#21
James Cole 

The brain’s biological age and neurodegeneration: Towards real-world and representative predictors of disease risk and health outcomes 

Ageing is a biological process that affects most aspect of human biology and results in a higher risk of morbidity and mortality. The brain is no exception, and many brain diseases have greater prevalence or poorer prognosis with increased age. However, the impact of brain ageing can vary greatly between people, as seen by the broad window of onset ages for cognitive decline or dementia. Here, I will recap efforts to measure brain ageing, focusing on machine-learning models of the brain’s “biological age”, so-called “brain age”. I will then give an overview of applications of brain-age to neurodegenerative diseases, highlighting recent efforts to move beyond selective and curated research datasets, towards more diverse datasets from hospital datasets and from underrepresented groups and present early data on brain age from ultra-low-field MRI.

#22
Somayeh Maleki Balajoo 

Disentangling heterogeneity in grey matter alterations in multiple sclerosis using generative normative modelling 

Background: Relapsing–Remitting Multiple Sclerosis (RRMS) involves widespread neuroinflammatory and neurodegenerative processes, with early thalamic and grey matter atrophy. However, substantial interindividual variability limits the sensitivi ty of group-level analyses to detect subject-specific alterations, and quantitative approaches for capturing such heterogeneity remain limited. Here, we address this gap by applying individual deviation analysis to characterize neuroanatomical variability and identify distinct brain structural subtypes in RRMS.

Methods: We developed a deep generative normative model using whole-brain parcel-wise grey matter volume (GMV) data, extending region-wise regression by capturing the full multivariate structure and nonlinear, covariate-dependent effects. The model was trained on 8,258 healthy adults (18–85 years) from UK Biobank, HCP, OASIS, and ADNI. Individual deviations from normative trajectories were assessed in 200 RRMS patients, with Z-scores 1.96 indicating extreme negative and positive deviations. Distinct RRMS neurotypes were identified by clustering GMV deviation patterns.

Results: We observed marked heterogeneity in both the spatial distribution and severity of structural abnormalities across patients with RRMS. The thalamus exhibited the highest prevalence of significant negative deviations (27% of patients). Cross-sectional analyses showed that a higher abnormality burden—defined as the number of regions with significant negative deviations per patient—was associated with greater baseline disability and longer disease duration. Two robust brain structural alterations subtypes were identified. Subtype 1 was characterized by more extensive and pronounced negative deviations (reflecting atrophy), particularly in the thalamus and widespread cortical regions. In contrast, Subtype 2 showed negative deviations mostly focused on the thalamus and parahippocampal regions, while other cortical and subcortical regions show positive deviance.

Conclusions: These results highlight substantial heterogeneity in brain structural alterations in RRMS and demonstrate the value of deep normative modelling for capturing individual-level spatial patterns. The identification of two neurotypes—both with typical thalamic atrophy but one with extensive cortical atrophy and the other with important positive deviations —suggests latent neuroanatomical variability beyond traditional classifications. While positive deviations may reflect transient inflammatory processes such as oedema, this framework provides a data-driven approach to characterizing disease heterogeneity and may inform future studies of RRMS progression and stratification.

Abstract: Alzheimer’s Disease causes pathology to spread along neural circuits that serve specific cognitive processes, such as pattern separation and pattern completion and familiarity recognition. On the basis of data from the DZNE DELCODE cohort, I will discuss how amyloid and tau pathology impact on synaptic function in these circuits and to what extent dysfunction can be observed in the absence of MRI visible neurodegeneration. These data have implications on the question whether synaptic function can partially recover after amyloid removal with disease modifying anti-amyloid treatments. Building on these observations, I will discuss how non-pharmacological interventions can be used to unlock reserve mechanisms in conjunction with amyloid-removal. Our data indicate that cognitive reserve can be associated with the ability to activate upstream visual areas and midline cortical areas of the episodic memory circuitry. Individuals who maintain high levels of encoding related activation in these areas can cope with hippocampal atrophy, amyloid and tau-pathology, such that they show slower cognitive decline in memory over several years of follow-up. I will discuss approaches for brain stimulation of these regions as well as cognitive training interventions targeting these regions to potentially improve cognitive reserve mechanisms and help individuals to better maintain cognitive performance.

Keywords: Amyloid pathology, neurodegeneration, episodic memory, synaptic dysfunction, anti-amyloid treatments, cognitive reserve, brain reserve, non-pharmacological interventions

#23
Sofia Gassó

Neuroimaging signatures across APOE genotype in Alzheimer’s disease revealed by explainable machine learning

Background: Apolipoprotein E (APOE), particularly ε4 allele (APOE-ε4), is a well-recognized genetic risk factor for Alzheimer’s disease (AD). Different genotypes confer varying levels of risk. This study aims to use explainable machine learning (XAI) techniques to identify neurodegenerative atrophy patterns associated with these genetic profiles, focusing on 3/ε3 individuals and APOE-ε4 heterozygotes and homozygotes.

Methods: Magnetic resonance imaging (MRI) data were collected from two independent cohorts: the Hospital Clínic de Barcelona (HCB) (N=339) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (N=438). Machine learning (ML) algorithms were performed using Random Forest and XGBoost, combined with Shapley Additive exPlanations (SHAP) for interpretability. We first classified healthy controls (CTR) vs AD. Then, analyses were restricted to AD individuals to classify: APOE-ε4 carriers vs non-carriers, as well as to compare specific genotypes (ε3/ε3 vs ε3/ε4, ε3/ε3 vs ε4/ε4, and ε3/ε4 vs ε4/ε4).

Results: CTR vs AD classification achieved excellent performance across both cohorts and algorithms, with accuracies around 90% and AUC around 0.95. The hippocampus, middle temporal and the amygdala emerged as the most important features. Table 1 shows all the APOE comparison results. The ML algorithm differentiates APOE-ε4 homozygotes vs ε3/ε3 individuals with high accuracy 85% (AUC 0.94) in HCB and 83% (AUC 0.89) in ADNI. SHAP analyses indicated that hippocampal and temporal regions (particularly amygdala), as the most relevant contributors for the discrimination. In contrast, the classification of ε3/ε3 versus APOE-ε4 heterozygotes yielded low accuracies (around or below 60%), indicating performance near chance. Conversely, APOE-ε4 homozygote vs ε3/ε4 heterozygote were classified with good accuracy, achieving 85% (AUC 0.95) in HCB and 80% (AUC 0.90) in ADNI. The amygdala and the insula emerged in this comparison.

Conclusions: These findings demonstrate that APOE-ε4 homozygote and heterozygote differentially modulate the neurodegenerative atrophy patterns observed in AD. Based on the algorithm classifications, APOE-ε4 heterozygotes are more similar to ε3/ε3 individuals than to APOE-ε4 homozygotes. Overall, the results emphasize the importance of considering genetic factors (APOE-ε4) in understanding and characterizing atrophy patterns in AD.

#24
Vera Komeyer

Age-sensitive structural hotspots as sparse, cross-cohort features for Alzheimer’s Disease prediction

Background: The human brain undergoes systematic structural changes across the lifespan measurable with MRI-derived gray matter volume (GMV). Specific patterns of age-related variation may reveal aging “hotspots” of increased vulnerability to neurodegeneration, such as Alzheimer’s disease (AD). Here, we investigate shared neurobiological alterations between aging and neurodegeneration through a multivariate modeling framework. Voxel-level structural aging hotspots are evaluated on their cross-cohort utility for AD classification beyond conventional GMV features.

Methods: We trained 371,139 voxel-wise one-class support vector machines (OCSVM) to learn young (≤ 50 years) with effect-size calculation and spatial separation(≧14mm) defined 264 age-outlier voxels. 264 Power nodes and effect-size-permuted nulls served as control voxelsets. For each, we defined mean GMV (r=5mm) and OCSVM distances – reflecting cross-cohort deviation from young normative structure – as features on an ADNI cohort. Nested cross-validation and out-of-sample prediction in four supervised classification analyses (voxelsets x features) aimed to discriminate cognitively normal and AD individuals.

Results: Age-outlier voxels (SALD) were predominantly located in frontal and temporal cortices (Fig. 2a). Cross-cohort OCSVM distances (ADNI) showed progressive deviation from young reference patterns consistent with disease burden across diagnostic groups (Fig. 2b) and exhibited stronger negative associations with chronological age (age-outlier voxels) than average GMV (both voxelsets) (Fig. 2c). Age-outlier voxels performed statistically comparable to Power nodes across cross-validation metrics and showed largely numerical advantages with both featuresets (Fig. 3a). In the out-of-sample evaluation, distance features slightly outperformed average GMV, with marginal advantage of age-outlier voxels (Fig. 3b). Power (ROC-AUC) and age-outlier (BA) voxelsets significantly exceeded permutation baselines but only with OCSVM distance features (Fig. 3c).

Conclusions: Age-sensitive structural deviations can be distilled into a sparse set of 264 voxels whose age-associative decision-distance representation generalizes across cohorts and achieves AD discrimination comparable to conventional voxel selections. Modeling divergence from younger-adult structural patterns provided advantages over average GMV while reducing model complexity, potentially enhancing clinical feasibility. This demonstrates that biologically informative representations can be derived from a single healthy reference group without requiring disease data, supporting scalable disease-agnostic modeling strategies