Towards real-time heart health monitoring in firefighting using convolutional neural networks

Abstract

A machine learning-based heart health monitoring model, named H2M, was developed. Twenty-four-hour electrocardiogram (ECG) data from 112 career firefighters were used to train the proposed model. The model used carefully designed multi-layer convolution neural networks with maximum pooling, dropout, and global maximum pooling to effectively learn the indicative ECG characteristics. H2M was benchmarked against three existing state-of-the-art machine learning models. Results showed the proposed model was robust and had an overall accuracy of approximately 94.3%. A parametric study was conducted to demonstrate the effectiveness of key model components. An additional data study was also carried out, and it was shown that using non-firefighters’ ECG data to train the H2M model led to a substantial error of ∼40%. The contribution of this work is to provide firefighters on-demand, real-time status of heart health status to enhance their situational awareness and safety. This can help reduce firefighters’ injuries and deaths caused by sudden cardiac events.