Computational psycholinguistic analysis of speech samples in PPA and AD and FTD

点击翻译按钮获取中文摘要

Abstract Primary Progressive Aphasia (PPA) is a clinical neurodegenerative syndrome characterized by abnormalities in language with initial relative sparing of other cognitive processes. The syndrome may result from several kinds of neuropathology, including Alzheimer's disease (AD) or Frontotemporal Lobar Degeneration (FTLD). The different neuropathological causes are associated with specific variants of the disease. Individuals with the non-fluent variant of PPA (nfvPPA) tend to show effortful speech and agrammatism, in some cases with motor speech dysfunction. Impairments in sentence repetition and lexical retrieval are exhibited by those with the logopenic variant of PPA (lvPPA). Difficulties in object naming and word comprehension are experienced in individuals with the semantic variant of PPA (svPPA). While widely used, the current system of classification is challenged by the occurrence of individuals with overlapping profiles of linguistic behavior and by an inconsistent alignment of linguistic profiles and patterns of cortical atrophy. In addition, some of these same linguistic or anatomic abnormalities can be seen in patients with non-PPA clinical phenotypes of AD or FTLD. That is, these PPA subtypes may represent one way of classifying a multidimensional spectrum of cognitive- behavioral anatomic abnormalities arising from a set of neurodegenerative pathologies; we need new ways of quantifying these abnormalities, and we need to consider alternative classification schemes. Here we introduce a new approach to accomplishing both of these possibilities. Recent developments in Natural Language Processing (NLP) and Machine Learning (ML) have now made possible the automated discovery and classification of linguistic features. Once established, these feature sets can be connected to distributions of cortical atrophy, thus enabling links between specific linguistic behavioral abnormalities and underlying neural networks. This approach to the analysis of PPA subtypes, and their contextualization with other clinical types of AD and FTLD, can be achieved through a sufficiently large number of language samples collected in ways that highlight both the production and comprehension aspects of the language system. In addition, such analyses require the use of the latest generation of artificial intelligence models, called transformer-networks. The result will be a new understanding of the PPA syndrome and the language network that it affects. In Aim 1, we will investigate the performance of an unsupervised artificial intelligence model for measuring and classifying language abnormalities in PPA patients. In Aim 2, we will investigate the how these models can be used to measure and classify language abnormalities in AD and FTD patients. In Aim 3, we will evaluate the reliability of these automated measures of language abnormalities in PPA, AD, and FTD. Through a finer-grained analysis of language in people with PPA and other forms of AD or FTLD, it should be possible to develop better understanding of the overlapping and dissociable features of these dementias, possibly leading to improved diagnostic classification and better prognostication.