Effect of the Initial Condition on SCG Clustering using Unsupervised Machine Learning

• Published in: IEEE Signal Processing in Medicine and Biology Symposium pp. 1 - 5
• Main Authors: Ahdy, S., Hassan, T., Rahman, B., Sandler, R., Mansy, H.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 07.12.2024
• Subjects: Accuracy; Clustering algorithms; Feature extraction; Machine learning; Signal processing algorithms; Sleep apnea; Testing; Timing; Valves; Vibrations
• ISSN: 2473-716X
• DOI: 10.1109/SPMB62441.2024.10842232

Seismocardiographic (SCG) signals refer to chest surface vibrations induced by the cardiac cycle [1] , [2] . SCG may be generated by valve closure, blood movement and cardiac muscle contraction [3] - [5] . SCG may serve as a non-invasive clinical tool to detect heart disease [6] and sleep apnea [7] and may be used to determine cardiac time intervals [8] . However, SCG signal variability can hinder accurate extraction of clinically useful SCG features [9] , such as cardiac time intervals [10] . SCG variability can be reduced by clustering SCG beats [1] , [9] . Respiration was found to be a source of SCG variability [1] , [11] . For example, respiration changes the heart rate (a phenomenon known as respiratory sinus arrhythmia) which affects the length of the cardiac cycle [1] , [12] . Respiration can also modulate the cardiac preload and afterload which can affect the intensity of heart sounds and cardiac time intervals [11] , [13] , [14] . Therefore, early efforts of SCG clustering grouped SCG beats based on the respiratory phase, e.g., high/low lung volume (HLV/LLV) or inspiration/expiration (INS/EXP) [15] . In recent studies, unsupervised machine learning (e.g., K-medoid) was used as it does not require any prior assumptions about SCG clusters [9] , [16] . In the K-medoid algorithm, cluster medoids are initially chosen either randomly [9] or based on a physiological basis (e.g., HLV/LLV or INS/EXP) [1] , then SCG beats are assigned to closest medoid, and cluster medoids are updated; the process is repeated until cluster assignments converge (i.e., stop changing) [1] . The quality of the clustering solution (i.e., final cluster assignment) can be assessed by calculating the intra-cluster and inter-cluster variabilities [1] , [16] . However, based on the choice of the initial medoids and cluster assignment (i.e., the initial condition), different clustering solutions could be obtained [9] .