Abstract Body:
Introduction
The term “headpulse” refers to imperceptible head movements associated with each cardiac contraction, as measured by a cranial accelerometry device with electrocardiogram leads. Prior research at another hospital used asymmetric arm weakness (AAW) and headpulse data to create an algorithm that was 91% sensitive and 93% specific for large vessel occlusion (LVO) stroke. This original study used convenience sampling of emergency department (ED) patients with suspected stroke and transfer patients with known LVO stroke. As a result, 23/68 (34%) of patients had LVO stroke diagnoses. External validation is needed in a separate, larger, and more representative sample of suspected stroke patients where the prevalence of LVO stroke is expected to be much lower (5-15%).

Methods
Headpulse recordings were prospectively obtained during stroke alerts for patients presenting directly to the UC Davis ED from 2020–2023 during business hours. A secondary analysis enriched with transferred LVO stroke and intracranial hemorrhage (ICH) patients was added in response to low enrollment of these diagnoses. Examinations and diagnoses were obtained by chart and imaging review. LVO stroke was defined as an acute symptomatic occlusion of the internal carotid, M1, M2, or basilar arteries. The algorithm being externally validated counted patients without AAW as negative for LVO stroke. For patients with AAW, the previously reported Support Vector Machine model used headpulse data to predict LVO stroke or not (Figure 1). Diagnostic statistics were compared to AAW alone, and Vision Aphasia Neglect (VAN) tests.

Results
Thirteen of 304 patients (4%) were diagnosed with LVO stroke in the primary analysis. The secondary analysis added 11 LVO stroke, 10 ICH, 1 non-LVO stroke, and 2 mimic patients (Table 1). The AAW plus headpulse algorithm undergoing external validation was less sensitive (54 vs 91%) and less specific (87 vs 93%) than the original analysis. AAW was more sensitive but less specific than the algorithm. VAN was more sensitive and slightly more specific (Table 2).

Conclusion
The lower sensitivity of this AAW plus headpulse algorithm during external validation suggests it is not a good candidate for prehospital validation. The small sample of LVO strokes led to wide sensitivity confidence intervals. The additional LVO stroke patients in the secondary analysis did not improve the sensitivity point estimate. Our results highlight the importance of external validation.