Journal of the Faculty of Engineering and Architecture of Gazi University, vol.39, no.4, pp.2023-2030, 2024 (SCI-Expanded)
Bowel sound (BS), a measure of bowel activity, can be observed through listening. By utilizing BS, many studies have been conducted for the early, harmless and practical detection of intestinal diseases. Basically, single (SB) and multiple (MB) burst-like bowel sounds, although observable with simple microphones, may not be accurately detected due to their abrupt character, long quiet periods (QP), and may be confused with other sounds such as stomach, muscle, breath. In this study, after the preprocessing steps, a distribution matrix (P) was formed by bringing together the characteristic time-frequency features specific to bowel sounds, and new reflection space (Q) stretching the indifference subspace was obtained from the eigenvectors corresponding to zero or near zero eigenvalues of this matrix. In order to determine which class a record belongs to; it will be sufficient to look at the common vector of which class its projection to the new space converges with the co-change matrix. In experimental studies, the average rates of SB, MB, QP, and non-BS classes in one-minute recordings were 2.3%, 0.3%, 92.9%, and 4.5%, respectively, while in tests performed with one-minute recordings that were never used in training, 87.5% of single bursts (SB), 35.7% of multiple bursts (MB), 84.3% of non-BS parts were assigned to the correct classes. As a result, by using this new reflection space (Q), which brings the in-class samples closer to each other and distances the inter-class samples from each other by looking at the distributions of all classes, bowel sounds can be largely separated from other sounds, regardless of the training set, without consulting medical professionals.