues, such as the link between functional and structural changes.

Probabilistic reward learning ref lects the ability to adapt choices based on probabilistic feedback. The dopaminergically innervated corticostriatal circuit in the brain plays an important role in supporting successful probabilistic reward learning. Several components of the corticostriatal circuit deteriorate with age, as it does probabilistic reward learning. We showed previously that D1 receptor availability in NAcc predicts the strength of anticipatory value signaling in vmPFC, a neural correlate of probabilistic learning that is attenuated in older participants and predicts probabilistic reward learning performance. We investigated how white matter integrity in the pathway between nucleus accumbens (NAcc) and ventromedial prefrontal cortex (vmPFC) relates to the strength of anticipatory value signaling in vmPFC in younger and older participants. We found that in a sample of 22 old and 23 young participants, fractional anisotropy in the pathway between NAcc and vmPFC predicted the strength of value signaling in vmPFC independently from D1 receptor availability in NAcc. These findings provide tentative evidence that integrity in the dopaminergic and white matter pathways of corticostriatal circuitry supports the expression of value signaling in vmPFC which supports reward learning, however, the limited sample size calls for independent replication. These and future findings could add to the improved understanding of how corticostriatal integrity contributes to reward learning ability.

The substantia nigra (SN), the subthalamic nucleus (STN), and the red nucleus (RN) are midbrain structures of ample interest in many neuroimaging studies, which may benefit from the availability of automated segmentation methods. The high iron content of these structures awards them high contrast in quantitative susceptibility mapping (QSM) images. We present a novel segmentation method that leverages the information of these images to produce automated segmentations of the SN, STN, and RN. The algorithm builds a map of spatial priors for the structures by non-linearly registering a set of manually-traced training labels to the midbrain. The priors are used to inform a Gaussian mixture model of the image intensities, with smoothness constraints imposed to ensure anatomical plausibility. The method was validated on manual segmentations from a sample of 40 healthy younger and older subjects. Average Dice scores were 0.81 (0.05) for the SN, 0.66 (0.14) for the STN and 0.88 (0.04) for the RN in the left hemisphere, and similar values were obtained for the right hemisphere. In all structures, volumes of manual and automatically obtained segmentations were significantly correlated. The algorithm showed lower accuracy on R2* and T2-weighted Fluid Attenuated Inversion Recovery (FLAIR) images, which are also sensitive to iron content. To illustrate an application of the method, we show that the automated segmentations were comparable to the manual ones regarding detection of age-related differences to putative iron content.

Alterations in iron concentration in certain deep gray matter regions are known to occur in aging and several clinical conditions. In vivo measurements of R2∗ transverse relaxation rates and quantitative susceptibility mapping (QSM) have been shown to be strongly correlated with iron concentration in tissue, but their calculation requires the acquisition of a multi-echo gradient recalled echo sequence (MGRE). In the current study, we examined the feasibility of approximating R2∗ rates using metrics derived from fMRI-EPI and T2-weighted FLAIR images, which are widely available. In a sample of 40 healthy subjects, we obtained these metrics (vEPI and vFLAIR), as well as R2∗ rates and QSM estimates, and found significant correlations between vEPI and vFLAIR and R2∗ rates in several subcortical gray matter regions known to accumulate iron, but not in a control corticospinal white matter region. These relationships were preserved after referencing vEPI and vFLAIR with respect to the values in the control region. Effect sizes (above 0.5 for some of the regions, particularly the largest ones) were calculated and put in relation to those of the correlation between QSM and R2∗ rates. We propose that the metrics described here may be applied, possibly in a retrospective fashion, to analyze datasets with available EPI or T2-weighted FLAIR scans (and lacking a MGRE sequence), to devise new hypotheses regarding links between iron concentration in brain tissue and other variables of interest.

Purpose: To use dynamic contrast-enhanced (DCE) and diffusion-weighted (DW) MRI at 3 Tesla (T) for early evaluation of treatment effects in breast cancer patients undergoing neoadjuvant chemotherapy (NAC), and assess the reliability of DW-MRI. Materials and Methods: DW- and DCE-MRI acquisitions of 15 breast cancer patients were performed before and after one cycle of NAC. MRI tumor diameter and volume, apparent diffusion coefficient (ADC) and kinetic parameters (Ktrans, ve) were derived. The reliability of ADC before NAC was assessed. Changes in MRI parameters after NAC were analyzed, and logistic regression analysis was used to find the best predictors for pathologic response. Results: The reliability for ADC values was high, with intraclass correlation coefficient of 0.84 (P ¼ 0.001). After one cycle of NAC, MRI tumor diameter (8\%, P ¼ 0.005) and tumor volume (30\%, P ¼ 0.008) was reduced for all patients, while ADC mean values increased (0.12 mm2/s, P ¼ 0.008). The best predictor for treatment response was a change in MRI tumor diameter with mean error rate of 0.167 (13\% for responders, 5\% for nonresponders, P ¼ 0.291). Conclusion: Changes in MRI derived tumor diameter and ADC after only one cycle of NAC could provide a valuable tool for early evaluation of treatment effects in breast cancer patients.

Background: A systematic comparison of magnetic resonance imaging (MRI) options for glioma diagnosis is lacking. Purpose: To investigate multiple MR-derived image features with respect to diagnostic accuracy in tumor grading and survival prediction in glioma patients. Material and Methods: T1 pre- and post-contrast, T2 and dynamic susceptibility contrast scans of 74 glioma patients with histologically confirmed grade were acquired. For each patient, a set of statistical features was obtained from the parametric maps derived from the original images, in a region-of-interest encompassing the tumor volume. A forward stepwise selection procedure was used to find the best combinations of features for grade prediction with a cross-validated logistic model and survival time prediction with a cox proportional-hazards regression. Results: Presence/absence of enhancement paired with kurtosis of the FM (first moment of the first-pass curve) was the feature combination that best predicted tumor grade (grade II vs. grade III–IV; median AUC ¼ 0.96), with the main contribution being due to the first of the features. A lower predictive value (median AUC ¼ 0.82) was obtained when grade IV tumors were excluded. Presence/absence of enhancement alone was the best predictor for survival time, and the regression was significant (P , 0.0001). Conclusion: Presence/absence of enhancement, reflecting transendothelial leakage, was the feature with highest predictive value for grade and survival time in glioma patients.