Deep behavioural representation learning reveals risk profiles for malignant ventricular arrhythmias

We aimed to identify and characterise behavioural profiles in patients at high risk of SCD, by using deep representation learning of day-to-day behavioural recordings. We present a pipeline that employed unsupervised clustering on low-dimensional representations of behavioural time-series data learned by a convolutional residual variational neural network (ResNet-VAE). Data from the prospective, observational SafeHeart study conducted at two large tertiary university centers in the Netherlands and Denmark were used. Patients received an implantable cardioverter-defibrillator (ICD) between May 2021 and September 2022 and wore wearable devices using accelerometer technology during 180 consecutive days. A total of 272 patients (mean age of 63.1 ± 10.2 years, 81% male) were eligible with a total sampling of 37,478 days of behavioural data (138 ± 47 days per patient). Deep representation learning identified five distinct behavioural profiles: Cluster A (n = 46) had very low physical activity levels and a disturbed sleep pattern. Cluster B (n = 70) had high activity levels, mainly at light-to-moderate intensity. Cluster C (n = 63) exhibited a high-intensity activity profile. Cluster D (n = 51) showed above-average sleep efficiency. Cluster E (n = 42) had frequent waking episodes and poor sleep. Annual risks of malignant ventricular arrhythmias ranged from 30.4% in Cluster A to 9.8% and 9.5% for Clusters D-E, respectively. Compared to low-risk profiles (D-E), Cluster A demonstrated a three-to-four fold increased risk of malignant ventricular arrhythmias adjusted for clinical covariates (adjusted HR 3.63, 95% CI 1.54-8.53, p < 0.001). These behavioural profiles may guide more personalised approaches to ventricular arrhythmia and SCD prevention.