Type 2 diabetes mellitus (T2DM) can induce impairments in vocal fold function and neural control, resulting in systematic changes in vowel articulation that may serve as objective biomarkers for speech-based disease detection. Traditional sentence-level approaches are susceptible to linguistic variability, limiting their ability to extract disease-specific acoustic features and reducing overall robustness. This study presents a vowel-based intelligent recognition framework for T2DM (VDT2) designed to capture stable and discriminative pathological speech features. The framework combines Lasso regression with recursive feature elimination to identify the most relevant acoustic indicators, and applies a dynamic logistic regression ensemble enhanced with an attention mechanism to strengthen feature representation. Experiments conducted on a self-constructed speech dataset demonstrate that VDT2 achieves a detection accuracy of 78%, outperforming conventional sentence-level methods. These findings highlight the potential of vowel-based analysis as a robust and non-invasive tool for T2DM detection.
A pilot feasibility study of a tablet-based virtual community application with shared avatars for promoting health behavior change in older adult care facilities
BackgroundMaintaining exercise and medication habits is crucial for older adults, but conventional reminder-based digital interventions often produce only transient effects.MethodsWe conducted a mixed-methods pilot feasibility