Comprehensive Summary
In this study, researchers developed an algorithm consisting of a convolutional neural network as a feature extractor (CNNE) and Boosting (BS) classifier to improve AED detection of sudden cardiac arrest (SAC). 57 ECG records from two separate databases were divided into a training (70%) or evaluation (30%) group. The databases were segmented and then processed with the modified variational mode decomposition (MVMD) technique. The MVMD technique separated each ECG segment into three input channels: preprocessed ECGs, shockable (SH) signals, and non shockable (NSH) signals. Then, using a grid search with five-folds cross validation, researchers optimized CNN learning parameters, and a final CNN algorithm was selected. The final proposed model was validated with a BS classifier. Overall, the model achieved an accuracy of 99.26%, a sensitivity of 97.07%, and a specificity of 99.44%.
Outcomes and Implications
Specificity and sensitivity are two vital figures used to assess for AED compliance. A false shock from an AED could result in an artificial SAC, while the absence of a shock when needed could delay life-saving defibrillation. The use of convolutional neural networks to identify shockable rhythms can prevent harmful clinical outcomes for both those who should not be shocked and those who should. Additionally, while past CNN models have not met AHA accuracy guidelines for AED compliance, the CNN proposed in this study does, implicating its efficiency over other models. However, the limitations of this study, including its very small dataset size, should be taken into account, and further validations on larger populations should be performed in the future.