Unsupervised segmentation of heart sounds

Digitized and recorded heart sounds contain significant diagnostic information, and are easily acquired in clinical practice. Part of a hypothetical diagnostic system is the segmentation subsystem which locates the boundaries between heartbeats. This thesis examines the design and performance of two approaches to segmentation: peak energy detection and sliding window autocorrelation. Emphasis is placed on synchronous detection of the heartbeats, so that subsequent subsystems can superimpose the heartbeats.An experimental database of heart sounds was compiled to assess the performance of the proposed algorithms. 206 sound files were gathered from 4 sources; the database includes 2709 heartbeats and covers a range of sound quality and complexity. A heart rate estimator is presented and tested, yielding a meaningful beat period in 90- 96% of typical heart sounds.The peak energy segmentation algorithm compares the signal energy against a threshold proportional to the signal's local average energy. It correctly segments 64% of the heartbeats from the experimental database, indicating the difficulty of segmentation in general.The novel sliding window autocorrelation segmenter is designed to take advantage of the self consistency of heart sounds. The algorithm operates by calculating multiple local autocorrelations in order to estimate the time offset to the next heart beat. A change in the time offset marks the boundary between heartbeats. It correctly segmented 83% of the heartbeats