Objective:

This study describes the development, refinement, and clinical validation of a virtual assistant (VA) based on AI for early stroke detection and appropriate emergency referral.

Background:

Many patients fail to access emergency services in time for acute stroke treatment. Artificial intelligence (AI) may help optimize prehospital triage.

Design/Methods:

A prospective cohort study was conducted between August 2024 and July 2025 in a tertiary care center in Buenos Aires, Argentina. The VA was applied to adult inpatients with acute stroke in a neurovascular unit. Prior to this, the tool had been optimized using a literature review and simulations with 1151 de-identified medical records. Clinical, demographic, and performance variables were recorded. The main outcomes were syndromic diagnostic accuracy, identification of the most probable diagnosis, appropriate emergency referral, and user satisfaction.

Results:

A total of 78 participants were included (median age: 73 years; 56.4% male). In 76.9% of cases, the VA was used by the patient, and in 23.1% by a companion. The mean time from symptom onset to VA use was 2 days. Final diagnoses were ischemic stroke (80.8%), transient ischemic attack (11.5%), subarachnoid hemorrhage (5.1%) and intracerebral hemorrhage (2.6%). Syndromic diagnosis matched the clinical standard in 89.7% of cases; top-1 match in 71.8% and top-3 in 91%. Emergency referral was adequate in 93.6%. Median use involved 10 questions and 4 minutes. Over 90% rated the experience 4 or 5 out of 5.

Conclusions:

The AI-based VA demonstrated strong diagnostic and referral performance, with high user satisfaction. The next step is its initial supervised implementation in clinical practice.