Multilingual prediction of Alzheimer’s disease through domain adaptation and concept-based language modelling

Proceedings of NAACL-HLT 2019, pages 3659–3670 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3659 Multilingual Prediction of Alzheimer’s Disease Through Domain Adaptation and Concept-Based Language Modelling Kathleen C. Fraser1, Nicklas Linz2, Bai Li3, Kristina Lundholm Fors4, Frank Rudzicz3,5, Alexandra K¨onig6, Jan Alexandersson2, Philippe Robert6, Dimitrios Kokkinakis4 1National Research Council Canada, Ottawa, Canada 2German Research Center for Artificial Intelligence (DFKI), Saarbr¨ucken, Germany 3University of Toronto and Vector Institute, Toronto, Canada 4University of Gothenburg, Gothenburg, Sweden 5St. Michael’s Hospital, Toronto, Canada 6University of Cˆote d’Azur, Nice University Hospital, and INRIA, Nice, France kathleen.fraser@nrc-cnrc.gc.ca, nicklas.linz@dfki.de, bai@cs.toronto.edu, kristina.lundholmfors@gu.se, frank@cs.toronto.edu, alexandra.konig@inria.fr, jan.alexandersson@dfki.de, probert@unice.fr, dimitrios.kokkinakis@gu.se Abstract There is growing evidence that changes in speech and language may be early markers of dementia, but much of the previous NLP work in this area has been limited by the size of the available datasets. Here, we compare sev- eral methods of domain adaptation to augment a small French dataset of picture descriptions (n = 57) with a much larger English dataset (n = 550), for the task of automatically distin- guishing participants with dementia from con- trols. The first challenge is to identify a set of features that transfer across languages; in addition to previously used features based on information units, we introduce a new set of features to model the order in which informa- tion units are produced by dementia patients and controls. These concept-based language model features improve classification perfor- mance in both English and French separately, and the best result (AUC = 0.89) is achieved using the multilingual training set with a com- bination of information and language model features. 1 Introduction According to the World Health Organisation, the largest global challenge facing the world today is the rapid increase of the population aged over 65 years. It is projected to increase from 524 million in 2010 to 1.5 billion in 2050, with the largest increase in the developing world (Suzman and Beard, 2011). This demographic trend has profound societal implications; for example, the number of persons affected by dementia will in- crease worldwide from 46 million in 2015 to 131.5 million in 2050 (Prince et al., 2015). The most common underlying condition causing dementia is Alzheimer’s disease (AD). Although no cure to this neurodegenerative disease has been found, ex- perts agree that intervention in early stages is cru- cial to delay onset (Dubois et al., 2016). AD is characterised by a global impairment of cognitive functioning, with specific deficits in episodic memory, executive functioning, percep- tual speed and language (B¨ackman et al., 2005; Weiner et al., 2008). Machine learning experiments using speech and language for the detection of dementia or re- lated disorders have been conducted in many lan- guages, including English (Roark et al., 2011; Mirheidari et al., 2016; Fraser et al., 2016; As- gari et al., 2017), French (Tr¨oger et al., 2017; K¨onig et al., 2018), German (Weiner et al., 2016), Hungarian (Szatloczki et al., 2015; Vincze et al., 2016), Spanish (Meil´an et al., 2014), Greek …

Content truncated to 500 words