Natural Language Processing for Cognitive Status Prediction, a Latent Semantic Analysis approach

Sofia De La Fuente Garcia, Saturnino Luz, R Olmos

Research output: Contribution to conferenceAbstract

Abstract / Description of output

Objective: Dementia prevention has become globally paramount across research fields. We present an automatic approach to identify language features associated with cognitive decline. The rationale for using linguistic features is that, in addition to being impaired in dementia, language can act as a proxy for other cognitive abilities (e.g. executive functions) that are also known to be impaired in the disease. Participants/Methods: 461 healthy participants from the HealthyAgeing dataset were included, aged 20-89 years old. They performed a storytelling task, engaging cognitive resources to generate a coherent structured plot. Latent Semantic Analysis (LSA), a Natural Language Processing method, was used to analyse the transcriptions of those tasks. Then, a stepwise regression with working memory (WM) as a dependent variable, was employed to test our hypothesis that some LSA indices would be predictive of WMscores (obtained through largely validated neuropsychological tests). We included the age covariate, as it is well established that it correlates negatively with WM. Results: Almost every LSA index had a statistically significant correlation with both age and WM. Also, the regression model explains 32.3% of the WM scores’ variability. These results hold after controlling for the effect of age. Conclusions: This is an exploratory analysis to assess the potential of narrative discourse to track cognitive decline. Overall, the importance of the model is that, even when controlling for age (known to explain around 25% of the variance in WM), some of our indices remain significant predictors of cognitive status. The novel contribution is the use of LSA, based on word meanings and word-context relations, as an automated way to extract these indices. Our findings could contribute to novel inexpensive and non-invasive screening technologies. We acknowledge Pope, Davis (MUSC) and Wright (ECU) for the HA dataset. The Medical Research Council supports our research.
Original languageEnglish
Publication statusPublished - 20 Jul 2018
EventThe International Neuropsychological Society: Bridging Science and Humanity - Prague, Czech Republic
Duration: 18 Jul 201821 Jul 2018


ConferenceThe International Neuropsychological Society
Abbreviated titleINS 2018
Country/TerritoryCzech Republic


Dive into the research topics of 'Natural Language Processing for Cognitive Status Prediction, a Latent Semantic Analysis approach'. Together they form a unique fingerprint.

Cite this