Machine learning-based detection of cognitive decline using SSWTRT: classification performance and decision analysis.

Area of Science:

• Neurology
• Cognitive Science
• Machine Learning

Background:

• Early detection of cognitive decline is crucial for managing dementia progression.
• Conventional screening tools like the Mini-Mental State Examination (MMSE) are time-consuming and require trained personnel.
• Dementia is linked to deficits in tactile and visual perception.

Purpose of the Study:

• To evaluate the Sound Symbolic Word Texture Recognition Test (SSWTRT) for identifying cognitive decline.
• To apply machine learning to SSWTRT responses for automated cognitive status classification.
• To assess the feasibility of a rapid, self-administered cognitive screening tool.

Main Methods:

• 233 participants with idiopathic normal pressure hydrocephalus (iNPH) completed the SSWTRT.
• The SSWTRT involves describing perceived texture from images using Japanese sound-symbolic words (SSWs).
• Machine learning classifiers (SVM, Random Forest, KNN) were trained to predict MMSE scores, using SSWTRT responses, age, and education.

Main Results:

• A balanced Support Vector Machine (SVM) model achieved moderate classification performance (accuracy, precision, recall, F1, AUC = 0.72).
• SHapley Additive exPlanations (SHAP) identified specific image textures (soft, coarse) as key predictors.
• Some responses indicated potential age-related sensory decline interference, not solely cognitive impairment.

Conclusions:

• Machine learning analysis of SSWTRT responses can moderately identify individuals with potential cognitive decline.
• The SSWTRT offers a non-invasive, resource-efficient screening approach.
• The framework provides a basis for developing scalable, language-specific cognitive screening tools, though further validation is needed.