Navigating the translational chasm: a scoping review of large language models and hybrid architectures in intensive care

Foundation models, particularly large language models, can process the unstructured data central to the intensive care record, yet their clinical readiness remains poorly defined. This study aimed to map foundation-model and large language model applications in adult intensive care, characterise their architectural paradigms, and chart their validation and deployment features. Empirical studies applying a large language model or other foundation model to adult intensive care tasks were eligible for inclusion. Traditional machine-learning-only, non-intensive care, and paediatric studies were excluded. PubMed, Web of Science, and Scopus were searched (January 2022-February 2026). A supplementary preprint and conference search through mid-2026 was analysed separately. Data were extracted in duplicate using an operationalised taxonomy (generative large language models, clinical language models, retrieval-augmented generation, and hybrid pipelines). Evidence-maturity mapping was performed without formal risk-of-bias appraisal. Sixty-four studies addressed decision support (51.6%), clinical NLP (23.4%), predictive analytics (20.3%), and ICU operations (4.7%). Standalone configurations predominated (79.7%); retrieval-augmented generation (6.3%) and hybrid (14.1%) architectures concentrated in higher-stakes tasks. Maturity was limited: 46.9% relied solely on retrospective MIMIC data, 6.3% reported external validation, 90.6% remained internally retrospective, and only 3.1% achieved prospective evaluation. Hybrid and retrieval-augmented architectures appear better aligned with high-risk ICU tasks, though no comparative synthesis was performed. Prospective validation and a risk-stratified 'clinical co-pilot' framework are needed.