Detecting cognitive impairments by agreeing on interpretations of linguistic features

Proceedings of NAACL-HLT 2019, pages 1431–1441 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1431 Detecting Cognitive Impairments by Agreeing on Interpretations of Linguistic Features Zining Zhu1,2, Jekaterina Novikova1, Frank Rudzicz3,5,2,4,1 1 Winterlight Labs 2 University of Toronto 3 Li Ka Shing Knowledge Institute, St Michael’s Hospital, Toronto ON, Canada 4 Vector Institute for Artificial Intelligence 5 Surgical Safety Technologies {zining, jekaterina}@winterlightlabs.com, frank@cs.toronto.edu Abstract Linguistic features have shown promising ap- plications for detecting various cognitive im- pairments. To improve detection accuracies, increasing the amount of data or the num- ber of linguistic features have been two ap- plicable approaches. However, acquiring addi- tional clinical data can be expensive, and hand- crafting features is burdensome. In this paper, we take a third approach, proposing Consensus Networks (CNs), a framework to classify after reaching agreements between modalities. We divide linguistic features into non-overlapping subsets according to their modalities, and let neural networks learn low-dimensional repre- sentations that agree with each other. These representations are passed into a classifier net- work. All neural networks are optimized itera- tively. In this paper, we also present two methods that improve the performance of CNs. We then present ablation studies to illustrate the effec- tiveness of modality division. To understand further what happens in CNs, we visualize the representations during training. Overall, using all of the 413 linguistic features, our models significantly outperform traditional classifiers, which are used by the state-of-the-art papers. 1 Introduction Alzheimer’s disease (AD) and its usual precur- sor, mild cognitive impairment (MCI), are preva- lent neurodegerative conditions that inhibit cog- nitive abilities. Cognitive impairments are tradi- tionally diagnosed only with standard clinical tests like MoCA (Nasreddine et al., 2005) and the Rey- Auditory Verbal learning Test (Rey, 1941), but hiring clinicians to administer these tests and an- alyze their results is costly. Fortunately, many cognitive impairments can be observable in daily life, because they impact one’s language abilities. For example, cognitively impaired people tend to use more pronouns instead of nouns, and pause more often between sentences in narrative speech (Roark et al., 2011). This insight makes automatic detection possi- ble. Machine learning classifiers can detect cogni- tive impairments given descriptive linguistic fea- tures. In recent work, linguistic features includ- ing pronoun-noun-ratios, pauses, and so on, are used to train classifiers to detect cognitive dis- eases in various tasks. For example, Fraser et al. (2015) achieved up to 82% accuracy on Demen- tiaBank1, the largest publicly available dataset on detecting cognitive impairments from speech, and Weissenbacher et al. (2016) achieved up to 86% accuracy on a corpus of 500 subjects. Yancheva et al. (2015) estimated Mini-Mental State Estima- tion scores (MMSEs), describing the cognitive sta- tus and characterizing the extent of cognitive im- pairment. To improve the accuracy of automated assess- ment using engineered linguistic features, there are usually two approaches: incorporating more clinical data or calculating more features. Tak- ing the first approach, Noorian et al. (2017) in- corporated normative data from Talk2Me2 and the Wisconsin …

Content truncated to 500 words