COGNITIVE SUBTYPES IDENTIFIED USING NONNEGATIVE MATRIX FACTORISATION IN FOUR LARGE ALZHEIMER'S DISEASE DEMENTIA COHORTS

Alzheimer’s disease (AD) dementia is a heterogeneous disorder with variability in cognitive presentation. We aimed to identify cognitive AD subtypes using a data-driven dual clustering approach. We repeated analyses in four large independent AD cohorts to assess robustness of the identified cognitive subtypes. We included probable AD dementia patients with Mini-Mental State Examination score (MMSE) >16/30 from the Amsterdam Dementia Cohort (ADC, n=496), the Alzheimer’s Disease Neuroimaging Initiative (ADNI, n=376), the German Dementia Competence Network (DCN, n=521), and the University of California San Francisco Memory Clinic (UCSF, n=589). The composition of neuropsychological test batteries differed between cohorts, but all covered the different cognitive domains. After imputation, normalisation, and scaling to include only positive values (0-1), we clustered neuropsychological test results for each cohort using the R package nonnegative matrix factorisation (NMF). This resulted in a clustering of neuropsychological tests (dimensionality reduction) and subjects (cognitive subtypes). Number of clusters was determined based on the highest cophenetic correlation. We compared identified cognitive clusters in terms of demographics, Apolipoprotein E [APOE] genotype, cerebrospinal fluid AD biomarkers, and cortical atrophy ratings. In all four cohorts, neuropsychological tests were clustered into two components (all cophenetic correlations >.8): One including mostly memory tests, and one largely including non-memory tests. Two clusters of patients were identified: One associated with low scores on the memory component (memory-impaired clusters, including 52-71% of patients) and one associated with low scores on the non-memory component (memory-spared clusters, including 29-48% of patients). Memory-spared clusters were younger, more often APOE e4 negative, had more severe parietal and frontal atrophy, and had lower MMSE scores than the memory-impaired clusters. Using dual-clustering approach NMF we robustly identified two cognitive AD subtypes in four independent large cohorts; one memory-impaired and one relatively memory-spared. The memory-spared subtypes included about one third of patients, and were characterized by younger age, APOE e4 negative genotype, more severe parietal or frontal atrophy, and lower MMSE scores compared to the memory-impaired clusters.