Cardiovascular diseases (CVDs) are one of the leading causes of death in the world. Long-term ambulatory electrocardiogram (ECG) monitoring enables timely medical interventions. Effective and efficient ambulatory ECG monitoring can be supported by compressive sensing, which can reduce battery consumption; however, the reconstruction of compressively sensed ECG could be a computationally intensive operation. Therefore, detection of CVDs in compressively sensed ECG is highly desirable. This thesis proposes a three-stage system based on machine learning for detecting atrial fibrillation (AF) in compressively sensed ECG while reducing false alarms (i.e., false positives) due to contamination and reducing reconstruction of the ECG. Stage 1 uses a novel signal quality index (SQI) to represent the quality of the compressively sensed ECG on a continuous scale and reject low-quality ECG. Stage 2 classifies the ECG as either AF or normal and generates an associated confidence score. Stage 3 selectively reconstructs the ECG when the Stage 2 confidence score is unacceptable and classifies the reconstructed ECG as either AF or normal. Clean ECG from Long-Term AF Database was artificially corrupted with simulated motion artifact to pre-set signal-to-noise ratios. The corrupted ECG was compressively sensed at 50% and 75% compression ratios (CRs). The system was evaluated using average precision (AP), the area under the curve (AUC) of the receiver operator characteristic curve, false-positive rate (FPR), and true positive rate (TPR). The system was optimized to maximize the AP and minimize ECG rejection and reconstruction ratios. The optimized system for 50% CR had 0.72 AP, 0.63 AUC, 0.38 rejection ratio, and 0.38 reconstruction ratio. The optimized system for 75% CR had 0.72 AP, 0.63 AUC, 0.40 rejection ratio, and 0.35 reconstruction ratio. The optimized system reduced FPR by 0.37 for both CRs compared to Stage 2 alone. At a fixed FPR of 0.10, the TPR of the proposed system was 0.43 and 0.42 for 50% and 75% CRs, respectively, an improvement of 0.10 and 0.14 over Stage 2. The proposed system reduced the probability of false alarms (FPR) and detected AF in compressively sensed ECG (improved TPR) while being computationally efficient (low reconstruction ratio).