Automated analysis of free speech predicts psychosis onset in high-risk youths.
Level 3 - non-randomized controlled study
Prospective cohort study tracking high-risk participants over time for psychosis onset
PubMed 27336038 · doi:10.1038/npjschz.2015.30
What was done
In a prospective proof-of-principle study, 34 youths at clinical high-risk (CHR) for psychosis (11 females) completed baseline interviews and were followed quarterly for up to 2.5 years. Baseline interview transcripts were evaluated using automated natural language processing for semantic coherence (via Latent Semantic Analysis) and syntactic complexity markers (maximum phrase length and determiner usage). A convex hull classification algorithm with leave-one-subject-out cross-validation was tested to predict psychosis transition, and canonical correlation assessed the relationship with prodromal symptom ratings.
What was found
Over up to 2.5 years, 5 of 34 participants transitioned to psychosis. The extracted speech features predicted later psychosis onset with 100% accuracy in cross-validation, outperforming classification based on clinical interviews. The speech features were significantly correlated with prodromal symptoms, although specific numerical correlation values, effect sizes, and confidence intervals were not reported in the abstract.
Why it matters
This study provides proof-of-concept that automated natural language processing of interview transcripts can identify objective linguistic markers of emerging psychosis in high-risk individuals.
Limits
The sample size is very small (n=34) with only 5 transition events, creating a substantial risk of model overfitting despite leave-one-subject-out cross-validation. There was no independent external validation cohort, and exact statistical values for the canonical correlation were omitted from the abstract.
Cited by
- supports Automated analysis of syntactic completion and reading patterns can predict the likelihood of psychosis.