A Neural-Network Model for Deriving Analogous Three-vector ECGs of S-ICD from Standard Nine-lead ECGs

With clinically reported problems of current subcutaneous implantable cardiac defibrillator (S-ICD), such as over-sense of ECG signals and correspondingly inappropriate shock, popularization of the surgical implantation process looks eagerly forward to a pre-operative ECG screening for doctor to determine whether a patient is suitable for an implantation or not. In this work, a nonlinear BP (back propagation) neural network model with three layers was proposed for deriving analogous three-vector ECGs from surface standard nine-lead ECGs of a patient, which can be recorded easily and non-invasively before the implantation. To evaluate its reconstruction performance, we trained and tested this model in 21 patients and 4 health subjects from two public standard 12-lead ECG databases. Over 320 thirty-seconds ECG segments, three analogous ECGs including the Primary, Secondary, and Alternate vectors (corresponding to standard leads I, II, and V2) can be derived well from the remaining nine-lead ECGs, and the obtained average R/T ratio between of the original and derived ECGs has a mean p-value 0.22 > 0.05. In addition, the correlation coefficients and the root mean square error are of around 0.82-0.93 and 55.9-92.3 μV, respectively. The obtained results indicate a great correlation between of the original and derived ECG signals. As such, this work would be of great significance for the future study on the clinical applicability of a pre-operative screening tool with the proposed method of this pilot work.