Optimal tuning of support vector machines and k-NN algorithm by using Bayesian optimization for newborn cry signal diagnosis based on audio signal processing features

• Published in: Chaos, solitons and fractals Vol. 167; p. 112972
• Main Authors: Lahmiri, Salim, Tadj, Chakib, Gargour, Christian, Bekiros, Stelios
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2023
• Subjects: Auditory-inspired amplitude modulation; Bayesian optimization; k-Nearest neighbors; Mel-frequency cepstral coefficients; Newborn cry; Prosody; Support vector machines; Support vector machines; Mel-frequency cepstral coefficients; Prosody; Newborn cry; k-Nearest neighbors; Auditory-inspired amplitude modulation; Bayesian optimization
• ISSN: 0960-0779; 1873-2887
• DOI: 10.1016/j.chaos.2022.112972

Recently, the number of machine learning models used to classify cry signals of healthy and unhealthy newborns has been significantly increasing. Various works have already reported encouraging classification results; however, fine-tuning of the hyper-parameters of machine leaning algorithms is still an open problem in the context of newborn cry signal classification. This paper proposes to use Bayesian optimization (BO) method to optimize the hyper-parameters of Support Vector Machine (SVM) with radial basis function (RBF) kernel and k-nearest neighbors (kNN) trained with different audio features separately or combined; namely, mel-frequency cepstral coefficients (MFCC), auditory-inspired amplitude modulation (AAM), and prosody. Particularly, the chi-square test is applied to each set of features to retain the ten most significant ones used to train optimal classifiers. The accuracy, sensitivity, and specificity of each experimental model are computed following the standard 10-fold cross-validation protocol. One of the contributions is an improvement over previous works on newborn cry signal classification used to distinguish between healthy and unhealthy ones over the same database, in terms of performance. The best model is the SVM trained with AAM ten most significant features achieved 83.62 % ± 0.022 accuracy, 59.18 % ± 0.0469 sensitivity, and 93.87 % ± 0.0190 specificity followed by kNN trained with ten most features from MFCC, AAM, and prosody to obtain 82.88 % ± 0.0144 accuracy, 55.34 % ± 0.0350 sensitivity, and 94.42 % ± 0.0075 specificity. These results outperformed existing works validated on the same database. In addition, optimally tuned SVM and kNN are fed with a restricted number of selected patterns so as the processing time for training and testing is significantly limited. This means that the RBF-SVM-BO classifier trained with AAM ten most significant features is more able to distinguish between healthy and unhealthy newborns. •Nonlinear SVM and kNN algorithm trained with different audio features to classify newborn cry audios.•Chi-square test is applied for selection of audio features.•Bayesian optimization is used to tune hyper-parameters of SVM and kNN.•The best model is the SVM trained with AAM ten most significant features•These results outperformed existing works validated on the same database.