A Multi-Input Neural Network for Early Detection of Neurological Disordersfrom Vocal and Sleep Data
TL;DR
A multi-input deep neural network that integrates vocal biomarkers and clinical sleep-related features to improve diagnostic accuracy and demonstrates consistent improvements over early and late fusion strategies, demonstrating the benefit of modality-specific representation learning.
Abstract
: Early detection of neurological disorders is critical for effective treatment planning and improved quality of life. This study proposes a multi-input deep neural network that integrates vocal biomarkers and clinical sleep-related features to improve diagnostic accuracy. The model processes each modality through separate neural branches before combining high-level representations for final classification. We evaluate the approach using two public datasets: a Parkinson’s disease dataset containing 1195 voice samples and a sleep-disorder dataset with 80 patient records. Experimental results show that the proposed model outperforms classical machine learning baselines and single-modality deep learning models, achieving an accuracy of 92.5%, precision of 90.2%, recall of 94.1%, F1-score of 92.1%, and an AUC-ROC of 0.953. The architecture demonstrates consistent improvements over early and late fusion strategies, demonstrating the benefit of modality-specific representation learning. This work provides a scalable and non-invasive diagnostic framework with strong potential for clinical screening and monitoring.