Abstract

The heartbeat-evoked potential (HEP), a cortical response to heartbeats and a neural marker of interoception, is increasingly considered clinically relevant, but is heavily contaminated by cardiac artefact (CA) on the scalp, making reliable distinction of HEP from CA challenging. Because the HEP and cardiac potentials are anatomically distinct, they may be separable via beamforming, a source localisation method that isolates brain activity at specific locations while suppressing external noise and interference. Here, the first known ground-truth validation of EEG beamforming for HEP source reconstruction was attempted, aiming to quantify source waveform recovery and spatial localisation accuracy using simulated EEG data. Using linearly constrained minimal variance (LCMV) beamforming, the following was investigated. (A) To test whether beamforming can recover a known signal, 128-channel EEG datasets were simulated for 3 models with a known HEP waveform: a single right insula (R-Ins) HEP (1), two temporally distinct HEPs in the R-Ins and right anterior cingulate cortex (R-ACC) (2), and two temporally overlapping HEPs in the same regions (3). Recovery was investigated by correlating the virtual electrode waveforms at the true location with the known true input waveform. (B) To test CA suppression, CA extracted from isoelectric EEG of brain-dead individuals providing CA with limited cortical activity, was integrated into the simulated EEG data. Source (-10 to -50dB) and sensor (0 to -30dB) signal-to-noise ratios (SNR) were systematically varied for each model with and without CA. (C) LCMV beamforming was then applied to retrospective empirical EEG data from hypertensive and anxiety individuals (n=106). Across simulations and empirical data, T-tests compared power in a HEP-dominated window to an earlier CA-dominated window. Null-space projection, using subject-specific QRS waveform and its temporal derivative, was applied to remove residual CA in reconstructed source waveforms. In model 1, beamforming achieved near-perfect recovery without CA (r>0.99, 0mm error) at source SNR of -30dB, remaining robust in the presence of CA (r=0.72-0.94, 0-5.7mm error). Recovery degraded at low SNR (<-20 dB; r<0.3, up to 27mm error). Models 2 and 3 showed similar patterns but introduced R-ACC to R-Ins leakage. Applied to empirical data, beamforming revealed significant HEP activity in the R-Ins and R-ACC across all pooled data in source space (p < 0.001). LCMV beamforming also revealed a significant HEP difference in the late R-ACC window (250-500 ms post R-peak) in hypertension versus controls (p = 0.040, d = 0.92) when poor SNR subjects were excluded. This effect strengthened after QRS cleaning (p = 0.035, d = 0.96). LCMV beamforming can reliably recover the simulated HEP while suppressing CA, provided source SNR is sufficiently high. Applied to empirical data, LCMV beamforming recovered HEP activity from interoceptive regions (R-Ins and R-ACC) and was sufficiently sensitive to detect clinically meaningful group differences. This study offers a source level approach to separate HEP activity from CA, a distinction that sensor-level analysis can struggle to make. Together, LCMV beamforming and scalp-based methods can provide converging evidence for genuine HEP activity.