Generative artificial intelligence in critical care medicine: A narrative review of applications, predictive analytics, documentation, and ethical imperatives

The ICU is a high‑acuity environment with overwhelming data and high‑consequence decisions. Generative AI can integrate clinical narratives, laboratory values, physiologic monitoring, imaging, and genomic data to generate actionable outputs. Challenges include safety, bias, transparency, regulation, and human–AI collaboration.

Generative AI and LLMs in Medicine

Foundation models such as GPT‑4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, and Meta Llama 3 introduce few‑shot learning, chain‑of‑thought reasoning, and multimodal integration, combining natural language processing, unsupervised pretraining, reinforcement learning from human feedback, and retrieval‑augmented generation (RAG). Key features include interpreting ambiguous scenarios, generating structured documentation, assisting diagnostic reasoning, producing individualized recommendations, and integrating multimodal data.

Prompt engineering: Strategies include zero‑shot/few‑shot prompting, chain‑of‑thought prompting for stepwise reasoning, role prompting to simulate expert consultation, multimodal prompting integrating labs, waveforms, ventilator parameters, and imaging, and context‑enriched prompting incorporating longitudinal trajectories and goals of care.

Applications in Critical Care

Clinical decision support: Generative AI processes unstructured data—notes, laboratory trends, consultant narratives, imaging reports—to generate patient‑specific recommendations, functioning as an AI medical consultant. GPT‑4 can match or exceed clinician performance on complex diagnostic vignettes. Multimodal systems combining vital signs, labs, orders, and nursing notes using RAG produce individualized risk assessments with narrative explanations. Ventilator weaning, delirium monitoring, and ARDS phenotyping remain conceptual. In acute kidney injury, models integrate labs, urine output, medication exposure, hemodynamics, and imaging to generate risk estimates and management options.

Knowledge management: Tools such as OpenEvidence use RAG to synthesize biomedical literature into evidence‑based summaries, supporting diagnosis of rare presentations and hypothesis generation.

Education: Generative AI supports automated curriculum development, AI‑generated clinical vignettes and simulation cases, adaptive learning, and virtual teaching assistants.

Documentation and EHR automation: Ambient AI scribes (Nuance Dragon Ambient eXperience, Abridge) transcribe conversations and populate EHR fields in outpatient settings; ICU use is limited because many patients are sedated or ventilated. Nevertheless, generative AI can synthesize labs, monitoring, imaging, and consult notes into daily progress notes, multi‑day summaries, discharge documentation, and structured handoff reports (SBAR, I‑PASS). Outputs remain subject to clinician review.

EHR integration and workflow: LLM‑embedded interfaces allow natural‑language queries—e.g., whether hypotension developed over 24 hours—by analyzing flowsheets, monitoring trends, labs, and notes. Ambient listening may capture team discussions and generate draft orders for clinician review.

Sepsis prediction and early warning: Real‑time monitoring translates abnormal trends into actionable narratives—e.g., rising lactate and heart rate suggesting evolving shock—rather than generic alarms. Diffusion models detect anomalies by reconstructing expected trajectories; LLMs augment risk scores with clinical context.

Real‑world barriers: Fragmented EHR integration, interoperability gaps, limited real‑time infrastructure, and downtime reliability remain challenging. Clinician acceptance depends on trust and medico‑legal accountability. Costs and reimbursement pathways complicate sustainability.

Data augmentation: LLMs generate synthetic interim notes to fill temporal gaps, improving mortality prediction and narrowing performance gaps between patients with sparse versus extensive documentation. LLMs also distill narratives into prognostically relevant features, improving readmission and length‑of‑stay prediction.

Synthetic data generation: GANs, diffusion models, and variational autoencoders create synthetic ICU datasets for rare events (ARDS, tamponade, adverse drug events). Synthetic time‑series added to limited real‑world data improved prediction of vasopressor requirements and mechanical ventilation. Synthetic data also supports simulation‑based evaluation and IT safety testing without compromising privacy.

ICU data summarization: Generative AI transforms waveforms, vital signs, ventilator parameters, labs, and medication records into concise narratives—e.g., declining mean arterial pressure requiring norepinephrine escalation with rising lactate—enhancing situational awareness and reducing cognitive burden.

Legal, Ethical, and Regulatory Considerations

Explainability: AI systems must provide transparent outputs; chain‑of‑thought reasoning and RAG citations improve interpretability. The FDA and EU AI Act require transparency.

Safety: Hallucinations—fabricated information with unwarranted confidence—risk incorrect diagnoses. Safeguards include RAG, external knowledge integration, post‑processing validation, performance monitoring, and independent verification.

Bias: Models trained on historical data may perpetuate disparities. Mitigation includes diverse datasets, bias auditing, subgroup evaluations, and fairness constraints.

Liability: Responsibility when AI recommendations contribute to adverse outcomes remains unresolved; clinicians are generally accountable as final decision‑makers, though shared liability is debated.

Regulation: The FDA regulates AI as medical devices; the EU AI Act classifies healthcare AI as high‑risk. China mandates algorithm registration; Singapore uses governance sandboxes; India prioritizes healthcare in its AI strategy. Institutional governance committees including clinicians, ethicists, data scientists, and patient representatives are recommended.

Consent and communication: Transparent communication about AI's role is warranted. Informed consent is challenging where patients lack decisional capacity. AI tools can translate complex findings into accessible narratives, supporting shared decision‑making.

Future Perspectives

AI agents may execute multistep workflows (sepsis bundles, antimicrobial stewardship, ventilator liberation) under clinician oversight. EHR‑embedded models may serve as clinical co‑pilots. Precision critical care integrating physiologic monitoring with genomic and biomarker data may enable individualized prognostication. Diffusion models, GANs, and RAG‑enhanced multimodal AI may enable real‑time CRRT adjustments, ARDS phenotyping, and bias‑mitigated sepsis prediction. Safe deployment requires ongoing evaluation and multidisciplinary collaboration.

Limitations

As a narrative review, no formal systematic search or quantitative synthesis was performed. The field evolves rapidly, and cited models may change. Many applications remain conceptual or supported by retrospective studies rather than prospective trials. Performance and generalizability vary across institutions, populations, and resource settings.

Conclusions

Generative AI represents a potential paradigm shift pending validation, with current evidence supporting documentation assistance, knowledge synthesis, and prototype EHR integration. Emerging applications remain largely speculative. With oversight and ethical safeguards, generative AI may augment efficiency, safety, and precision care. Generative AI will not replace intensivists but may become a foundational co‑pilot in high‑acuity care.

Full text

https://www.xiahepublishing.com/2590-3438/JTCCM-2025-00022

The study was recently published in the Journal of Translational Critical Care Medicine.

Journal of Translational Critical Care Medicine (JTCCM), an open access publication academically supported by The First Affiliated Hospital of Harbin Medical University, is a peer-reviewed online journal with continuous print on demand compilation of issues published. The journal allows free access (Open Access) to its contents and permits authors to self-archive final accepted version of the articles on any OAI-compliant institutional/subject-based repository.

The journal covers technical and clinical studies related to health, ethical and social issues in field of Critical Care Medicine, Translational Medicine, Respirology, Infection, Tumor, Surgery. Articles with clinical interest and implications will be given preference.

Follow us on X: @xiahepublishing

Follow us on LinkedIn: Xia & He Publishing Inc.