A Cross-language Dementia Classifier: a Preliminary Study

Life expectancy increased globally, and the increasing prevalence of age-related issues is a major societal challenge. In particular, the World Health Organisation estimates that people suffering from dementia worldwide will grow up to 150 million by mid-century. Current pharmacologic treatments are only symptomatic, and therapies are ineffective to slow down or cure the degenerative process. An automatic and standardised tool for early screening of the signals of Alzheimer’s disease is thereby of the utmost importance to rapidly respond. Recently literature suggests that language impairment is a promising sign to reveal early signs of cognitive decline. However, most of the proposed methods require manual activities to extract language features resulting in not scalable, non-standardised and time-consuming screening tools. In this paper, we propose a cross-language Alzheimer’s disease classifier based on spontaneous speech, which is an end-to-end deep learning model able to achieve better results than manual and semi-automatic approaches. In particular, we investigated the capability of our method to achieve good classification results regardless of the language used for the training of the model. To evaluate the method, we used the English audio files of the Pitt Corpus from DementiaBank with 180 subjects: 43 healthy controls subjects and 137 Alzheimer’s disease patients, and the Italian audio files from 96 subjects: 48 healthy controls subjects and 48 Alzheimer’s disease patients. The proposed method obtained good classification results with an accuracy of 93.30% and 90.57%, respectively, for the English and Italian languages separately. Moreover, the classifier achieved good results when training and testing languages differ, obtaining an accuracy of 89.89%, using the English language for training the network and the Italian language for testing, and 89.32% vice-versa.