Subtle linguistic features in paragraph-recall responses may help identify cognitive impairment and provide information beyond conventional memory scoring.
Investigators analyzed digital voice recordings from 710 participants in the Long Life Family Study. Recordings came from participants at the Boston field site during the second study visit and included immediate and delayed responses to the Logical Memory paragraph-recall task. Among the participants, 598 had normal cognition and 112 had cognitive impairment.
The investigators used Linguistic Inquiry and Word Count software to characterize language in the recall transcripts and identified features associated with cognitive status after accounting for age, sex, education, and family relationships. They combined associated features into weighted polyfeature scores for immediate (PFS-IR) and delayed recall (PFS-DR). The investigators examined whether these scores distinguished cognitive impairment from normal cognition and how their classification performance compared with traditional Logical Memory scoring. A longitudinal analysis examined whether baseline PFS-DR was associated with subsequent performance on the modified Telephone Interview for Cognitive Status (TICS-m), and exploratory analyses assessed linguistic patterns in amnestic and dysexecutive mild cognitive impairment.
Eight immediate-recall and 7 delayed-recall linguistic features were associated with cognitive impairment. During immediate recall, participants with cognitive impairment used more negations, cognitive-process terms, and present-focused words, as well as fewer prepositions and words related to biological processes, work, space, and leisure. During delayed recall, cognitive impairment was associated with greater use of negations and first-person pronouns as well as fewer past-focused, number-related, biological-process, work-related, and drive-related words.
Both polyfeature scores were associated with cognitive impairment. When combined with demographic characteristics, PFS-DR achieved an area under the precision-recall curve of 0.77 vs. 0.81 for traditional delayed Logical Memory scoring, with no statistically significant difference between the models. Combining PFS-IR and PFS-DR with demographic characteristics yielded an area under the precision-recall curve of 0.78. The linguistic scores were moderately correlated with traditional Logical Memory scores, suggesting that the approaches captured overlapping but not identical information.
The longitudinal analysis included 514 participants who completed at least 2 TICS-m assessments following baseline. Higher baseline PFS-DR was associated with poorer cognitive performance during follow-up. A 30-point higher PFS-DR, about 2 standard deviations in the study population, corresponded to a 2-point lower TICS-m score over 7 years. PFS-DR remained associated with cognitive performance when traditional Logical Memory scores were included.
PFS-DR did not predict a faster rate of cognitive decline. The interaction between PFS-DR and time was not statistically significant, indicating that higher baseline scores were associated with lower cognitive performance across follow-up. In an analysis restricted to 449 participants with normal cognition at baseline, the association between PFS-DR and subsequent TICS-m performance was only marginal.
Exploratory analyses identified differences by mild cognitive impairment subtype. Among 24 participants with amnestic mild cognitive impairment, delayed recall was associated with greater use of first-person pronouns, negations, present-focused terms, and cognitive-process words, among other features. Among 46 participants with dysexecutive mild cognitive impairment, delayed recall was associated with greater use of present-focused words and negations and fewer past-focused language and words reflecting clout. The investigators cautioned that the subgroup analyses were underpowered and require replication.
Several limitations may affect interpretation. Traditional Logical Memory scores contributed to the consensus determination of cognitive status, while the linguistic measures were derived from the same recall responses, creating potential circularity that likely inflated predictive accuracy. Feature selection was performed using the full sample rather than an independent training sample, which may have produced overly optimistic estimates. In addition, the linguistic software was not designed for paragraph-recall assessment, did not capture syntax or contextual relationships between words, and did not incorporate acoustic or temporal features of speech. The investigators emphasized the need for validation in more diverse cohorts and clinical samples.
"[T]his study identified novel, subtle features of verbal responses on a brief assessment of episodic memory that were informative for detecting cognitive impairment and predicting cognitive performance over time,” wrote lead study author Seho Park, of the Department of Medicine at Boston University Chobanian & Avedisian School of Medicine, and colleagues.
Full disclosures of the study authors can be found in the study.
Source: Journal of the International Neuropsychological Society
