Spontaneous Language May Hold Clues Into Alzheimer’s Disease

Older woman talking laughing exercise yoga mat

Just as eyes are windows to the soul, so too words are windows to the mind. We can easily infer a person’s mood from the adjectives she uses, her age from her syntax, her education level from her vocabulary. Yet, it is far more difficult to detect linguistic clues into specific brain conditions, such as Alzheimer’s disease (AD) dementia. Relevant signs can be identified by highly trained experts, but these few individuals are hugely outnumbered by persons with AD, surpassing 43 million worldwide and growing steadily. This is an unfortunate scenario since, unlike other approaches, linguistic assessments offer a low-cost, non-invasive, culture-sensitive framework to support clinical diagnosis. Would it not be great if we had tools to do this objectively and automatically?

This is the question underlying our team’s recent paper in Alzheimer’s & Dementia: Diagnosis, Assessment & Disease Monitoring (a journal of the Alzheimer’s Association). We recruited 21 persons with AD as well as two control groups, composed of 16 healthy individuals and 18 patients with Parkinson’s disease. All participants were Spanish speakers, from Chile, and they were diagnosed by expert multidisciplinary teams. Using a novel battery of our group, we recorded them as they performed everyday verbal tasks, such as describing their routine, recounting a pleasant memory, describing a picture, and retelling a video.

We transcribed their discourse automatically and used computerized tools to examine specific aspects known to be compromised in AD. First, we focused on semantic granularity, that is, the level of specificity of verbal concepts. For example, in naming the picture of a short-legged, long-bodied barking pet, one may use a highly specific noun, such as dachshund, or increasingly vaguer ones, such as dog or animal. Second, we examined ongoing semantic variability, namely, the conceptual closeness of successive words. For example, words are more conceptually close in the utterance The cook was in the kitchen and took a knife than in its counterpart Someone was in a place and took a... What was it? Well, something. To our knowledge, this is the first time such phenomena are captured automatically in AD.

Compared to healthy persons, individuals with AD exhibited significant differences in both measures. In particular, their discourse involved greater use of unspecific concepts, reduced use of specific concepts, and more discontinuous conceptual choices overall. Importantly, no systematic differences were observed between healthy persons and individuals with Parkinson’s disease. Also, using machine learning algorithms, we showed that these features robustly identified individual persons with ADD relative to healthy persons, but they failed to identify patients with Parkinson’s disease.

Our study shows that well-established aspects of AD, so far evident only to selected specialists, can be captured with automated tools that do not require elevated clinical expertise. No less important is the absence of such results in Parkinson’s disease. Indeed, other linguistic aspects more systematically assessed in DD, such as verbal fluency, are also frequently affected in this other condition, limiting their use for disease differentiation.

This breakthrough is part of our ongoing efforts to establish early linguistic markers of diverse brain diseases. Further work in this direction may open new avenues to support mainstream diagnostic tests while circumventing the limitations of specialist-dependent assessments. In line with the GBHI philosophy of equity in brain health, we trust that this research will favor more equitable clinical opportunities for countless individuals across the globe.

Reference: Sanz, C., Carrillo, F., Slachevsky, A., Forno, G., Gorno Tempini, M. L., Villagra, R., Ibáñez, A., Tagliazucchi, E. & García, A. M. (2022). Automated text-level semantic markers of Alzheimer’s disease. Alzheimer’'s & Dementia: Diagnosis, Assessment & Disease Monitoring, 14(1), e12276. doi: 10.1002/dad2.12276