Bipartite Dynamic Pattern and Penta Pooling-Based Valvular Heart Disorder Detection Model

Abstract

Valvular heart disorders (VHD) have high mortality rates, making early detection essential. Machine learning offers a strong solution for improving diagnosis. This study presents a self-organised feature engineering model designed for high classification accuracy. A large dataset of respiratory sounds, with over 10,000 samples in 10 classes, was used. A new dynamic pattern, called bipartite dynamic pattern (BDP), was introduced. A novel pooling method, penta pooling (PP), was also proposed. The model extracted features using a multilevel approach by combining BDP and PP. BDP applied three feature kernels to generate three feature vectors, leading to seven combined feature sets. Feature selection used neighbourhood component analysis (NCA) and Chi-squared (Chi2) methods, producing 14 selected feature vectors. Classification was performed using k-nearest neighbours (kNN) and support vector machine (SVM). A 10-fold cross-validation process created 28 classification outcomes. An information fusion step was applied, using iterative majority voting (IMV) to refine the final decision. In total, 54 results were generated (28 classifier-based, 26 voted). The proposed BDP and PP-based model achieved 99.76% classification accuracy, demonstrating its effectiveness.