Hybrid SVM and CNN System for Cough Detection and Health Classification

Abstract

This study introduces a hybrid architecture designed to address the challenges of automated cough sound analysis by combining unsupervised detection with supervised classification. The system integrates a One-Class Support Vector Machine (SVM) trained exclusively on positive cough recordings to establish an acoustic boundary that isolates true cough events (inliers) while rejecting out-of-distribution background noise (outliers) from raw audio streams, and a convolutional neural network (CNN) to classify the extracted cough sounds into three clinically relevant categories. The preprocessing pipeline is integral to the system's effectiveness. The process involves extracting static Mel-Frequency Cepstral Coefficients (MFCCs), followed by standardization and dimensionality alignment to ensure compatibility with the CNN. These steps ensure that the spectral and temporal characteristics of cough sounds are accurately represented while reducing redundancy and enhancing separability in the feature space. Experimental evaluation demonstrates the reliability of this hybrid framework. As a standalone classifier on clean, isolated cough sounds, the DenseNet achieves an accuracy of 96.0% and an F₁-score of 0.96. When evaluated on complex, unsegmented audio streams, the end-to-end hybrid system achieves an overall accuracy of 93.0% and a corresponding F₁-score of 0.91. These results underscore the system's resilience and capacity for generalization, even when confronted with noisy or heterogeneous data. The incorporation of One-Class SVM for unsupervised detection ensures that cough events are effectively isolated, while the CNN provides precise classification, making the system suitable for real-world applications. Generalizability requires further validation on prospective external cohorts. Beyond its clinical relevance, this work also holds pedagogical value for engineering education. It illustrates the integration of signal processing, machine learning, and system design in a coherent pipeline, serving as a practical case study for students and researchers. By integrating theoretical concepts with applied methodologies, the system illustrates how hybrid architectures can be utilized to address complex challenges in health monitoring. Overall, the proposed framework offers a reliable foundation for preliminary screening and educational purposes.