Data‐driven neuroanatomical subtype trajectories of primary progressive aphasia.

Background: Primary progressive aphasia (PPA) is an atypical neurodegenerative dementia with three main clinical variants: semantic (svPPA), non‐fluent/agrammatic (nfvPPA) and logopenic (lvPPA). While svPPA is typically associated with left anterior temporal lobe atrophy, neuroimaging findings in nfvPPA, lvPPA and PPA not otherwise specified (PPA‐nos) are more variable. We therefore applied unsupervised machine learning to one of the largest databases of PPA magnetic resonance imaging (MRI) scans ever acquired. Method: Baseline MRI scans were collected from 273 people with PPA participating in the longitudinal research programme in the Queen Square Dementia Research Centre (svPPA n = 97; nfvPPA n = 109; lvPPA n = 51; PPA‐nos n = 16). Longitudinal scans were collected from 138 participants. We used Subtype and Stage Inference (SuStaIn), a model that combines disease progression modelling with unsupervised machine learning to identify subgroups of patients following a similar trajectory. We standardised the data with respect to cognitively normal controls (n = 360). Based on clinicians' input we chose 19 brain regions of interest to model from those generated by GIF (Cardoso et al. 2012). The model was fit using baseline data, selecting the appropriate number of subtypes using cross‐validation (Young et al. 2018). Result: Cross‐validation supported four subtypes of spatiotemporal atrophy (Figure 1). Clinical variant representation per subtype assignment (1/2/3/4) was: svPPA (n = 74/15/4/4), nfvPPA (n = 3/35/21/44), lvPPA (n = 1/18/30/2) and PPA‐nos (n = 7/4/3/2) (Figure 2). Subtype 1 showed early decline in the left anterior temporal lobe and sparing of the left frontal lobe. Subtype 2 showed early decline of the left insula and sparing of the temporal pole. Subtype 3 showed initial atrophy of the left temporoparietal junction; Subtype 4 showed early atrophy in the left precentral gyrus. Statistical analysis of posterior distributions found low similarity between subtypes (Hellinger distance > 0.66) (Figure 3), validating the presence of four distinct subtypes. Conclusion: While the data‐driven atrophy subtypes did not align perfectly with clinical variants, we found a correspondence between Subtype 1 and svPPA, the most coherent PPA variant. Subtypes 2,3 and 4 were associated with nfvPPA and lvPPA, corroborating work suggesting these diagnostic terms encompass multiple clinical phenotypes. Future work will investigate the association between clinical diagnosis, longitudinal subtype, and clinical measures. [ABSTRACT FROM AUTHOR]