Prediction of psychosis across protocols and risk cohorts using automated language analysis.
Level 3 - non-randomized controlled study
Cohort validation study evaluating a prognostic machine learning classifier across prospective cohorts
PubMed 29352548 · doi:10.1002/wps.20491
What was done
Researchers tested an automated machine learning and natural language processing classifier on transcribed speech from English-speaking youths at clinical high risk for psychosis and patients with recent-onset psychosis. The model, which measures semantic coherence, coherence variance, and possessive pronoun usage, was developed and cross-validated across two separate high-risk cohorts and evaluated for its ability to discriminate recent-onset psychosis patients from healthy controls.
What was found
The automated speech classifier achieved 83% accuracy in predicting psychosis onset within the primary cohort (intra-protocol) and 79% accuracy when cross-validated on the original independent risk cohort (cross-protocol). The model also discriminated the speech of recent-onset psychosis patients from healthy controls with 72% accuracy and correlated closely with manual linguistic ratings.
Why it matters
Automated speech processing offers an objective, non-invasive biomarker to identify high-risk individuals likely to convert to full psychosis, providing potential targets for early clinical monitoring and preventive intervention.
Limits
The abstract does not report the exact sample size (n) for the cohorts evaluated. The classifier was trained and validated exclusively on English-speaking participants, meaning findings cannot be generalized to other languages or linguistic cultures without further cross-linguistic validation and testing in larger cohorts.
Cited by
- supports Automated analysis of syntactic completion and reading patterns can predict the likelihood of psychosis.