Pulse detection in Out-of-Hospital Cardiac Arrest (OHCA) is crucial to identify cardiac arrest and to detect Return of Spontaneous Circulation (ROSC). Carotid pulse check or checking for signs of life have been proven time consuming and inaccurate. Pulse detection remains challenging for both lay rescuers and healthcare personnel, and automatic methods are needed for universal use in automated external defibrillators, where the only signal always available is the ECG.
Aim: To develop an automatic algorithm for organized rhythms to be discriminated into Pulsed-Rhythm (PR), associated with ROSC, and Pulseless Electrical Activity (PEA) using only the ECG.
Materials and methods: A total of 3914 ECG segments (2372 PR and 1542 PEA) of 4 s were extracted from 279 OHCA patients (134 with ROSC and 145 without ROSC). Segments were annotated as PR or PEA based on clinical ROSC marks from clinicians on scene, and corroborated with the thoracic impedance and the capnogram. The ECG was downsampled to 100 Hz and bandpass filtered 0.5-30 Hz. The designed algorithm extracted ECG features using a single convolutional layer followed by a pooling layer and a bidirectional gate recurrent unit. In the classification stage, a single neuron computed the likelihood of PR. Data were partitioned patient-wise into 3038 segments (223 patients) for training and 876 (56 patients) for testing. Performance was evaluated in terms of correctly detected PR/PEA, Sensitivity(Se)/ Specificity(Sp) respectively, weighted per patient.
Results: The method showed a Se/Sp of 89.4%/97.2% in the training set and 91.7%/92.5% in the test set.
Conclusions: A simple deep neural network can accurately discriminate between PR and PEA rhythms. More advanced regularization techniques or the inclusion of more data in the training of the model would reduce overfitting and hence improve the results.