Aims: Early warning signatures of sudden cardiac arrest (SCA) remain poorly characterised in long-duration electrocardiograms (ECG). We aimed to quantify pre-event instability using a trajectory-based framework that models ECG as a continuous dynamical system rather than discrete events. Methods: Continuous Holter ECG from 23 patients with annotated SCA events was analysed using a sliding-window approach (10 s windows). ECG was represented using a motif-based paradigm, in which representative (motif) cardiac cycles provide a compact, interpretable, personalised description of morphology, enabling continuous tracking of deviation. Within each window, motif-based metrics were computed, including motif instability (window-wise distance), motif consistency (within-window distance), motif dispersion, and personalised-baseline distance. Trajectories were aligned to SCA onset and normalised using robust z-scores relative to an early baseline. Instability burden was defined as the proportion of windows exceeding z ≥ 3 within a 10-minute rolling window, from which peak instability and timing metrics were derived. Results: Motif-based trajectories exhibited multi-stage dynamics prior to SCA, with high early detection. Motif instability achieved ≥1-hour early warning in 95% of patients and ≥2 hours in 76%, with peak instability at a median of 112 minutes before SCA. Motif consistency showed the strongest late-stage signal, with 100% detected ≥30 min and 85% ≥120 min, and 70% abnormal burden in the final 10 min. Motif dispersion demonstrated the highest consistency (100% ≥60 min, 89% ≥120 min), with peak instability at a median of 127 minutes before SCA. Personalised-baseline distance provided stable early detection (94% ≥30 min, 75% ≥120 min). Conclusion: Trajectory analysis reveals progressive increases in morphological burden preceding SCA, not captured by conventional Holter metrics. Modelling ECG as a continuous dynamical system provides a sensitive, interpretable, personalised framework for early detection of cardiac deterioration. This approach requires no labelled data and is well-suited to wearable ECG platforms for continuous risk monitoring.