Digital speech‐based measures separate those with cognitive impairment from those without. (31st December 2021)

• Record Type: Journal Article
• Authors: Stegmann, Gabriela M; Hahn, Shira; Liss, Julie; Berisha, Visar; Mueller, Kimberly D
• Is Part Of: Alzheimer's & dementia. Volume 17(2021)Supplement 6
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-12-31
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.056454
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Measures of language have demonstrated value in the detection of cognitive decline. Here we automatically extracted speech (audio) and language (transcripts) features from Cookie Theft picture descriptions (BDAE) to develop two classification models separating healthy participants from those with mild cognitive impairment (MCI) and dementia. Method: We used Dementia Bank and Wisconsin Registry for Alzheimer's Prevention (WRAP) datasets and evaluated the first observation of Cookie Theft descriptions from 1011 healthy participants: 51 with MCI (based on MMSE scores) and 193 with dementia. From the samples, we automatically extracted metrics to fit two classification models, which: (1) separated healthy from MCI participants; and (2) separated healthy from dementia participants. To avoid overfitting and spurious patterns in the data, we used k ‐fold cross‐validation, restricted the feature space to only clinically meaningful metrics, and used simple logistic regression models with no more than 5 predictors. Result: The MCI classifier separated healthy and MCI groups (ROC AUC = .80; Figure 1). Predictors were: MATTR (proportion of unique words on 10‐word moving average; b = ‐16.0, z = ‐4.7, p‐val < .05); the ratio of pronouns to nouns (b = 28.1, p ‐value < .05), and propositional density (b = ‐15.0, p‐value < .05). The dementia classifier separated healthy and dementia groups (ROC AUC = .89; Figure 2). The predictors were: parse tree height (sentence complexity; b = ‐1.3, p ‐value < .05), mean length of word (surface form; b = ‐3.9, p‐val < .05), the proportion of relevant details correctly identified in a picture description (b = 10.7; p < .05), type‐to‐token ratio (proportion of unique words; b = ‐11.0; p ‐value < .05), and the duration of pauses relative to the total speaking duration (b = 1.1, p‐value = .22). Conclusion: Both classifiers separated groups with ROC AUC >= 0.80, with the dementia classifier performing better than MCI classifier. Features in the models were clinically meaningful and interpretable, and relate to vocabulary (MATTR, propositional density, TTR, mean word length), syntax (parse tree height), language processing (pause duration/speaking duration), and ability to convey relevant picture details.