Presentation (CTI)

OBJECTIVES: The CPR composite error action recognition system based on artificial intelligence technology can effectively assist doctors in CPR skill assessment, w
METHODS: A vision-based system was constructed through literature research and expert consultation to define 13 single and 74 composite error actions of extracorporeal cardiac compression in CPR. From July to September 2023, we collected 500 CPR action videos in Zhongshan hospital and constructed a fine-grained composite error action set named CPR-Coach. In addition, we proposed a neural network that can predict composite errors while training single-error samples. The control group rated videos through traditional methods, others with the proposed model. T test was used for comparison among groups. This study uses incremental cost effectiveness ratio to analyze the effectiveness of the system.
RESULTS: In the CPR-Coach dataset, our model achieved 88.79% Top-1 Acc and 99.40% Top-3 Acc, which suggests that it can effectively handle composite error recognition tasks. There were no differences in age, gender, educational background and clinical competence between the evaluation experts in groups. Results showed that there is no difference in accuracy between the control group (99.56 %) and the experimental group (99.72 %). But the average time consuming in experimental group (57 mins) was saved by nearly four times compared with the control group (15 mins) (P<0.05). The proposed system has preliminary capabilities to assist decision-making in CPR assessment.
CONCLUSIONS: The CPR composite error action recognition system based on AI, which can support fine-grained action recognition and composite error action recognition tasks under restricted supervision, can effectively assist doctors in CPR skill assessment, which can alleviate the time-consuming and labor-intensive issues of traditional assessment methods. In addition, it can effectively reduce labor costs and time costs without increasing the cost.