Diagnostic and transition accuracy of natural language processing in high risk for psychosis individuals: A...

Schizophrenia spectrum disorders often emerge in adolescence or early adulthood and are a leading cause of global disability. Early identification of clinical high‑risk for psychosis (CHR‑P) can reduce comorbidity and shorten untreated psychosis duration, yet clinician‑administered tools (e.g., SIPS/SOPS, CAARMS, PSYCHS) are time‑consuming and only moderately predictive.

To systematically review the diagnostic and prognostic accuracy of natural language processing (NLP) applied to speech in CHR‑P populations, and to map methodological trends and gaps.

We searched PubMed, Scopus, and Embase through May 2025 for English or Spanish studies enrolling CHR‑P individuals by validated criteria, applying NLP to speech transcripts, and reporting quantitative metrics (accuracy, sensitivity, specificity, AUC‑ROC). Two reviewers independently screened studies, extracted data, and assessed bias with QUADAS‑2; disagreements were resolved by a third reviewer.

Results: From 356 records, nine studies (eight unique cohorts; N = 353 CHR-P, 197 controls) met inclusion. Four case-control studies and one prospective cohort assessed cross-sectional discrimination of CHR-P from healthy controls, reporting accuracies of 56-95 % (AUC-ROC 0.86-0.99). Four prospective studies examined transition prediction, with accuracies ranging from 83 % to 100 %. Studies covered five languages and employed diverse NLP pipelines (e.g., LSA, Word2Vec, USE, SBERT, graph metrics, sentiment analysis). However, feature heterogeneity, small samples (≤ 50 CHR-P), varied speech tasks, and inconsistent validation limited comparability.

NLP‑based speech analysis shows promise as an objective biomarker for early psychosis detection and risk stratification. To advance clinical utility, future research should adopt standardized protocols, recruit larger and more diverse cohorts, and implement multicenter validation.