A personalized pipeline to reduce ECGi-AF biomarkers disparity: a clinical-computational study

Gabriel Costa1, Camila Restivo1, Italo Sandoval Ramos de Oliveira2, Maria de la Salud Guillem Sánchez3, Joao Salinet1
1HEartLab, Federal University of ABC, 2Universidade Federal do ABC, 3Universitat Politècnica de València


Aims: This study aimed to proposed a pipeline of techniques to reduce disparity in ECGI-AF frequency and phase biomarkers, due to signals noise, reduced number of leads layout and 3D geometries modification.

Method: 771 torso signals from each AF mathematical models (n =5) are contaminated with different levels of noise (SNR between – 3dB to 60 dB), and filtered with a Butterworth bandpass filter (BPF) at different cut-off frequencies (0.5-30, 3-30 and HDF±1 Hz). Through the Tikhonov Zero Order and 3D torso/atria geometries, 2048 epicardial atrial electrograms (AEGs) are estimated, and 13 biomarkers calculated from the respective 3D frequency (n=7) and phase (n=6) maps. The impact of reduced number of leads (from 771 to 131, 62, 32, and 16) and 3D torso/atria morphologies change, according to the respiratory and cardiac cycles, on the biomarkers are also investigated. The above methodology was extended to two AF patients (EDGAR database). Torso and atria geometries were segmented using SEG3D, and meshes created on Cleaver.

Results: The cut-off frequency around the HDF reduced the disparity of frequency and phase biomarkers, for the different SNRs. Moreover, in overall, the different morphological 3Ds (torso and/or atria) did introduce small variations on the biomarkers, showing the importance of considering the cardiac and respiratory cycles on ECGi. Reduction of leads increased disparity within biomarkers, but the impact is lower using Laplacian interpolation rather than Linear interpolation. Analysis from patients showed a higher sensibility of phase biomarkers due to the choice of BPF cut-off frequencies.

Conclusion: ECGi in AF may be customized with (i) a narrowed BPF around HDF on the torso’s signals, (ii) applying Laplacian interpolation to estimate unknown sources and, (iii) using the 3D torso/atria geometries obtained according to the cardiorespiratory cycle.