Large language models

Where the papers sit

50 papers from the last 6 months study large language models directly · updated from PubMed 1 Nov 2026

• Healthcare AI Safety and Equity

  Healthcare LLM work is shifting from headline accuracy toward bias, privacy, equity and oversight. Evaluations span prescription review, dermatology, MRI findings, EHR symptom extraction, spiritual pain and suicidal ideation, alongside education and therapy. Sensitivity, specificity and human agreement recur, but performance varies by domain and deployment setting. Local models and multidisciplinary or multimodal systems are being explored to protect sensitive data without losing clinical utility.

  13 papers · 26%

• Clinical Education and Support

  Clinical education is moving from static text toward interactive virtual patients, 3D avatars, automated feedback and shared-decision training. LLMs are also being tested for readability, symptom assessment, treatment planning, radiology impressions and emergency recommendations. Early evaluations emphasize feasibility, acceptance and agreement with human supervisors, while cross-sectional designs dominate. Reliability, domain adaptation and safe integration into clinician and therapist workflows remain open questions.

  10 papers · 20%

• Clinical Decision Support

  Clinical decision support is moving toward specialty-aware assistants that combine case reasoning, structured extraction and patient communication. Work spans hepatopancreatobiliary oncology, oncology nursing, rheumatic-disease infections, audiology, postoperative complications and cancer survivorship. Tool use for trial eligibility and coding aims to connect unstructured records with actionable decisions. Protocols and feasibility studies still predominate, leaving prospective validation, safety and effects on care unresolved.

  10 papers · 20%

• Clinical Record Language Processing

  Clinical language processing is moving toward longitudinal, structured use of electronic health records rather than isolated text generation. Applications include chronic kidney disease diagnosis, infectious-disease extraction, anonymization, semantic interoperability and catheter-associated infection surveillance. Fine-tuning, rule-based hybrids and quantum-enhanced architectures are compared with clinician or specialist assessments using sensitivity, specificity and F1 score. The recurring goal is reliable data conversion for prevention and value-based care, with generalization and privacy still open.

  7 papers · 14%

• Clinical Reliability and Safety

  Clinical evaluation is increasingly testing whether LLM answers are safe, complete, appropriate and calibrated for real patient-facing use. Applications include urological and andrological health management, rare-disease triage, reproductive counselling and drug-induced liver injury prognosis. High safety can coexist with missing information, and confidence does not always track accuracy. Clinician-centered frameworks and technology-acceptance measures point toward deployment decisions based on usability and evidence, not accuracy alone.

  6 papers · 12%

• Research Analysis Automation

  LLMs are being tested as research co-pilots for statistical analysis, thematic coding and risk-of-bias assessment. The emphasis is shifting from effortless “vibe coding” toward benchmarking failures, reproducibility and explicit oversight. Prompt engineering changes performance, but it does not remove errors in code, interpretation or methodological judgment. Human review remains necessary before analyses or evidence assessments enter clinical research workflows.

  4 papers · 8%