Natural language processing application in electronic health records studies for psychosis: a systematic review

Natural language processing (NLP) enables the use of electronic health records (EHRs) for psychosis research in real-world settings and develops precision psychiatry. We conducted a PRISMA-guided systematic review of five databases from inception to 9 October 2025 and identified 62 eligible studies with over 1 million participants. Most studies (62.9%) used data from the UK Clinical Record Interactive Search (CRIS) platform. Methodologically, 37.1% applied dictionary-based NLP, 53.2% representation-based machine learning, 9.7% transformer-based NLP, and none applied generative large language models (LLMs). Clinically, 74.2% targeted symptom/information extraction, 19.2% classification, and 16.1% risk prediction. However, 56.5% studies self-reported limited generalizability, 32.3% reported data quality/collection challenges, and only two studies performed external validation. Despite promising proof-of-concept performance, evidence of real-world clinical utility remains limited. Current work remains concentrated on retrospective phenotyping within a single healthcare ecosystem. We highlight priorities for multi-system validation, temporal trajectory modeling, and responsible LLMs deployment in psychosis EHRs research.