A large-scale multi-centre study characterising atrophy heterogeneity in Alzheimer’s disease

Background Previous studies reported on the existence of atrophy-based Alzheimer’s disease (AD) subtypes that associate with distinct clinical symptoms. However, the consistency of AD atrophy subtypes across approaches remains uncertain. This large-scale study aims to assess subtype concordance in individuals using two methods of data-driven subtyping. Methods We included n = 10,011 patients across the clinical spectrum from 10 AD cohorts across Europe, regional volumes using Freesurfer v7.1.1. To characterise atrophy heterogeneity in the AD continuum, we introduced a hybrid two-step approach called Snowphlake (Staging NeurOdegeneration With PHenotype informed progression timeLine of biomarKErs) to identify subtypes and sequence of atrophy-events within each subtype. We compared our results with SuStaIn (Subtype and Stage Inference) which jointly estimates both, and was trained and validated similarly. The training dataset included Aβ+ participants ( n = 1,195), and a control group of Aβ-cognitively unimpaired participants ( n = 1,692). We validated model staging within each subtype, in a held-out clinical-validation dataset ( n = 6,362) comprising patients across the clinical spectrum irrespective of Aβ biomarker status and an independent external dataset ( n = 762). Furthermore, we validated the clinical significance of the detected subtypes, in a subset of Aβ+ validation datasets with n = 1,796 in the held-out sample and n = 159 in the external dataset. Lastly, we performed concordance analysis to assess the consistency between the methods. Results In the AD dementia (AD-D) training data, Snowphlake identified four subtypes: diffuse cortical atrophy (21.1%, age 67.5 ± 9.3), parieto-temporal atrophy (19.8%, age 60.9 ± 7.9), frontal atrophy (24.8%, age 67.6 ± 8.8) and subcortical atrophy (25.1%, 68.3 ± 8.2). The subtypes assigned in Aβ+ validation datasets were associated with alterations in specific cognitive domains (Cohen’s f: [0.15 - 0.33]), while staging correlated with Mini-Mental State Examination (MMSE) scores (R: [-0.51 to - 0.28]) in the validation datasets. SuStaIn also identified four subtypes: typical (55.7%, age 66.7 ± 7.8), limbic-predominant (24.2%, age 72.2 ± 6.6), hippocampal-sparing (14.6%, age 62.8 ± 6.9), and subcortical (0.8%, age 68.2 ± 7.6). The subtypes assigned in Aβ+ validation datasets using SuStaIn were also associated with alterations in specific cognitive domains (Cohen’s f [0.17 - 0.34]), while staging correlated with MMSE scores in the validation datasets (R: [-0.54 to - 0.26]). However, we observed low concordance between Snowphlake and SuStaIn, with 39.7% of AD-D patients consistently grouped in concordant subtypes by both the methods. Conclusion In this multi-cohort study, both Snowphlake and SuStaIn identified four subtypes that were associated with different symptom profiles and atrophy-severity measures that were associated with global cognition. The low concordance between Snowphlake and SuStaIn suggests that heterogeneity may rather be a spectrum than discretised by subtypes.