Aims: The overall goal is to develop methods to understand the relationship between therapy and patient response during CardioPulmonary Resuscitation (CPR). Among the subsidary goals are chest compression and ventilation detection, shock outcome prediction, compression artifact removal and rhythm interpretation. Patients undergoing CPR will respond through rhythm transitions between ventricular fibrillation (VF), ventricular tachycardia (VT), asystole (AS), pulseless electrical activity (PEA) and pulse generating rhythm (PR). In this work we add ress rhythm interpretation with the aim to recognise the five rhythm types applying deep neural networks on electrocardiogram (ECG) segments recorded during resuscitation.
Methods: The database was extracted from patients of an out-of-hospital cardiac arrest study. Artifact-free four- second segments were extracted from 100 patients. A total of 2833 segments were included: of 643 VF, 166 VT, 423 AS, 912 PE, and 689 PR. In this study, a convolutional neural network (CNN) was used, experiments conducted with increasing number of layers, with and without padding, varying filter kernel and pooling strategies. For the best model, training was repeated 10 times to explore variations in the results. The data was split into training, test, and validation sets.
Results: For the best performing five layer network the following metrics (median(25th,75th quartiles) were achieved on the test set: Total accuracy 0.81(0.77,0.84); Rhythm specific recalls AS 1.00(1.00,1.00), PEA 0.57(0.52,0.62), PR 0.68(0.58,0.78), VF 0.87(0.82,0.98), VT 0.91(0.87,0.95); Rhythm specific precisions AS 0.88(0.71,0.88), PEA 0.80(0.78,0.90), PR 0.96(0.91,1.00), VF 0.77(0.73,0.81), VT 0.75(0.70,0.81).
Conclusion: We have proposed a deep learning approach to automatically recognise five cardiac arrest rhythms common during resuscitation. The results are promising and comparable to our previous studies using classical machine learning approaches with feature engineering.