Mapping individualised patterns of atrophy in Alzheimer’s disease using neuroanatomical normative models

Abstract Background Alzheimer’s Disease (AD) is highly heterogeneous, with marked individual differences in clinical presentation and aspects of its neurobiology. Neuroanatomical normative modelling is an emerging technique that reveals individual patterns of neuroanatomy by quantifying deviations from normative ranges (Verdi et al., 2021). Using Alzheimer’s Disease Neuroimaging Initiative (ADNI) data, we applied a hierarchical Bayesian regression (HBR) normative model to account for site effects (Kia et al., 2021). We assessed individual patterns of cortical thickness (CT) heterogeneity in AD patients, people with mild cognitive impairment (MCI), and healthy controls. Method CT across 148 regions of interest (ROIs) was generated using FreeSurfer from 1492 T1‐weighted MRI scans across 63 sites. A reference HBR normative model was trained on a separate healthy dataset of 34,490 to index population variability, predicting CT at each ROI using age and sex. This generated CT z‐scores for each ROI, per participant. Here z‐scores < ‐1.96 were defined as outliers. Outliers were also summed across 148 ROIs to give a total outlier score per participant. The influence of the total outlier score on converting from MCI to AD within 3 years was tested using Cox proportional hazards regression. Result AD subjects had higher total outlier scores than MCI and control participants (β=10.99,95%CI=[5.55,16.42], p<0.001), with marked differences in temporal regions. Total outlier score predicted memory performance (R 2 =0.39,p<0.001) and executive function (R 2 =0.35,p<0.001). Total outlier score was predicted by amyloid‐beta (R 2 =0.06,p<0.001) and phospho‐tau (R 2 =0.16,p<0.001). The patterns of neuroanatomical outliers varied markedly between AD patients (Fig.1) ; the parahippocampal gyrus had the highest rate of AD patients with outliers (47%). Outlier maps of AD patients matched age, sex, cognition and APOE4 genotype, illustrate spatially heterogeneous patterns of outliers (Fig.2) . Greater within‐group dissimilarity was seen in AD patients, relative to MCI or control participants (Fig.3) . Every 10 points of total outlier score contributed to an 31.4% risk of converting from MCI to AD within 3years (HR=1.028,95%CI=[1.016,1.039],p<0.001) (Fig.4) . Conclusion Our novel quantitative estimate of the neuroanatomical heterogeneity of CT suggests the impact of AD on the brain is not consistent between patients. Furthermore, quantify outliers in CT may help predict individuals with MCI at risk of converting to AD.