Enhancing Early Dementia Detection with AI and Character-Level Speech Biomarkers

Abstract

The rising prevalence of neurocognitive disorders, such as Alzheimer’s disease (AD), poses a significant global health challenge. Traditional diagnostic methods, including clinical interviews and paper-based tests like the Mini-Mental State Examination (MMSE), Mini-Cog Test, and Montreal Cognitive Assessment (MoCA), are limited by subjectivity, memory biases, and interviewer variability. To address these limitations, there is an increasing need for noninvasive, real-time, and data-driven diagnostic tools. This study introduces an interpretable artificial intelligence (AI) approach to analyze speech transcripts for the early detection of dementia signs. Initially, each unique character in the speech transcripts is encoded as a distinct number. Recurrence Quantification Analysis (RQA) is then applied to generate recurrence plots, capturing dynamic linguistic patterns associated with cognitive decline. From these plots, we extract salient features, termed linguistic biomarkers, using deep metric learning with Siamese networks, which effectively represent essential linguistic characteristics. These biomarkers are transformed into numerical embeddings and processed using an XGBoost classifier, recognized for its robustness, to differentiate between dementia and healthy subjects. Utilizing stratified K-Fold cross-validation, our model achieves a mean ROC AUC of 88.5% ± 2.9%, demonstrating strong and consistent performance across diverse data subsets. Additionally, SHAP (SHapley Additive exPlanations) analysis identifies the most important linguistic features influencing the model's predictions, providing clear insights into the language patterns linked to cognitive impairment. This research not only enhances the accuracy of dementia diagnosis but also offers valuable understanding of the linguistic indicators of cognitive disorders, enabling the development of more effective and interpretable diagnostic tools in clinical settings.