Classification of primary progressive aphasia: Do unsupervised data mining methods support a logopenic variant?

Abstract

Our objective was to test whether data mining techniques, through an unsupervised learning approach, support the three-group diagnostic model of primary progressive aphasia (PPA) versus the existence of two main/classic groups. A series of 155 PPA patients observed in a clinical setting and subjected to at least one neuropsychological/language assessment was studied. Several demographic, clinical and neuropsychological attributes, grouped in distinct sets, were introduced in unsupervised learning methods (Expectation Maximization, K-Means, X-Means, Hierarchical Clustering and Consensus Clustering). Results demonstrated that unsupervised learning methods revealed two main groups consistently obtained throughout all the analyses (with different algorithms and different set of attributes). One group included most of the agrammatic/non-fluent and some logopenic cases while the other was mainly composed of semantic and logopenic cases. Clustering the patients in a larger number of groups (k > 2) revealed some clusters composed mostly of non-fluent or of semantic cases. However, we could not evidence any group chiefly composed of logopenic cases. In conclusion, unsupervised data mining approaches do not support a clear distinction of logopenic PPA as a separate variant.

Key words::

• Primary progressive aphasia
• logopenic variant (lvPPA)
• non-fluent variant (nfvPPA)
• semantic variant (svPPA)
• data mining

Acknowledgements

The authors thank the patients for their participation and also acknowledge the facilities provided by Memoclínica. They also thank Isabel Pavão Martins, head of the Laboratory of Language Research, as well as speech therapists of this department involved in the language assessments of some patients (Luísa Farrajota, Gabriela Leal and José Fonseca), and Jason Warren (Dementia Research Centre, UCL Institute of Neurology, University College of London, UK) for the thorough revision of the manuscript. CM and TP are supported by Fundação para a Ciência e Tecnologia (FCT) PhD Fellowships (SFRH/BD/75710/2011 and SFRH/BD/95846/2013, respectively). AdM and MG also receive funding from FCT. TP and SM receive funding from NEURO- CLINOMICS (PTDC/EIA/111239/2009) and UID/CEC/50021/2013, also funded by FCT.

Declaration of interest: The authors declare no conflicts of interest.