Praise for clinical applications of linguistic analysis

« previous post | next post »

From the abstract of Sunghye Cho et al., "Lexical and Acoustic Speech Features Relating to Alzheimer Disease Pathology", published in Neurology on 4/29/2022:

Background and Objectives: We compared digital speech and language features of patients with amnestic Alzheimer’s disease (aAD) or logopenic variant primary progressive aphasia (lvPPA) in a biologically confirmed cohort and related these features to neuropsychiatric test scores and CSF analyses.

[…]

Discussion: Our measures captured language and speech differences between the two phenotypes that traditional language-based clinical assessments failed to identify. 

From an editorial by Federica Agosta and Massimo Felippi, "Natural Speech Analysis: A Window Into Alzheimer Disease Phenotypes", published in Neurology on 5/4/2022:

Compared to a standard language assessment, the automated analysis of natural speech is more complex and requires a larger amount of time to be post-processed. On the other hand, as is well demonstrated by this study, analysis of natural speech provides information at several levels of language production. Even though data are extracted from only one recorded minute of speech, the tool is able to detect subtle differences among groups, reflecting the patient’s daily experience in a more realistic way than the standard speech and language assessment. Its use has already produced important achievements in distinguishing different language phenotypes. Furthermore, differently from other studies, the work of Cho et al proposed an automated and reproducible method that highly reduces the time of speech analysis and increases the inter-rater reliability.

Since Neurology advertises itself as "The most widely read and highly cited peer-reviewed neurology journal", this editorial is an important indication of a general trend. As I wrote in the abstract for a presentation at UT Dallas in 2019, "Clinical Applications of Human Language Technology: Opportunities and Challenges" —

We infer a lot from the way someone talks: personal characteristics like age, gender, background, personality; contextual characteristics like mood and attitude towards the interaction; physiological characteristics like fatigue or intoxication. Many clinical diagnostic categories have symptoms that are manifest in spoken interaction: autism spectrum disorder, neurodegenerative disorders, schizophrenia, and so on.

The development of modern speech and language technology makes it possible to create automated methods for diagnostic screening or monitoring. More important is the fact that these diagnostic categories are phenotypically diverse, representing (sometimes apparently discontinuous) regions of complex multidimensional behavioral spaces. We can hope that automated analysis of large relevant datasets will allow us to do better science, and learn what the true latent dimensions of those behavioral spaces are. And we can hope for convenient, inexpensive, and psychometrically reliable ways to estimate the efficacy of treatments.

I'll present some suggestive preliminary results, and discuss future research opportunities as well as the existing barriers to progress.

The main "barrier to progress" is  the lack of adequate data. The cited Neurology paper was based on picture-description recordings from just 93 subjects (44 with aAD, 21 with lvPPA, 28 healthy controls), with about a minute of speech from each subject. Such small numbers are unfortunately typical.

There are efforts underway to improve the situation, though we should be concerned that premature commercialization may lock data up in the grip of a few large companies with limited ideas. The Human Genome Project managed to prevent such an outcome for DNA data more than 30 years ago. The outcome for "digital biomarkers", linguistic and otherwise, remains uncertain.

Full disclosure: I'm one of the eight co-authors of the cited Neurology paper. But there are many other researchers around the world who have been working for years on computational approaches to the linguistic analysis of clinical data, and who are responsible for a growing record of success, and for gradual acceptance by the wider community of researchers and funders.

 



1 Comment

  1. Jerry Packard said,

    May 11, 2022 @ 9:56 pm

    Impressive!

RSS feed for comments on this post