<![CDATA[Cardiac arrhythmia is a critical cardiovascular disorder associated with high global mortality, while electrocardiogram (ECG)-based diagnosis remains time-consuming and susceptible to inter-observer variability. Although recent artificial intelligence approaches have improved automated ECG analysis, many existing studies rely on single time-frequency representations and uniform classification strategies, limiting their ability to capture complementary ECG characteristics. This study proposed a hybrid time-frequency ECG arrhythmia classification framework incorporating feature–model compatibility and ensemble decision fusion to improve classification robustness and reliability. ECG signals from the MIT-BIH Arrhythmia Database were preprocessed using a fourth-order Butterworth bandpass filter and segmented using overlapping windows. To prevent data leakage, patient-wise cross-validation was employed, ensuring that ECG segments originating from the same patient were assigned exclusively to either training or testing folds. Three complementary time-frequency representations, namely Mel-Spectrogram, Short-Time Fourier Transform (STFT), and Discrete Wavelet Transform (DWT), were paired with their most compatible classifiers: Convolutional Neural Network (CNN); Random Forest (RF); and Support Vector Machine (SVM); respectively. Decision fusion was performed using stacking, hard voting, and soft voting strategies. Experimental results showed that the DWT-SVM model with stacking achieved the best overall performance, attaining 94.82% accuracy and a 94.59% F1-score, while STFT-RF with hard voting achieved comparable performance with 94.75% accuracy and a 94.50% F1-score. In contrast, Mel-Spectrogram-CNN produced substantially lower performance with 81.45% accuracy, indicating limited suitability of Mel-scale representations for ECG morphology analysis. Statistical analysis confirmed significant performance differences among models (p < 0.001). The findings demonstrate that integrating hybrid time-frequency representations with feature–model compatibility and model-dependent decision fusion provides a robust framework for automated ECG arrhythmia classification with strong potential for clinical decision support and real-time cardiac monitoring applications]]>
