Multimode Vector Modalities of HMM-GMM in Augmented Categorization of Bioacoustics' Signals

Y usefultool in classification of dominant characteristics in biological data. In particular, the aimof this approach was to enhance classifications of lung sounds (LS) and heart sounds (HS). In order to achieve these objectives, the LS and HS signals were expressed in terms of Mel-frequency cepstral coefficients (MFCCs) and Quantile acoustic vectors. Once the signals were vectorized, a clusters’ quantity analysis for the LS and HS signals was executed for both classes, representing normal abnormal sounds, in such a way that a criterion for the model’s size was obtained. The clusters’ quantity analysis was carried out applying dendrograms, silhouettes and the Bayesian Information Criterion (BIC). Starting from these computations, the HMM-GMM model architecture for the normal and abnormal classes were conceptualized.The models for the LS signals using Quantile vectors, specifically Quartile, yielded excellent results, while for HS signals, the best results for the HMM-GMM models were obtained with MFCC vectors. In both cases, i.e., LS and HS signals, a close to 100% classification efficiency was achieved for studied cases. Furthermore, the evaluations were assessed in terms of sensitivity and specificity defined as a true positive rate and a true negative rate respectively; LS signals achieved a 100% in sensitivity and specificity, while HS signals also reached a 100%,excludingthe normal vs stenosis case, which obtained a 85% in specificity. The importance of this approach lies is the possibility of implementing automated assessment diagnostics for patients with respiratory and cardiac disorders, and essentially the ability to bring this diagnostic capability to remote and limited medical resource areas utilizing low cost technologies.