Abstract
Survival prediction based on an electrocardiogram (ECG) is crucial for assessing cardiovascular disease risk. Existing deep survival models mostly rely on convolutional neural networks, which effectively extract local morphological features but struggle to capture dynamic evolution patterns across different time scales in ECGs. To address this issue, this paper proposes a multi-scale spatio-temporal fusion network (MSTF-Net), which jointly models local waveform morphology, mid-term rhythm changes, and long-term temporal dependencies through a hierarchical structure and directly outputs individualized survival probabilities within the framework of discrete-time survival analysis. Experiments on the CODE dataset (233,645 ECGs) and the independent SaMi-Trop cohort (1,631 ECGs) demonstrate that the proposed method achieves a C-index of 0.7828 on the internal test set and 0.7869 on the external validation set, with an integrated Brier score (IBS) of 0.02841 and 0.02996, respectively. Compared to the state-of-the-art ResNet baseline, MSTF-Net significantly improves calibration performance (IBS: 0.02841 vs. 0.02870, p< 0.01) and exhibits superior cross-cohort generalization, achieving a generalization ratio close to 1.0. Ablation studies further validate that each multi-scale temporal component contributes indispensably to the model's generalization capability.