Local LLM-based detection of iodinated contrast media allergy mentions in CT reports
In the emergency department, a reported allergy to iodinated contrast media frequently leads to withholding contrast injection, although true allergies are rare and delayed allergy work-ups are seldom performed. The result is lost time and lost opportunity for patients who could have benefited from a contrast-enhanced examination. ALLIANCE uses locally hosted large language models (LLMs) to analyze CT reports from the hospital's clinical data warehouse and identify mentions of iodinated contrast media allergy, in order to objectify them and open a dedicated allergology referral pathway.
~253,000 CT reports
CHU de Lille clinical data warehouse
Pre-selection filter (Word2Vec) followed by local LLM analysis; no data leaves the institution
Human review of a report sample to estimate sensitivity and specificity
CHU de Lille internal call for projects
€30,000 requested (research engineer + operating costs)
Primary: Automatically and reliably detect mentions of iodinated contrast media allergy in CT reports.
Secondary:
The STaR-AI group contributes emergency medicine clinical expertise: analysis of management changes linked to presumed allergies, induced delays and throughput times, based on timestamps available in the clinical data warehouse.