Changes in conversational pattern as a clinical trial outcome: a proof-of-concept study using the I-CONECT data

Recent advancements in natural language processing show that linguistic measures can detect subtle cognitive decline. However, most analyses rely on restrictive tasks (eg, picture descriptions). In this proof-of-concept study, we use semantic noise (SN) to quantify conversational patterns and monitor changes following a behavioral intervention.

We analyzed text derived from semi-structured conversations of the I-CONECT project. Participants in the intervention group, who were either cognitively normal (normal cognition [NC]) or had mild cognitive impairment (MCI), engaged in frequent conversational interactions (4 times/week) via an internet/webcam for 6 months. The mean semantic noise (MSN) from the second week and 24th week was used to represent the baseline and month 6, respectively.

Using a 2-tailed t-test, baseline MSN was significantly higher in the MCI group than in the NC group (MCI: 2.09 [SD = 0.34], NC: 1.87 [SD = 0.35], p = .024). MCI and NC did not differ significantly 6 months post-intervention (MCI: 1.81 [SD = 0.47], NC: 1.95 [SD = 0.47], p = .280). Using the paired 2-tailed t-test, MSN in the MCI group decreased significantly from baseline to month 6 (p = .009), while no significant difference was observed in the NC group (p = .426).

After 6 months of conversational engagement, the MCI group's MSN came to resemble that of the NC group. These findings suggest that linguistic measures derived from spontaneous, naturalistic conversations may be sensitive to cognitive changes over time. Further research on extracting linguistic measurements from spontaneous conversations could facilitate cognitive monitoring in daily life.

NCT02871921.