Tracing semantic trajectories in attention deficit hyperactivity disorder

Large language models (LLMs) enable the quantification of speech as embeddings in a multidimensional semantic space. By dividing the speech into utterances, we can embed complete-thought-like units on a high-dimensional semantic space. This approach proves particularly relevant for psychiatric research, as neurodivergent conditions like attention deficit hyperactivity disorder (ADHD) have been associated with differences in discourse organisation and pragmatic language ability, including more variable topic range and topic jumps. ADHD shows problems in organisation of structured behaviour by disturbance of prefrontal/executive circuitry. This can manifest in tangential thinking and disturbed organisation of thought content.

The primary aim of this study was to determine whether individuals with ADHD exhibit more frequent topic shifts, operationalised as “jumpiness rate”, compared to healthy controls. A secondary objective of the study was to investigate how different qualities of speech beyond semantic jumpiness in a discourse setting compare between ADHD and healthy controls.

We collected interview data of an ADHD group (n = 74; 40 men and 34 women; mean age = 35.8 years) and healthy controls (n = 23; 14 men and 9 women; mean age = 33.6 years). All participants completed the German version of the Wender–Reimherr Interview (WRI), a structured diagnostic instrument for adult ADHD. Audio recordings were transcribed and segmented into utterances representing complete thought units. Each utterance was embedded using a Multilingual-E5-large-model. Semantic similarity between consecutive utterances was quantified using cosine distance. Topic shifts were defined as distances exceeding a threshold of mean + 1.5 standard deviations. The proportion of such shifts relative to the total number of utterances constituted the jumpiness rate. Additionally, trajectory length was calculated as the cumulative cosine distance across all utterances, representing overall semantic dispersion.

Associations between the captured speech metrics and WRI-subscales were assessed using Spearman rank correlations. Group differences were evaluated using Mann–Whitney-U-tests. To account for variation in interview length, analyses of covariance (ANCOVA) were conducted with group as the independent variable, semantic trajectory metrics as dependent variables, and number of utterances as a covariate.

The ADHD group demonstrated a significantly higher jumpiness rate compared to healthy controls after controlling for utterance count (b = −0.021, t = −3.84, p < 0.001). Trajectory length was greater in the ADHD group descriptively but did not differ significantly after adjustment for interview length (b = −0.028, t = −0.16, p = 0.876). WRI-subscales Attention, Hyperactivity and Disorganisation showed significant correlations with some speech measures, including jumpiness rate and trajectory length and lexical, temporal and prosodic metrics, particularly total word count and total speaking time.

These findings indicate reduced semantic stability in ADHD, reflected in increased variability between consecutive utterances. While global semantic dispersion was not significantly different after controlling for verbosity, multiple speech-derived features were associated with clinical symptom severity. The structured nature of the diagnostic interview may have influenced discourse patterns and represents a limitation. Future studies should extend this approach to other discourse settings to assess generalisability and potential clinical utility.

References

No references