Exploring the Diagnosis of Frontotemporal Dementia by Analyzing Neuropsychological Data With K-Means Clustering

Background: Differential diagnosis of dementia syndromes is difficult. Treatments are available for Alzheimer’s dementia (AD), but effects do not translate to other dementia syndromes. Therefore, early differential diagnosis is necessary to administer appropriate interventions and to boost drug development and research of therapeutic strategies that may lower patient and care-giver burden. The focus of the current study is to disentangle the clinical picture of frontotemporal dementia (FTD) from AD spectrum disorders. Findings could support an early, cheap, and more accurate diagnosis. Methods: K-means clustering, an unsupervised machine learning algorithm, was used on neuropsychological data from neurologic patients of the FTLD consortium databank. The analysis was performed twice, once including only neuropsychological test scores and a second time combining the neuropsychological variables with questionnaire scores assessing behavioral changes. In total n = 484 and n = 469 participants were included in the analysis with and without questionnaires, respectively. Participants included were either healthy controls with no family relation to patients in the dataset, or patients diagnosed with one of the following dementia disorders: AD, a behavioral variant of FTD (bvFTD), or one of three possible primary progressive aphasia (PPA) syndromes - a semantic variant (svPPA), a non-fluent variant (nfvPPA) or a logopenic variant (lvPPA). Results: Agreement of results from the various analyses performed was relatively high. Homogeneous clusters of diagnostic groups emerged. Homogeneity seemed higher for bvFTD and svPPA than for the other patient groups. NfvPPA and lvPPA patients were particularly likely to cluster together. Exploring neuropsychological patterns of cluster results demonstrated high variability between patients of the same diagnostic groups which could partly be explained by differences in disease severity. Tests that might prove particularly relevant to distinguish diagnostic subgroups are the FTLD-CDR sub-scores, the repeat and point task as well as questionnaires assessing apathy. Conclusion: K-means clustering proved to be a useful technique to explore various diagnostic syndromes that show overlapping clinical pictures. This study helped to formulate specific hypotheses based on the observation of patterns in multidimensional data. Disease severity showed to impact k-means clustering results considerably and should therefore be accounted for in future studies. Future studies will need to test the formulated hypotheses and inspect the meaning of impure clusters. Keywords: k-means clustering, frontotemporal dementia, primary progressive aphasia, Alzheimer’s dementia, differential diagnosis, neuropsychological tests