Non-invasive classification of coronary perfusion pressure during CPR using smartphone-based skin video and deep learning.

BACKGROUND AND

OBJECTIVE

Coronary perfusion pressure (CPP) is an important determinant of myocardial blood flow and an indicator during cardiopulmonary resuscitation (CPR). However, conventional CPP monitoring methods are invasive and unsuitable for out-of-hospital settings. This study proposes a non-invasive approach to classify CPP levels using skin video recorded with a smartphone camera and deep learning.

METHODS

Video and biosignal data were collected simultaneously from 15 pigs during CPR. An integrated deep learning model that combines backbone model (CNN, EfficientNetV2-B0, ConvNeXt-Nano, FastViT-T8) with gated recurrent unit (GRU) was developed to classify whether CPP exceeded the clinically relevant threshold of 20 mmHg. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to identify regions attended to by the model, and statistical analysis conducted.

RESULTS

Among four different backbone models used for training and evaluation, the EfficientNetV2-B0-GRU architecture demonstrated the best performance, achieving accuracy of 84.60 % and F1-score of 76.04 %. Statistical analysis revealed significant differences in YCrCb channel values between correctly classified groups. In addition, active regions of Grad-CAM showed greater variances than inactive regions, indicating that the model focused on regions with higher color variabilities related to perfusion.

CONCLUSION

A deep learning-based approach using skin video enables the non-invasive classification of CPP during CPR. Grad-CAM and quantitative YCrCb analyses improve interpretability, while the proposed method provides proof-of-concept for cost-effective and accessible perfusion-related assessment using skin video recorded with a smartphone, with potential future utility for real-time CPR quality assessment and decision-making in emergency or out-of-hospital settings.