This paper looks for early signs of sudden cardiac arrest in long electrocardiogram (ECG) recordings. The authors used a motif-based method to summarise the shape of heartbeats over time and searched for when those shapes began to deviate from each person’s own baseline. In 23 Holter recordings, they found sustained, measurable changes that typically began hours before ventricular fibrillation (VF), the rhythm that often leads to sudden cardiac arrest.

The team analysed 23 two-channel Holter ECGs from the Sudden Cardiac Death Holter Database (SDDB) on PhysioNet. Recordings were sampled at 250 Hz and split into non-overlapping 10-second windows. For each window they picked a representative “motif” — a single heartbeat shape that best summarized that window — after detecting R-peaks and normalising each beat to a fixed length. Ectopic beats (extra or irregular beats) were kept so the trajectories would reflect the full range of changes.

They turned the sequence of window motifs into five simple metrics: personalised-baseline distance (how far current motifs are from the first 15 minutes), motif consistency (how similar beats in a window are to the motif), motif dispersion (how spread out beats are around the motif), motif heterogeneity (distance to the most unusual beat), and motif instability (how much the motif changes between windows). Each metric was converted to a z-score relative to the person’s first 15 minutes. A window was labelled “abnormal” when the z-score reached 3 or higher. Abnormal burden was defined as the proportion of abnormal windows inside a rolling 10-minute span; burden onset required the abnormal burden to persist for at least 2 minutes.

Across the cohort, the median time when sustained abnormal burden began fell between about 5.8 and 8.4 hours before VF, depending on the metric. Motif dispersion (the spread of beats around the motif) gave the most consistent long-range signal: sustained abnormal burden appeared at least 1 hour before VF in 100% of recordings and at least 2 hours before VF in 89% of recordings. Peak morphological deviation from baseline tended to occur around 2 hours before VF. Motif consistency showed the largest change in the final minutes before VF, with a median 70% abnormal windows in the last 10 minutes.