Brain-Age: Advancing Dementia Biomarkers in Clinical Settings

1163 

Abstract Submission 

Marcella Montagnese^1,2, Serena Verdi³, James Cole³, Timothy Rittman¹

¹Department of Clinical Neurosciences, University of Cambridge, Cambridge, United Kingdom, ²Christ's College, University of Cambridge, Cambridge, United Kingdom, ³Centre for Medical Image Computing, Department of Computer Science, University College London, London, United Kingdom

Marcella Montagnese  
Department of Clinical Neurosciences, University of Cambridge|Christ's College, University of Cambridge
Cambridge, United Kingdom|Cambridge, United Kingdom

Serena Verdi  
Centre for Medical Image Computing, Department of Computer Science, University College London
London, United Kingdom

James Cole, PhD  
Centre for Medical Image Computing, Department of Computer Science, University College London
London, United Kingdom

Timothy Rittman  
Department of Clinical Neurosciences, University of Cambridge
Cambridge, United Kingdom

The escalating incidence of dementia poses a great challenge to global healthcare, with projections indicating 135 million affected individuals by 2050 [1]. Neuroimaging has become an important tool in diagnosing dementia, with the potential to use imaging-derived metrics to revolutionise the field. An innovation in this domain is the 'brain-age' metric [2], a computational estimation of the brain's biological age, which is based on MRI scans and is indicative of brain atrophy. A brain that appears biologically 'older' than its chronological counterpart signals increased risk of age-related deterioration and death [3]. However, implementing these computational solutions in healthcare presents several barriers, including data-privacy and governance issues - limiting the application of such approaches to routinely collected clinical data. Computational tools for dementia are therefore often developed and tested on data that does not accurately represent real-life patients [4]. One important study to date has examined the prognostic value of brain-age within clinical cohorts [5]. Our work seeks to extend this exploration to encompass its diagnostic usefulness alongside routine cognitive tests in multiple diagnoses within memory clinics.

We used T1w MPRAGE scans from participants in the QMIN-MC cohort - recruited from memory clinics and community-based psychiatry-led clinics within the UK National Health Service (NHS) – encompassing a representative range of individuals. Patients were categorised into groups according to their diagnoses: Alzheimer's Disease (N=111, 74.1 ± 8.6 years); Mild Cognitive Impairment (N=53, 74.5 ± 7.7 years); Non-Alzheimer's Dementia (N=40, 76.1 ± 7.9 years), and Functional/Attentional Memory Symptoms (N=50, 59.5 ± 9.2 years). The brainageR tool (v2.1, https://github.com/james-cole/brainageR) was used to estimate brain age from raw T1w scans. For the initial pre-processing phase, SPM12 software was used for data segmentation and normalisation, followed by a predictive analysis. Brain-predicted age difference (brain-PAD), was determined by subtracting the individual's age at scanning from predicted brain age. See schematics in Figure 1. For multi-diagnosis classification, a Gradient Boosting approach was employed, using brain-PAD and routine cognitive scores. Multicollinearity was evaluated with Variance Inflation Factors and feature importance with Recursive Feature Elimination. Model robustness was ensured through k-fold cross-validation, with model efficacy quantified by weighted F1-scores to account for class imbalance. Comparative analyses were then done to evaluate the added diagnostic value of brain-PAD, including permutation tests for p-value calculations.

Comparative analyses between models (with vs without brain-PAD) revealed classification improvements across all diagnoses, alongside shifts in feature importance. Mean F1-score improved from 0.51 to 0.56 (p=.22). For Non-AD Dementia, precision improved from 0.23 to 0.41 (p<0.001), and F1-scores from 0.11 to 0.25. AD precision went from 0.61 to 0.63 (p<0.05), and F1-score to 0.71. The model also showed gains for Functional/Attentional Memory Symptoms (F1-score 0.71; precision from 0.62 to 0.74, p<0.001), and MCI (F-score 0.39, precision from 0.39 to 0.41, p<0.05). The brain-PAD model was more parsimonious (3 vs 6 features), suggesting a simplification in the cognitive subdomains, with more reliance on brain-PAD. Shapley Additive exPlanations (SHAP)[6] values echoed the feature importance findings, with brain-PAD showing higher impact on model output, particularly for the correct classification of AD and Non-AD Dementia. See Figure 2.

The integration of brain-PAD showed a relatively small but consistent boost in predictive performance across diagnoses. Looking at features importance showed that brain-PAD has the potential to be a crucial biomarker in the heterogeneous and multiclass diagnostic landscape of dementia even in smaller samples.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Aging ¹

Classification and Predictive Modeling

Other Methods

Anatomical MRI

Aging

Computational Neuroscience

Degenerative Disease

Machine Learning

Neurological