Longitudinal changes of automated speech markers in MCI and mild AD

Abstract

AbstractBackgroundMild cognitive impairment (MCI) and mild Alzheimer’s disease (mAD) are characterized by decline in memory and other cognitive domains, including language. Previous studies have shown that MCI and mAD patients produce impaired speech. In this study, we investigated longitudinal changes in lexical and acoustic speech features from natural speech of MCI and mAD patients using highly reliable automated methods.MethodWe analyzed 116 digitized 1‐minute picture descriptions produced by 24 MCI (65.8±9.2y, 9 females, mean mini‐mental state exam (MMSE) score at the first recording (T1)=25.4±1.1) and 19 mAD patients (66.9±8.8y, 5 females, mean MMSE at T1=21.9±1.1). The groups did not differ in demographic characteristics, disease duration, and mean inter‐sample interval. Using automated pipelines, we tagged the part‐of‐speech categories of all words, rated words for word frequency, familiarity, concreteness, and semantic ambiguity, and counted partial words and total number of words. Speech samples were segmented into speech and silent pause segments to calculate mean speech segment duration, pause rate, percent of speech out of total time, and articulation rate. Linear mixed‐effects models examined changes over time for each measure, controlling for disease duration at T1. The difference between the first and the last recordings (mean=24±10.9 months) in each measure was related to changes in MMSE, covarying for disease duration at T1.ResultPatients’ adjectives (beta=‐0.04, p=0.006) and determiners (beta=‐0.06, p=0.01) decreased over time, whereas pronouns increased (beta=0.06, p=0.014). Patients produced more frequent (beta=0.004, p=0.043) and more familiar (beta=0.001, p=0.046) words over time. Patients articulated more slowly (beta=‐0.02, p=0.001) and produced more partial words (beta=0.04, p=0.004) and fewer total words (beta=‐0.9, p=0.018) over time. Patients produced shorter speech segments (beta=0.008, p=0.021), paused more frequently (beta=0.32, p=0.001), and produced less speech (beta=‐0.35, p<0.001) over time. All measures showed no group interaction. Patients whose MMSE scores decreased spoke more slowly (beta=2.25, p=0.04) and produced higher‐frequency words (beta=‐6.4, p=0.043) and more partial words (beta=‐0.72, p=0.028) compared with patients whose MMSE scores did not change.ConclusionMCI and mAD patients’ language changed over time, and this was significantly related to patients’ cognitive decline. These findings support the use of automated speech analyses in studying longitudinal changes in neurodegeneration.