P1‐535: PREDICTING DEMENTIA FROM WRITTEN TEXTS USING FEATURE EXTRACTION AND MACHINE LEARNING

Natasha Clarke, Tom Barrick, Peter Garrard

Alzheimer s & Dementia · 2018 · 인용 1

The effects of early pathological brain changes due to Alzheimer's disease (AD) may be detectable in written language. Idea density was lower in essays written by Nuns in their twenties and thirties who would be diagnosed with dementia decades later (Snowden et al., 1996), and correlated with post mortem brain weight, atrophy and neurofibrillary pathology (Riley et al., 2005). Studies of the published works of Iris Murdoch found a shrinking vocabulary and reliance on higher frequency words years prior to the detection of disease (Garrard et al., 2005), as well as a decline in grammatical and syntactic complexity (Pakhomov et al., 2011).

Masrani et al. (2017) was able to classify people with dementia compared to controls using text scraped from online blog posts, achieving 0.84 accuracy. Further research is required to explore these findings in a wider population.

Written text samples were collected and transcribed from 81 members of the public (34 dementia, 47 controls) spanning an average of 28 years per person (sd = 12.7). 165 linguistic features encompassing parts of speech, lexical and syntactic complexity, and sentiment were extracted from 814 texts. Five machine learning algorithms were trained to classify patients or control feature vectors using 10-fold cross validation.

Logistic regression was found to have the highest accuracy. Texts were then split in to four groups according to time-from-diagnosis, and four algorithms were trained to classify texts in each group. Classification accuracy for texts written post-diagnosis was comparable to previous studies at 0.83.

When comparing texts written by patient and control groups in periods up to 20 years before diagnosis, classification accuracies were more variable and depended on combinations of selected features. There was no clear pattern in features that increased the odds ratio for dementia across the time span. AD pathology may be detectable through computational analysis of written language, but a single feature or group of features were not indicative of disease status.

Future studies should aim to discover improved methods for extracting, detecting and quantifying linguistic change years prior to diagnosis.

📄 이 논문을 인용한 Paperis 글

이 논문이 근거 목록에 올라 있는 Paperis 글입니다.