This paper presents an efficient detection and classification algorithm of multiple-class myocardial infarction (MI) (i.e., prior and acute), which is one of mortality diseases for humans. However, feature extraction is one of the challenges in MI classification as the extracted features may not be optimized for class separation. To this end, we propose a new deep feature learning based MI detection and classification approach. It seeks to learn a representation of the extracted features that optimize the classification performance. Moreover, to further enhance the feature learning process, we incorporate multi-scale discrete wavelet transformation into the feature learning process to facilitate the extraction of MI features at specific frequency resolutions/scales. Finally, softmax regression is employed to build a multi-class classifier based on the learned optimal representation of the features. Experimental results using public ECG datasets obtained from the PTB diagnostic database show that the proposed approach can achieve better performance than other state-of-the-art approaches in terms of sensitivity and specificity. The effectiveness and good performance of the proposed approach may serve as an attractive alternative to MI classification or other related applications.