A Novel AI-Driven Markov Chain Approach to Dementia Detection

Abstract

Dementia, one of the most prevalent and devastating neurodegenerative diseases, affects millions worldwide. Identifying early markers of dementia is crucial for timely intervention and management. Current non-invasive diagnostic methods, including the Montreal Cognitive Assessment (MoCA) and the Saint Louis University Mental Status (SLUMS), are paper-and-pencil-based, manually interpreted, and time-consuming. This paper introduces a first-of-its-kind, interpretable artificial intelligence (IAI) approach that utilizes first-order Markov Chain models to analyze linguistic patterns for detecting early-stage dementia. By computing steady-state probabilities for each character in speech transcripts from both dementia subjects and healthy controls, our analysis revealed that the space character “ ”, representing pauses and voice breaks, and specific letters such as “n” and “i”, exhibited statistically significant differences between the groups. Principal Component Analysis (PCA) visualization demonstrated clear clustering of the two groups based on these features, indicating effective differentiation by our method. These linguistic markers were validated through Kolmogorov-Smirnov (KS) testing, confirming distinct probability distributions between dementia subjects and healthy controls. Furthermore, employing Lasso Logistic Regression on all available features, our best model achieved a state-of-the-art (SOTA) cross-validated AUC of 80.6%, emphasizing the robustness and effectiveness of the selected linguistic features in early dementia detection. Our findings not only reinforce the potential of automated speech analysis as a non-invasive, cost-effective tool for cognitive health assessments but also demonstrate the utility of character-level Markov Chain analysis in enhancing the sensitivity and timeliness of early dementia detection.