Unsupervised segmentation of heart sounds from abrupt changes detection.

Background And Objective: Heart auscultation enables early diagnosis of cardiovascular diseases. Automated segmentation of cardiograms into fundamental heart states can guide physicians to analyze the patient's condition more effectively. In this work, we propose an unsupervised method of segmentation into heart sounds and silences based on the detection of abrupt changes in the signal.

Methods: Our procedure involves two steps. First, the abrupt changes, which correspond to the beginning and end of the heart sounds, are localized. Heart sounds and silences are then identified by calculating the signal power in each interval defined by the change points. The parameters of our algorithm are adjusted on the basis of estimated heart rate alone.

Results: We evaluate our method on three independent open-access database (PhysioNet 2016, CirCor DigiScope and PASCAL) for healthy and pathological populations, with or without murmurs. It achieves mean F score detection performance of 91.2%, 94.3% and 96.3% respectively, outperforming most of the competing unsupervised approaches.

Conclusion: By providing top ranking detection performance for three different types of heart sounds database, the proposed algorithm is reliable and robust, yet easy to implement.

Significance: This paper presents a simple and effective alternative segmentation method that can help improve the physiological interpretation of heart sounds recordings.