Clinical Language Models in Precision Psychiatry: A Review of Treatment Prediction, Causal Inference, and Deployment Barriers Clinical natural language processing; language models; treatment recommendation; individualized treatment effects; precision psychiatry; causal inference
Main article
Abstract
Treatment recommendation systems in psychiatry remain limited by the modest individual-level predictability achievable from conventional clinical predictors. This review examines the proposition that clinical language, processed through modern language models and integrated within causally aware estimators, supplies a workflow-native substrate for raising that ceiling. We synthesise the recent literature across four layers: language sources, feature extraction, modelling, and decision support. The evidence suggests that the addition of clinical text to structured predictors produces a consistent though moderate gain in discrimination across treatment-response prediction tasks, with the largest benefits emerging in calibration and net benefit rather than raw discrimination. We further examine the distinct issues raised by language inputs for causal validity, distributional drift, sociolinguistic fairness, privacy, and interpretability, drawing on the deployment-assurance literature for clinical artificial intelligence. The review concludes that language-informed treatment recommendation has crossed the threshold from speculative to credible, but that translation to durable clinical impact depends on prospective multi-site validation, joint deployment of language models with causal estimators, and the disciplined exercise of an assurance stack that the field is still consolidating.
