Privacy-Preserving Clinical Decision Support for Emergency Triage Using LLMs: System Architecture and Real-World Evaluation


KARAMANLIOĞLU A., Demirel B., Tural O., Doğan O. T., ALPASLAN F. N.

Applied Sciences (Switzerland), cilt.15, sa.15, 2025 (SCI-Expanded, Scopus) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 15 Sayı: 15
  • Basım Tarihi: 2025
  • Doi Numarası: 10.3390/app15158412
  • Dergi Adı: Applied Sciences (Switzerland)
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Agricultural & Environmental Science Database, Applied Science & Technology Source, Communication Abstracts, INSPEC, Metadex, Directory of Open Access Journals, Civil Engineering Abstracts
  • Anahtar Kelimeler: clinical decision support, fair data principles, federated learning, generative AI, healthcare architecture, infectious disease triage, large language models, privacy preservation, sepsis alerting, triage
  • Orta Doğu Teknik Üniversitesi Adresli: Evet

Özet

This study presents a next-generation clinical decision-support architecture for Clinical Decision Support Systems (CDSS) focused on emergency triage. By integrating Large Language Models (LLMs), Federated Learning (FL), and low-latency streaming analytics within a modular, privacy-preserving framework, the system addresses key deployment challenges in high-stakes clinical settings. Unlike traditional models, the architecture processes both structured (vitals, labs) and unstructured (clinical notes) data to enable context-aware reasoning with clinically acceptable latency at the point of care. It leverages big data infrastructure for large-scale EHR management and incorporates digital twin concepts for live patient monitoring. Federated training allows institutions to collaboratively improve models without sharing raw data, ensuring compliance with GDPR/HIPAA, and FAIR principles. Privacy is further protected through differential privacy, secure aggregation, and inference isolation. We evaluate the system through two studies: (1) a benchmark of 750+ USMLE-style questions validating the medical reasoning of fine-tuned LLMs; and (2) a real-world case study (n = 132, 75.8% first-pass agreement) using de-identified MIMIC-III data to assess triage accuracy and responsiveness. The system demonstrated clinically acceptable latency and promising alignment with expert judgment on reviewed cases. The infectious disease triage case demonstrates low-latency recognition of sepsis-like presentations in the ED. This work offers a scalable, audit-compliant, and clinician-validated blueprint for CDSS, enabling low-latency triage and extensibility across specialties.