Article
Author
This research aims to develop a non-invasive and accurate system for detecting and classifying laryngeal pathologies – especially cancerous and precancerous lesions – by combining audio and laryngeal bioimpedance signals in a multimodal deep learning classification architecture. It addresses the central research question: Can a multimodal deep learning framework using audio and electroglottographic (EGG) signals outperform traditional and unimodal approaches in classifying laryngeal conditions?
To answer this, a novel dataset was collected, featuring simultaneous recordings of audio and laryngeal bioimpedance from healthy individuals and patients with a variety of laryngeal disorders. The methodology includes feature extraction using Equivalent Rectangular Bandwidth-based methods, and the design of both unimodal and multimodal deep learning models based on Convolutional Neural Networks (CNNs). The performance of Recurrent Neural Networks (RNNs) for laryngeal pathology classification is also investigated.
The final system adopts a late fusion architecture: two one-dimensional CNNs independently process each modality, and their outputs are combined using a stacked generalisation approach with an ECOC-based meta-classifier. Models were trained and evaluated across two datasets – a custom database and the Saarbruecken Voice Database (SVD) – to ensure generalisability of the developed
system.
The results show that the multimodal classification approach significantly outperforms unimodal baselines, particularly in malignant case detection, achieving 89.23% ± 1.95 accuracy, 82.32% ± 2.78 precision, and 83.27% ± 3.43 sensitivity. For general pathology detection, the system achieved even higher classification metrics (94.92% ± 2.82 accuracy and 96.67% ± 2.90 precision). Continuous speech outperformed the sustained phonation, and ERB-based feature provided richer pathological discrimination than other feature representations.
This study contributes a robust, accurate, and clinically relevant framework for automated laryngeal pathology screening. Its findings support
the integration of multimodal voice analysis into future diagnostic tools and lay the groundwork for real-time, non-invasive applications in clinical settings.
Keywords: laryngeal pathologies, bioimpedance, multimodal deep learning classification architecture
How to Cite: Tomaszewska, J. Z. (2026) “Use of audio, bioimpedance signals, and multimodal deep learning in classification of laryngeal pathologies”, New Vistas. 12(1). doi: https://doi.org/10.36828/newvistas.404