Objective:

To develop a novel method for predicting large vessel occlusion (LVO) stroke using cranial accelerometry (CA) without reliance on electrocardiogram (ECG) or neurological examination data.

Background:

CA has shown promise in LVO stroke diagnosis for prehospital triage to thrombectomy centers, but previous methods often relied on ECG for waveform segmentation and neurological examination findings, which can be unreliable due to artifacts and subjective assessments.

Design/Methods:

CA recordings were collected from consecutive stroke alert patients presenting to the UC Davis Emergency Department from 2020–2023, enriched with transferred LVO stroke and intracranial hemorrhage patients. 10 second left and right CA recordings were analyzed using continuous wavelet transform (CWT) and cross wavelet transform (XWT). Valid recordings were defined as those with 30 consecutive seconds with amplitude under 30 milli-g. The wavelet transform data was featurized and processed through principal component analysis before being input into a polynomial support vector machine classifier. Diagnoses were obtained through chart and imaging review, and LVO stroke was defined as an acute symptomatic occlusion of the internal carotid, M1, M2, or basilar arteries. 

Results:

Out of 304 patients, 13 were diagnosed with LVO stroke. 219 patients had valid left and right recordings, and 10 of those had LVO stroke. CWT analysis revealed clear periodic activity between 2 and 10 Hz that oscillated roughly every second, corresponding with a patient’s heartbeat. XWT analysis of left and right CA data showed strong correlation, with a phase shift between them that varied in each patient. On the 219 patients, this approach achieved 90% sensitivity and 97.12% specificity in diagnosing LVO stroke.

Conclusions:

When analyzed using signal processing techniques and machine learning algorithms, cranial accelerometry accurately diagnosed LVO stroke without using ECG or neurological examination findings. If validated, this would improve diagnostic performance and simplify data acquisition for prehospital triage.