Systems and methods for reconstructing heart sounds from heart sound samples taken under a sub-optimal condition, such as at a low sampling rate, are discussed. An exemplary system receives acceleration information from a patient sensed at a first sampling rate, and generate a heart sound ensemble of portions of acceleration information over multiple cardiac cycles. The system can reconstruct a heart sound segment to have a second sampling rate, higher than the first sampling rate, using the generated heart sound ensemble. A heart sound metric can be generated using the reconstructed heart sound segment, and used for detecting a cardiac event, such as a cardiac arrhythmia episode, or a worsening heart failure event.