This systematic review and meta-analysis evaluated speech-based computational approaches for diagnosing schizophrenia and assessing symptom severity. This study is a systematic review and meta-analysis conducted in accordance with PRISMA guidelines. A comprehensive search of PubMed, Scopus, and Web of Science databases was performed from inception to August 2025. Eligible studies included individuals diagnosed with schizophrenia and employed computational analyses of speech or language, such as natural language processing, acoustic feature extraction, or large language models, and reported diagnostic accuracy metrics or associations with validated clinical scales. Two independent reviewers screened titles, abstracts, and full texts, extracted data, and assessed methodological quality. When sufficient quantitative data were available, random-effects meta-analysis models were applied to pool diagnostic performance and correlation estimates. Diagnostic performance refers to supervised classification accuracy relative to clinician-assigned reference diagnoses rather than autonomous clinical diagnosis. Thirty-nine studies met inclusion criteria, with 16 contributing data to meta-analysis. Pooled diagnostic performance was high (AUC≈0.84), while speech-derived features showed moderate association with symptom severity (r ≈ 0.34). Large language models and multimodal pipelines outperformed traditional machine-learning and unimodal systems. Speech measures were most sensitive to negative symptoms, including alogia and reduced affect, whereas coherence-based metrics captured thought disorder. Sensitivity analyses confirmed robustness, and no meaningful publication bias was observed. Computational speech analysis demonstrates strong potential as an objective, scalable adjunct for schizophrenia assessment. Future research should prioritize standardized protocols, larger multicenter cohorts, and real-world clinical integration to support translation into practice. Not applicable. Not applicable.
Naeim et al. (Thu,) studied this question.