Cardiovascular disease (CVD) remains a leading cause of global mortality, requiring accurate and early diagnosis. This study proposes a ConvNeXt-based multi-module feature fusion approach that enhances feature extraction, interpretability, and spatial-temporal representation in ECG classification. Using the MIT-BIH dataset, ECG signals are denoised with Discrete Wavelet Transform (DWT) and balanced using SMOTE, then encoded into GASF, GADF, and MTF images. These images are fused to improve generalization within the 2D module. The first module uses a ConvNeXt architecture with CBAM for feature extraction from the fused 2D representations, while the second combines CNN, SENet, and BiLSTM to analyze 1D ECG signals. This fusion maximizes the benefits of both data types, leading to a robust cardiac abnormality detection. The approach achieves training and validation accuracies of 99.97% and 99.90%, demonstrating its potential for practical cardiovascular diagnostics.