Data-driven neuroanatomical subtypes of primary progressive aphasia

Abstract

The primary progressive aphasias are rare, language-led dementias, with three main variants: semantic, non-fluent/agrammatic and logopenic. Although the semantic variant has a clear neuroanatomical profile, the non-fluent/agrammatic and logopenic variants are difficult to discriminate from neuroimaging. Previous phenotype-driven studies have characterized neuroanatomical profiles of each variant on MRI. In this work, we used a machine learning algorithm known as SuStaIn to discover data-driven neuroanatomical 'subtype' progression profiles and performed an in-depth subtype-phenotype analysis to characterize the heterogeneity of primary progressive aphasia. Our study included 270 participants with primary progressive aphasia seen for research in the UCL Queen Square Institute of Neurology Dementia Research Centre, with follow-up scans available for 137 participants. This dataset included individuals diagnosed with all three main variants (semantic, n = 94; non-fluent/agrammatic, n = 109; logopenic, n = 51) and individuals with unspecified primary progressive aphasia (n = 16). A dataset of 66 patients (semantic, n = 37; non-fluent/agrammatic, n = 29) from the ARTFL LEFFTDS Longitudinal Frontotemporal Lobar Degeneration (ALLFTD) Research Study was used to validate our results. MRI scans were segmented, and SuStaIn was used on 19 regions of interest to identify neuroanatomical profiles independent of the diagnosis. We assessed the assignment of subtypes and stages, in addition to their longitudinal consistency. We discovered four neuroanatomical subtypes of primary progressive aphasia, labelled S1 (left temporal), S2 (insula), S3 (temporoparietal) and S4 (frontoparietal), exhibiting robustness to statistical scrutiny. S1 was correlated strongly with the semantic variant, whereas S2, S3 and S4 showed mixed associations with the logopenic and non-fluent/agrammatic variants. Notably, S3 displayed a neuroanatomical signature akin to a logopenic-only signature, yet a significant proportion of logopenic cases were allocated to S2. The non-fluent/agrammatic variant demonstrated diverse associations with S2, S3 and S4. No clear relationship emerged between any of the neuroanatomical subtypes and the unspecified cases. At first follow-up, subtype assignment was stable for 84% of patients, and stage assignment was stable for 91.9% of patients. We partially validated our findings in the ALLFTD dataset, finding comparable qualitative patterns. Our study, leveraging machine learning on a large primary progressive aphasia dataset, delineated four distinct neuroanatomical patterns. Our findings suggest that separable spatiotemporal neuroanatomical phenotypes do exist within the primary progressive aphasia spectrum, but that these are noisy, particularly for the non-fluent/agrammatic non-fluent/agrammatic and logopenic variants. Furthermore, these phenotypes do not always conform to standard formulations of clinico-anatomical correlation. Understanding the multifaceted profiles of the disease, encompassing neuroanatomical, molecular, clinical and cognitive dimensions, has potential implications for clinical decision support.

Keywords: atypical dementia; longitudinal; machine learning; phenotype; progression modelling; subtype and stage inference.

Figure 1

Positional variance diagrams for the data-driven subtypes. Along the y-axis are the regions of interest used in the model, grouped by location in the brain. The data-driven stages correspond to the sequence in which brain regions become abnormal, with colour representing the degree of abnormality (w-score 1, red; w-score 2, pink; w-score 3, blue) and colour density representing model certainty.

Figure 2

Stages of regional atrophy in the outer cortical left hemisphere for the data-driven subtypes. Colour represents the degree of abnormality (w-score 1, red; w-score 2, pink; w-score 3, blue). Brain regions that are dark grey were not included in the model.

Figure 3

Comparison between data-driven subtype assignment and clinical diagnosis. (A) The stacked bar chart shows the number of patients with each clinical diagnosis by data-driven subtype assignment. (B) The stacked bar chart shows number of patients who were assigned to each data-driven subtype by clinical diagnosis. lvPPA = logopenic variant PPA; nfvPPA = non-fluent/agrammatic variant PPA; PPA-nos = PPA not otherwise specified; svPPA = semantic variant PPA.

Figure 4

Boxplots comparing MMSE and neuropsychological scores between subtypes at baseline in the Queen Square discovery dataset. The full set of neuropsychological scores per subtype is available in Supplementary Table 5. MMSE = Mini-Mental State Examination; PALPA-55 = Psycholinguistic Assessments of Language Processing in Aphasia, subtest 55.

Figure 5

Mini-Mental State Examination (MMSE) versus stage at baseline in the Queen Square discovery dataset.

Figure 6

Longitudinal consistency of subtype and stage assignment. (A) Sankey diagram of subtype assignment between baseline and first follow-up visits. (B) Data-driven stage assignment at baseline and first follow-up by data-driven subtype. (C) Data-driven stage assignment at baseline and first follow-up by clinical subtype. Those above the diagonal progressed at follow-up, whereas those below regressed. lvPPA = logopenic variant PPA; nfvPPA = non-fluent/agrammatic variant PPA; PPA-nos = PPA not otherwise specified; svPPA = semantic variant PPA.