Stroke detection in medical emergency calls: a retrospective analysis and exploratory evaluation of an AI decision support model.

BACKGROUND: Timely identification of acute stroke during medical emergency calls is critical for optimizing patient outcomes. We aimed to (i) characterize the prehospital trajectories of patients with stroke and (ii) explore the ability of an artificial intelligence (AI) model developed within the Artificial Intelligence Support in Medical Emergency Calls project to support decision-making at Emergency Medical Communication Centres (EMCCs). METHODS: We conducted a retrospective analysis using data from 1,164 patients diagnosed with stroke from an EMCC (2018-2019), focusing on those primarily assessed by EMCC operators. We categorized patients into optimal and nonoptimal trajectory groups and applied logistic regression to explore the factors associated with an optimal trajectory. An AI model was trained using a dataset of 2,980 emergency calls (2019 and 2022), integrating transcribed audio logs and structured clinical data. The model was evaluated on the full dataset and subgroups that were incorrectly assessed by the EMCC. RESULTS: Aphasia/dysarthria was the only factor associated with optimal trajectories. The AI model achieved a sensitivity of 81.0%, specificity of 79.6%, and F1-score of 64.7%. The subgroup analysis included 14 false-negative and 41 false-positive cases. The model correctly predicted stroke in 9 of 14 false-negative cases and ruled out stroke in 9 of 41 false-positive cases. CONCLUSIONS: This explorative evaluation indicates that a multimodal AI pipeline combining audio and structured clinical data may show potential to support decision making in medical emergency calls. Internal results are promising, but studies on larger and external datasets are needed to validate these findings.