Sensitivity Analysis of ECG Features to Computational Model Input Parameters

Jenny Venton1, Karli Gillette2, Matthias Gsell3, Axel Loewe4, Claudia Nagel5, Benjamin Winkler6, Louise Wright1
1Data Science, National Physical Laboratory, 2Gottfried Schatz Research Center - Medical University of Graz, 3Medical University of Graz, 4Karlsruhe Institute of Technology (KIT), 5Institute of Biomedical Engineering - Karlsruhe Institute of Technology (KIT), 6Physikalisch-Technische Bundesanstalt (PTB)


Abstract

Aim: Cardiac models of electrophysiology generating simulated electrocardiogram (ECG) signals are an increasingly valuable tool for both personalised medicine and understanding cardiac pathologies. Knowledge of how simulation parameters affect clinical features of the simulated ECG is crucial. This study used sensitivity analysis (SA) methods to determine the impact of cardiac model input parameters on measured ECG R peak amplitudes.

Method: A ventricular model was used to generate QRS complexes. Twelve input parameters relating to stimulation sites were varied according to a Saltelli sampling scheme, resulting in 14000 simulated signals. Each QRS complex was appended to a generic P-wave, generated using a non-corresponding atrial model, creating a complete heartbeat as required by ECG feature extraction software. R amplitude was calculated using two methods: ECG feature extraction software (ECGdeli), and finding the absolute maximum of the signal. First order Sobol coefficients were calculated using two SA methods: direct numerical evaluation of integrals and polynomial chaos expansion (PCE).

Results: Sobol coefficients calculated using PCE and direct methods were in good agreement (correlation coefficient 0.97). The agreement between the ECGdeli and maximum methods for determining R amplitude varied depending on which ECG lead was considered. Correlation coefficients ranged from 0.70 (lead aVR) to 1.00 (leads aVL, V4, V5, V6).

Conclusion: Sensitivity analysis provides valuable information about the relationship between simulated ECG morphology and cardiac model input parameters. This provides valuable insight on the quality of the simulated signals and can allow for more nuanced patient-specific simulation changes.