Multi-Modal Voice Biomarker Analysis For Early Asthma Detection Using Hybrid Deep Learning Models

Abstract

Early diagnosis of asthma is critical, particularly where access to diagnostic equipment is limited. A non-invasive, intelligent model for early asthma detection is presented, using multimodal voice biomarkers and a hybrid deep learning architecture. Physiological audio cuesphonated vowels, respiratory acoustics, and coughswere recorded from asthma patients and healthy controls, capturing subtle pathological changes in airway function. Preprocessing includes artifact suppression, frequency-domain filtering, time-domain segmentation, and amplitude normalization, followed by extraction of discriminative descriptors: cepstral and spectral representations, perturbation-based voice quality indices, and respiration-specific temporal markers of abnormal airflow. A hybrid convolutional long short-term memory network learns local spectral patterns and temporal dynamics across speech and breathing cycles to distinguish asthmatics from non-asthmatics. Performance is evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and sensitivity. This approach offers a scalable, clinically applicable screening tool with potential in telemedicine, wearable health monitoring, and remote respiratory diagnostics.