Electrocardiogram (ECG) is one of the most important tools that are used to diagnose heart arrhythmias in the medical field. The primary purpose of the document is to autonomously detect ECG arrhythmias by recognizing the cardiac conditions indicated by the patient’s ECG readings. For this purpose, a deep learning framework that has learned from a general picture data set in advance will be utilized. This basic back propagation neural network is then used on the recovered features to do the last classification. Three distinct ECG (Electrocardiogram) waveform scenarios are chosen from the MIT-BIH dataset. This study’s primary goal is to apply a deep learning approach that is straightforward, dependable, and simply adaptable for classifying the three distinct heart diseases that were chosen. The obtained findings showed that very high-performance rates might be achieved by cascading a standard back propagation neural network with a transferred deep learning feature extractor. The highest accurate recognition rate recorded was 98.51%, with a testing accuracy of about 92%. In accordance with these findings, transfer learning emerged as an effective automated strategy for detecting cardiac arrhythmias. Use of transfer learning removed the need to construct a deep CNN model from the beginning, making it a readily implementable method.