The electrocardiogram (ECG) is an important measurement for diagnosing heart disease. Transmission of continuous ECG over a wireless network can be taxing; therefore, compressing the ECG can reduce the load on wireless networks. On the other hand, reconstructing the ECG for analysis can be computationally intensive. As such, diagnosing heart diseases from compressed ECG is desired. Abnormal beat detection using machine learning in the compressed domain is proposed. The ECG was compressed using a wavelet-based morphological feature preserving compression algorithm The compression algorithm was applied on 84 ECG records available in Long Term Atrial Fibrillation Database (LTAFDB) achieving an average compression rate of 4.17:1. Abnormal beats in the compressed signals were classified using a Random Forest trained using a randomly under-sampled training set. The achieved true positive rate was 69.5% and the false positive rate was 32.8%. The results indicate that identification of abnormal beats in compressed ECGs is possible. Future work will explore detection of abnormal beats in compressively sensed ECG.