Towards Domain-Agnostic and Domain-Adaptive Dementia Detection from Spoken Language

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11965–11978 July 9-14, 2023 ©2023 Association for Computational Linguistics Towards Domain-Agnostic and Domain-Adaptive Dementia Detection from Spoken Language Shahla Farzana and Natalie Parde Natural Language Processing Laboratory Department of Computer Science University of Illinois Chicago {sfarza3, parde}@uic.edu Abstract Health-related speech datasets are often small and varied in focus. This makes it difficult to leverage them to effectively support healthcare goals. Robust transfer of linguistic features across different datasets orbiting the same goal carries potential to address this concern. To test this hypothesis, we experiment with do- main adaptation (DA) techniques on heteroge- neous spoken language data to evaluate general- izability across diverse datasets for a common task: dementia detection. We find that adapted models exhibit better performance across con- versational and task-oriented datasets. The feature-augmented DA method achieves a 22% increase in accuracy adapting from a conver- sational to task-specific dataset compared to a jointly trained baseline. This suggests promis- ing capacity of these techniques to allow for productive use of disparate data for a complex spoken language healthcare task. 1 Introduction Data-driven models of diverse conditions affect- ing spoken language abilities offer promising real- world benefits (Amini et al., 2022; Girard et al., 2022). However, the datasets available for devel- oping these models are often small and disparate, spanning varied diagnostic and non-diagnostic tasks mapped to different taxonomies at conflicting granularities (Graham et al., 2020). This has largely constrained progress to models excelling in special- ized settings (e.g., individuals with homogeneous language background describing a standardized im- age (Luz et al., 2020)). At the same time, it has created challenges in building more generalizable knowledge about language patterns associated with the condition of interest (Guo et al., 2021). Outside healthcare applications, domain adap- tation (DA) has long been applied to increase the capacity of NLP systems to leverage meaningful information from diverse data (Kouw and Loog, 2018). These techniques generally seek to har- ness data from one domain (the source) to improve performance in another (the target). Usually the target domain has little or no labeled data, while the source has a relatively large amount of labeled data. Despite the advantages offered by DA for many NLP problems, it has remained under-studied for healthcare tasks due to numerous complexities of healthcare data (Laparra et al., 2020). Nonetheless, most healthcare problems offer the ideal learning settings in which DA is designed to thrive. We present a systematic analysis of the use of DA for a low-resource healthcare problem that has recently been popular in the NLP community: de- mentia. We adopt a wide definition of dementia in our work, encompassing datasets pertaining to Alzheimer’s disease or related dementia (ADRD) and age-related mild cognitive impairment (MCI), in line with current NLP community norms (Amini et al., 2022). Our research questions include: Q1. Can DA be used to exploit spoken language data pertaining to dementia from one domain, to improve its detection in other domains? …

Content truncated to 500 words