Dynamic prediction of malignant ventricular arrhythmias using neural networks in patients with an implantable cardioverter-defibrillator

Multimodal explainable artificial intelligence identifies patients with non-ischaemic cardiomyopathy at risk of lethal ventricular arrhythmias

“…Raw-format 12-lead 10-s resting ECGs were collected retrospectively at both sites ( Supplementary Methods p1). After downsampling to 250 Hz, raw signals underwent noise filtering and baseline wander removal using a Savitzky-Golay Filter for smoothing via high-order polynomial fitting and a low-resolution Fourier series subtraction for eliminating baseline wander 16 , 21 . Individual heartbeats were isolated for each ECG lead by automatic marking of individual R-peak locations, and subsequent extraction of heartbeat templates given a list of these locations (Fig.…”

“…Individual heartbeats were isolated for each ECG lead by automatic marking of individual R-peak locations, and subsequent extraction of heartbeat templates given a list of these locations (Fig. 1 D) 16 . Individual P-QRS-T segments were aligned, after which mean waveforms were calculated by averaging individual waveforms per unique lead.…”

“…Improved arrhythmic risk prediction may be achieved through comprehensive modelling of electrophysiological mechanisms and myocardial substrate characteristics using neural networks, that can learn non-linearities and complex patterns from high-dimensional data 12 , 13 . In relatively homogeneous cohorts of patients with an ischaemic cardiomyopathy, neural networks have been shown to capture high-level representations from cardiac magnetic resonance imaging (CMR) and electrophysiological signals, that reflect ventricular arrhythmia risk 14 – 16 . This approach could prove beneficial in non-ischaemic cardiomyopathy, which encompasses heterogeneous inflammatory, infiltrative and genetic pathologies, each of which likely has a distinct SCD risk profile 9 .…”

Digitale Zwillinge in der kardialen Elektrophysiologie – aktueller Entwicklungsstand und kommende Herausforderungen

“…The significant increase in computational resources and the staggering expansion of the amount of available data have enabled the development of advanced data-driven models [ 11 , 15 ]. These models include classical risk scores derived from multivariable statistics, e.g., CHA 2 DS 2 -VASc (congestive heart failure, hypertension, age ≥75 [doubled], diabetes, stroke/thromboembolism [doubled], vascular disease, age 65 to 74, and sex category [female]), as well as artificial intelligence- or machine learning-based models [ 19 ]. These are able to map patient-specific input parameters to clinically relevant output parameters, but are typically static and do not explicitly integrate pathophysiological mechanisms, reducing their ability to generalize beyond the training space and limiting their explainability.…”

“…Finally, even if high-quality patient-specific data can be obtained for model personalization, these data are typically only available at a single moment in time, ignoring the dynamic nature of cardiac electrophysiology and the sudden occurrence of arrhythmias. Recent work has shown how the performance of a static machine learning-based VT risk predictor derived from the baseline ECG dropped over time, whereas a dynamic model incorporating time-varying ECG data showed increased performance over time [ 19 ]. More research is needed to identify methodologies to incorporate dynamic changes in arrhythmogenic risk in digital twins, e.g., based on blood biomarkers or wearables.…”