Comprehensive Summary
This paper systematically reviews studies that applied natural language processing (NLP) to speech from individuals at clinical high risk for psychosis (CHR-P) to evaluate both diagnostic accuracy (CHR-P vs healthy controls) and prognostic accuracy (prediction of transition to psychosis). The authors searched PubMed, Scopus, and Embase through May 20, 2025, screened 356 records, and included nine studies representing eight cohorts (total ~353 CHR-P and 197 controls). Reviewed work used a variety of NLP pipelines — from earlier latent semantic analysis to newer contextual embeddings (Word2Vec, USE, SBERT), graph metrics, and semantic/coherence measures — and a range of classifiers and validation strategies. Across cross-sectional discrimination studies reported accuracies ranged widely (≈56–95%, AUCs often high), while longitudinal transition prediction studies reported high but variable accuracies (≈83–100%), with prognostic performance more sensitive to methodological choices and validation strategy. The review highlights consistent promise for speech-based NLP as a biomarker but stresses that heterogeneity in methods, small sample sizes, language and corpus differences, and limited external validation limit generalizability and clinical translation.
Outcomes and Implications
For clinicians and researchers, the review suggests that automated speech analysis could become a practical, low-cost tool for early detection and risk stratification in psychosis — potentially helping target early interventions and shorten untreated psychosis. However, the authors caution that current evidence is preliminary: many studies used small or overlapping cohorts, varied feature pipelines, and internal validation that risks optimistic estimates. To safely move toward clinical use, the field needs larger, multilingual, and longitudinal cohorts, standardized speech elicitation and feature-extraction protocols, stronger external validation, and clear governance around privacy and bias. Until those steps happen, NLP should be considered a promising adjunct for research and screening, not a standalone diagnostic or prognostic tool.