Machine Learning-Based Detection of Cognitive Decline Using SSWTRT: Classification Performance and Decision Analysis

Yuji Nozaki^1*Chihiro Kamohara²Ryota Abe¹Taiki Ieda¹Madoka Nakajima²Maki Sakamoto^1*

• ¹The University of Electro-Communications, Chofu, Japan
• ²Juntendo University School of Medicine, Tokyo, Japan

In high-income countries, early detection of dementia is a pressing public health priority. The Mini-Mental State Examination (MMSE) is the most widely used screening tool, but its administration requires trained personnel, limiting large-scale, cost-effective deployment. Building on prior findings that patients with dementia are prone to tactile perceptual impairments, we previously proposed the Sound-Symbolic Word Texture Recognition Test (SSWTRT) as a low-cost, low-resource method for the early detection of cognitive decline. Because SSWTRT has minimal time and resource demands, it may offer practical value when used to predict MMSE-defined risk groups. In this study, we applied machine-learning models to experimental SSWTRT data to classify individuals with suspected cognitive decline, such as those with possible MCI (e.g., MMSE ≤27), and evaluated model performance and outputs. We further analyzed model interpretability using SHapley Additive exPlanations (SHAP) to quantify each image's contribution to classification. The model achieved an accuracy of 0.71, precision of 0.72, recall of 0.72, and F1-score of 0.72, demonstrating potential utility for screening cognitive decline. SHAP analyses indicated that responses to specific images— particularly those depicting soft and coarse textures—had strong influence on predictions, suggesting an association between texture perception and cognitive decline.