$\texttt{M}^3$Stroke: $\textbf{M}$ulti$\textbf{M}$odal $\textbf{M}$obile AI for Emergency Triage of Mild to Moderate Acute Strokes

Tongan Cai, Kelvin Wong, James Z. Wang, Sharon X Huang, Xiaohui Yu, JOHN JULIUS VOLPI, Stephen T. C. Wong

Keywords: Stroke, Artificial Intelligence, Computer Aided Diagnosis, Mobile Computing

TL;DR: We propose an iOS-based multimodal mobile AI tool for mild to moderate stroke triage. The novel core model significantly outperforms triage clinicians and prior SoTA models with high accuracy and efficiency.

Abstract: Over 22\% of ischemic stroke patients are overlooked during triage in the emergency departments, particularly those with mild or moderate stroke which resembles stroke mimics in symptoms. While pronounced neurological conditions can be captured with existing AI solutions, identifying stroke patients with minor symptoms remains under-explored due to data scarcity, noise complexity, and feature subtlety. We propose $\texttt{M}^3$Stroke, a $\textbf{M}$ulti$\textbf{M}$odal $\textbf{M}$obile AI tool, to enhance the accuracy and efficiency of stroke triage for these patients. As the first stroke screening tool to integrate novel audio-visual multimodal AI into efficient mobile computing, $\texttt{M}^3$Stroke runs seamlessly on common iOS devices and significantly outperforms prior methods. Trained and evaluated on a dataset of 269 patients suspected of stroke (191 stroke/78 non-stroke), $\texttt{M}^3$Stroke model achieves 80.85\% accuracy, 60.00\% specificity, and 90.63\% sensitivity, demonstrating 14.29\% gain in specificity and 20.44\% higher sensitivity compared with traditional stroke triage methods. The tool's performance, robustness, and fairness across diverse demographics confirm its potential to improve general emergency triage, aiding tele-stroke detection and self-diagnosis, and enhancing life quality for elderly patients.

Track: 4. AI-based clinical decision support systems

Registration Id: FPNKWCG6JF9

Submission Number: 56