Prediction of countershock success using single features from multiple ventricular fibrillation frequency bands and feature combinations using neural networks.

Targeted defibrillation therapy is needed to optimise survival chances of ventricular fibrillation (VF) patients, but at present VF analysis strategies to optimise defibrillation timing have insufficient predictive power. From 197 patients with in-hospital and out-of-hospital cardiac arrest, 770 electrocardiogram (ECG) recordings of countershock attempts were analysed. Preshock VF ECG features in the time and frequency domain were tested retrospectively for outcome prediction. Using band pass filters, the ECG spectrum was split into various frequency bands of 2-26 Hz bandwidth in the range of 0-26 Hz. Neural networks were used for single feature combinations to optimise prediction of countershock success. Areas under curves (AUC) of receiver operating characteristics (ROC) were used to estimate prediction power of single and combined features. The highest ROC AUC of 0.863 was reached by the median slope in the interval 10-22 Hz resulting in a sensitivity of 95% and a specificity of 50%. The best specificity of 55% at the 95% sensitivity level was reached by power spectrum analysis (PSA) in the 6-26 Hz interval. Neural networks combining single predictive features were unable to increase outcome prediction. Using frequency band segmentation of human VF ECG, several single predictive features with high ROC AUC>0.840 were identified. Combining these single predictive features using neural networks did not further improve outcome prediction in human VF data. This may indicate that various simple VF features, such as median slope already reach the maximum prediction power extractable from VF ECG.