Summary
A medRxiv preprint reports that StrokeAgent matched expert-adjudicated final decisions in 86% of a curated set of 100 acute-stroke cases. The system processes clinical records, laboratory results, ECG waveforms and CT images as information becomes available.
A medRxiv preprint describes StrokeAgent, an AI decision-support system that processes clinical records, laboratory results, raw electrocardiogram (ECG) waveforms and CT images during the acute stroke pathway. In a curated evaluation of 100 patients, the researchers report that its final decisions agreed with expert-adjudicated decisions in 86% of cases. The system is designed to update its recommendations as new information arrives, rather than make a single assessment from a fixed set of inputs.
Following decisions across the stroke pathway
Acute ischemic stroke care can involve several linked decisions under emergency time pressure. The study describes StrokeAgent as supporting emergency triage, the initial decision about thrombolysis and final reperfusion planning. Thrombolysis uses medication to dissolve a clot; reperfusion treatment aims to restore blood flow to the brain.
The system combines different forms of patient information, including raw ECG signals and source CT images, and updates its recommendations at successive decision points. The researchers say each recommendation is accompanied by an auditable chain of supporting findings and evidence. This design is intended to make the basis for a recommendation traceable as the clinical picture develops.
Results from a curated 100-patient evaluation
The researchers evaluated StrokeAgent on 100 patients with multimodal clinical and imaging data. The set was deliberately enriched with clinically complex reperfusion scenarios alongside standard presentations. Against the expert-adjudicated final decisions, StrokeAgent achieved 86% concordance. Under matched information content and timing, this was 13 percentage points above the mean performance of three tool-free large language models (LLMs) and 17 points above the mean for six resident and attending neurologists.
Agreement across every applicable decision point in the pathway was 76%. The researchers also report that StrokeAgent completed the full decision pathway, on average, 73.9 seconds faster than the case-level median neurologist time.
A separate set of 1,200 blinded clinician assessments rated the system’s outputs. StrokeAgent received higher overall clinical-utility ratings than each LLM comparator, with reported common odds ratios ranging from 10.8 to 16.3. Its reasoning-quality ratings were also higher, and fewer outputs were judged to pose severe potential harm.
These findings describe a preprint evaluation on a curated case set, not a prospective assessment of care delivered to patients. The authors say prospective studies are needed to determine whether physician-supervised use improves clinical care. The results therefore indicate a potential role for auditable, multimodal decision support, while its value in routine clinical workflows remains to be tested.