Automated arrhythmia detection from electrocardiogram (ECG) signals is crucial for the early prevention and diagnosis of cardiovascular diseases. However, conventional ECG classification methods usually employ 1D convolutions to extract features from individual leads separately, neglecting the intrinsic correlations inherent in 12-lead ECGs. This paper proposes DSTNN-X, a deep neural network that performs spatio-temporal learning by effectively aggregating multi-lead contextual information for ECG arrhythmia classification. Specifically, a depth-stationary cross convolution (DSC-Conv) is introduced to synergistically integrate horizontal convolutions with vertical convolutions, effectively extracting temporal patterns and spatial dependencies of intra-group leads. Besides, a residual structure is designed, combining temporal convolution with cross-lead features to enhance inter-group lead correlations. Finally, multi-head attention and a linear layer are employed to map arrhythmia diagnosis labels. Experimental results on public multi-label ECG datasets PTB-XL and CPSC2018 demonstrate that the proposed lead-fusion network achieves consistent performance improvements across diverse arrhythmia classification tasks, outperforming current state-of-the-art models.