SLINet: Dysphasia detection in children using deep neural network. (July 2021)

• Record Type: Journal Article
• Authors: Kaushik, Manoj; Baghel, Neeraj; Burget, Radim; Travieso, Carlos M.; Dutta, Malay Kishore
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Developmental dysphasia -- Envelop modulation spectra -- Diagnosis; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102798
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Deep learning model to diagnose SLI at the early stage. Approach consists of utterances that contained seven types of vocabulary. A rigorous analysis based on different age-group. Approach is gender, age and speaker-independent Accuracy achieved up to 99.09 %. Abstract: A child has specific language impairment (SLI) or developmental dysphasia (DD) when the speech is delayed or has disordered language development for no apparent reason. As it may be related to loss of hearing, speech abnormality should be diagnosed at an early stage. The existing methods are mainly based on the utterance of vowels and have a high misclassification rate. This article proposes an automatic deep learning model that can be an effective tool to diagnose SLI at the early stage. In the proposed work, raw audio data is processed using Short-time Fourier transform and converted to decibel (dB) scaled spectrograms which are classified using the proposed convolutional neural network (CNN). This approach consists of utterances that contained seven types of vocabulary (vowels, consonant and different syllable Isolated words). A rigorous analysis based on different age-group was performed and a 10-fold Cross-Validation (CV) was done to test the accuracy of the classifier. A comprehensive experimental test reveals that 99.09 % of the children are correctly diagnosed by the proposed framework, which is superior when compared to state-of-the-art methods. The proposed scheme is gender and speaker-independent. The proposed model can be used as a stand-alone diagnostic tool that can assist automatic diagnosis of children for SLI and will be helpful for remote areas where professionals are not available. The proposed model is robust, efficient with low time complexity which is suitable for real-time applications.