Demo: A Local Multi-Pass LLM System for Automated Code Stroke Identification from Emergency Department Triage Notes

Delayed Code Stroke activation worsens outcome, and 10–20% of strokes are still missed at emergency department (ED) triage. We present a working, fully local multi-pass large language model (LLM) pipeline that reads free-text triage notes and classifies whether institutional Code Stroke criteria are met. Sequential model calls handle translation of abbreviations, focal-symptom detection, mimic exclusion, baseline functional exclusions, onset-time windowing, and symptom-resolution checking; deterministic logic then emits CAT-1, CAT-2, or NO-CODE. Inference runs on a single consumer GPU via Ollama with 4-bit open-weight models (8–14B) and no internet, fine-tuning, or patient-data exposure. On 3,023 consecutive adult ED notes from one month at a large Australian tertiary hospital, the best chained configuration (Qwen2.5 14B) reached sensitivity 0.890 and specificity 0.993 against a blinded neurologist reference standard. Collapsing the same criteria into one prompt collapsed specificity for Llama 3.1 8B from 0.952 to 0.581. An interactive demo is at https://scanstroke.ai