Chronic heart failure (CHF) is a complex clinical syndrome characterised by the inability of the heart to provide sufficient perfusion to meet the body’s metabolic demands. It occurs primarily in the elderly and currently affects 64.3 million people worldwide. Heart failure is associated with significant morbidity and mortality as well as with prohibitive utilization of healthcare resources. Novel technologies that would improve patient management and reduce the burden of HF on healthcare resources are thus urgently needed. We assessed the performance of machine learning algorithms for predicting decompensation in CHF using heart sound data obtained by two different setups. The most accurate model was a decision tree classifier that achieved accuracy, precision, recall, F1 score, and area and the receiver operating curve of 0.896, 0.797, 0.812, 0.801, and 0.898, respectively. We also identified the most relevant predictor features extracted from different frequency bands of the recordings. Our analysis suggests that the low-frequency abnormal heart sounds do not play a critical role in detecting decompensation episodes in CHF patient cohort.