Atrial fibrillation (AF) is the most common and prevalent cardiac arrhythmia that leads to cardiovascular mortality, blood clots, stroke, heart failure, other heart-related complications, and death. Manual screening of AF from an electrocardiogram (ECG) record is time-consuming and labor-intensive and leads to misdiagnosis due to medical professional skill gaps and sight problems. There is also a high possibility that small changes in the ECG signal may be left undetected by the human eye due to a weak heart signal. Therefore, automatic early-stage detection of AF is crucial to addressing the health problem. Machine learning models have been formulated to detect AF automatically, but they need more performance analysis to implement. In this work, we analyze the performance of deep learning algorithms like convolutional neural networks (CNN) and long shortterm memory (LSTM) to detect AF at an early stage from an ECG signal. To analyze the two deep learning algorithms for AF detection from an ECG signal, the thesis is organized into three main stages. ECG data collection and reshaping were done in stage one. In the second stage, ECG is preprocessed (ECG normalization and denoising using Butterworth low pass filter, median filter, and notch filter), and an AF detection model is created using deep learning algorithms (CNN and LSTM) and tested using accuracy performance metrics. The experimental results of the proposed AF detection method showed that the CNN algorithm had better performance with 98% accuracy using balanced ECG data. However, the LSTM algorithm performs poorly, with only 77% accuracy when compared to the CNN algorithm. As a result, CNN's algorithm can help medical professionals detect or diagnose diseases. Keywords: AF, CNN, ECG, Deep Learning, LSTM