Development and Evaluation of Multimodal Universal CPR AI Assistance and Response Engine (U-CARE)

Each year, 350,000 people in the United States experience out-of-hospital cardiac arrest, and 90% do not survive. Early cardiopulmonary resuscitation (CPR) can double or triple survival rates.

Bystander CPR is performed correctly in fewer than 40% of cases, with major challenges persisting: lack of widespread training, decision paralysis, and recall bias during an emergency.

Universal CPR Assistance and Response Engine (U-CARE) is a low-cost smartphone-based system that serves as a universal CPR guide to address these challenges. A large language model-powered verbal guidance system provides objective next-best steps multilingually, coupled with a computer vision model that continuously assesses CPR technique. U-CARE was evaluated using holistic large language model evaluation and computer vision metrics, and outperformed out-of-the-box frontier models on relevance and safety in human evaluation, demonstrating its viability for real-time feedback.

U-CARE demonstrates the feasibility of a smartphone-based multimodal CPR feedback and response agent for emergency and training guidance.