Feasibility of Patient-Specific Magnetohydrodynamic Pattern Recognition for Correcting and Simulating 12-Lead ECGs During Cardiac MRI

Background: During cardiac MRI (CMR), a reliable and clinically interpretable 12-lead electrocardiographic (ECG) recording is essential for monitoring of patients with cardiac implantable electronic devices (CIEDs) or during MRI-guided cardiac procedures. However, ECG recordings are severely distorted by the magnetic field, time-varying gradients, and radiofrequency pulses. In particular, the magnetohydrodynamic (MHD) effects, predominantly resulting from the pulsatile aortic electrically conductive blood flow within the magnetic field, produce ECG distortions that potentially obscure important ECG-features, like acute changes in the ST-segment and T-wave. Currently available 12-lead in-bore ECG systems insufficiently correct for these MHD-related artifacts, thereby limiting their clinical applicability. This study aimed to isolate the MHD effect in 12-lead in-bore ECGs and to explore its feasibility for MHD correction and MHD simulation.

Methods: Continuous 12-lead ECGs were recorded in patients w/o CIEDs undergoing a routine clinical CMR scan (1.5T, MAGNETOM Sola, Siemens Healthineers, Forchheim, Germany), using an MRI-compatible ECG recording system (MiRLTE Medical, North Andover, MA, USA). Clean ECGs were acquired outside the MRI scanner and MHD-affected ECGs inside the bore in the absence of active imaging for 60 seconds (Figure 1). Beat-to-beat MHD patterns were isolated by subtracting clean ECG beats from MHD-affected beats to derive a patient-specific average MHD pattern per lead. This patient-specific MHD pattern was subtracted from all MHD-affected beats to remove the MHD-induced distortions and added to clean ECG beats to simulate MHD-affected signals. Maximum absolute cross-correlation quantified the temporal similarity of full beats.

Results: A total of 20 patients were included, with a mean of 49.2±15.8 heartbeats per patient analyzed (Table 1). High variability in MHD pattern was observed per beat and per patient (Figure 2). The average cross-correlation between clean and MHD-affected beats was 0.63±0.04 for all leads, which was improved by the MHD-removal method to 0.82±0.07, with precordial leads showing the greatest increase (Table 1). Synthetic MHD beats showed a mean cross-correlation of 0.92±0.02 with the patient’s measured MHD-affected beats during CMR.

Conclusion: This study shows that using a patient-specific MHD-pattern removal technique, the MHD effects in 12-lead ECG recordings can consistently be removed and substantially improve correspondence between ECGs obtained within and outside the MRI bore.

Nevertheless, interpatient variability and beat-to-beat changes in heart rate or stroke volume (i.e., aortic flow) limit robustness, as substantial hemodynamic fluctuations directly alter MHD patterns, complicating MHD correction. While deep learning methods such as denoising autoencoders could learn this MHD variability, paired datasets of 12-lead MHD-affected and clean ECGs are required to improve ECG quality, which the proposed MHD simulation method enables.

Table 1: Clinical characteristics of the study population and cross-correlation of heartbeats. Values for patient characteristics are presented as mean ± standard deviation (SD), or number (percentage). Maximum normalized cross-correlation coefficients are reported for clean ECG beats versus MHD-affected beats, clean ECG beats versus corrected beats obtained with the proposed method, and MHD-affected beats versus simulated MHD beats generated using this approach. Values are expressed as mean ± SD, derived from 20 patients with a total of 984 heartbeats.

Patient characteristics
Age – yr	63.2±9.0
BMI – kg/m2	25.7±4.3
Male sex – no. (%)	12 (60%)
Device present – no. (%)	9 (45%)
Paced during CMR – no. (%)	6 (30%)
Heart rate – bpm	70.9±13.2
LVSV – mL	88.76±27.0
LVEF – %	47.9±9.2
Cross-correlations
	Clean vs MHD	Clean vs Corrected	MHD vs Simulated
Limb leads	0.60±0.02	0.77±0.04	0.91±0.01
Precordial leads	0.66±0.04	0.88±0.03	0.93±0.01
Total	0.63±0.04	0.82 ± 0.07	0.92 ± 0.02

Figure 1: Flowchart of the methodology. ECG and MRI data were recorded simultaneously during a routine diagnostic scan. MHD pattern extraction involved three main steps: 1) R-peak detection using independent component analysis, 2) beat extraction using a fixed beat size based on the mean of the median RR intervals of both clean and MHD-affected signals with false beats manually removed when present, and 3) beat-per-beat subtraction of the MHD-affected and clean beats. For MHD correction, the mean MHD pattern was subtracted from the MHD-affected beats. For simulation of MHD-affected beats, the mean MHD pattern was added to the clean beats.
Figure 2: Example of MHD correction and simulation. A) Effective MHD correction with close correspondence to the clean ECG. B) Reduced MHD correction quality for another beat of the same patient. C) MHD simulation closely matches the measured artifact. D) Less accurate MHD simulation in another beat from the same patient. Due to beat-to-beat MHD pattern differences lower cross-correlations are for some beats observed (B and D).