Introduction: Monitoring the heart rhythm during out-of-hospital cardiac arrest (OHCA) is essential to improve treatment quality. OHCA rhythms fall into five categories: asystole (AS), pulseless electrical activity (PEA), pulse-generating rhythms (PR), ventricular fibrillation (VF) and ventricular tachycardia (VT). OHCA patients are monitored using defibrillators that record the ECG and the thorax impedance (TI) concurrently through the defibrillation pads. The aim of this study was to develop an algorithm to classify OHCA rhythms using the signals recorded by the defibrillation pads.
Materials and methods: Defibrillator data from 100 OHCA patients from three European locations (42 Akershus, Norway; 30 Stockholm, Sweden; 28 London, UK) were analysed and 2833 4-second segments were extracted: 423 AS, 912 PE, 689 PR, 643 VF, and 166 VT. The stationary wavelet transform was used to obtain 95 features from the ECG and the TI. Random Forest (RF) classifiers were used, features were ranked during training using RF-importance, and models with increasing number of features were evaluated. Data were partitioned patient-wise in a quasi-stratified way into training (70%) and test (30%) sets. The process was repeated 100 times to statistically characterise the results. The models were evaluated using the per-class sensitivity (Se), and the unweighted mean of sensitivities (UMS) as global performance metric.
Results: The best classifier was obtained combining 20 ECG and TI features, achieving a median (90% confidence interval) UMS value of 87.4% (82.5-91.0), 0.7 percentage points above that of the best model using only the ECG (needing 30 features). The Se for AS/PEA/PR/VF/VT were 97.1% (94.1-99.3), 79.1% (64.6-89.9), 87.2% (74.7-94.5), 93.3% (87.9-97.7) and 82.5% (61.3-93.5), respectively.
Conclusion: A robust and accurate approach for multiclass OHCA rhythm classification has been presented. Adding the TI improves the accuracy of current rhythm classifiers based exclusively on the ECG.